A support agent stops mid-task on complex queries. Its loop checks `response.content[0].type == "text"` to decide it is finished. What is the fix?

1:"$Sreact.fragment" 3:I[339756,["/_next/static/chunks/33t46atd3n2zd.js","/_next/static/chunks/1uoo0zstz-tka.js","/_next/static/chunks/1gmtbr53vpwcw.js","/_next/static/chunks/0hwl0k3t96f8i.js","/_next/static/chunks/07dyy4_l8-v8w.js","/_next/static/chunks/1a28s5qpw8ubo.js","/_next/static/chunks/3nx9-unhdiib7.js","/_next/static/chunks/1l3wx9j9z2da9.js","/_next/static/chunks/10prs2o5tycig.js","/_next/static/chunks/02ocli407gubc.js","/_next/static/chunks/2r--443asccvg.js","/_next/static/chunks/0rjt9gb10d05i.js","/_next/static/chunks/3lr20ikf1jmri.js","/_next/static/chunks/2mak3tsoa7kik.js","/_next/static/chunks/39kizkc9pkpb1.js","/_next/static/chunks/2uicbuh3o1h8o.js","/_next/static/chunks/1e-wsdiuv_p_u.js","/_next/static/chunks/12syjpiwnqlnq.js","/_next/static/chunks/0dnfb3y43dtv8.js"],"default"] 4:I[837457,["/_next/static/chunks/33t46atd3n2zd.js","/_next/static/chunks/1uoo0zstz-tka.js","/_next/static/chunks/1gmtbr53vpwcw.js","/_next/static/chunks/0hwl0k3t96f8i.js","/_next/static/chunks/07dyy4_l8-v8w.js","/_next/static/chunks/1a28s5qpw8ubo.js","/_next/static/chunks/3nx9-unhdiib7.js","/_next/static/chunks/1l3wx9j9z2da9.js","/_next/static/chunks/10prs2o5tycig.js","/_next/static/chunks/02ocli407gubc.js","/_next/static/chunks/2r--443asccvg.js","/_next/static/chunks/0rjt9gb10d05i.js","/_next/static/chunks/3lr20ikf1jmri.js","/_next/static/chunks/2mak3tsoa7kik.js","/_next/static/chunks/39kizkc9pkpb1.js","/_next/static/chunks/2uicbuh3o1h8o.js","/_next/static/chunks/1e-wsdiuv_p_u.js","/_next/static/chunks/12syjpiwnqlnq.js","/_next/static/chunks/0dnfb3y43dtv8.js"],"default"] 6:I[347257,["/_next/static/chunks/33t46atd3n2zd.js","/_next/static/chunks/1uoo0zstz-tka.js","/_next/static/chunks/1gmtbr53vpwcw.js","/_next/static/chunks/0hwl0k3t96f8i.js","/_next/static/chunks/07dyy4_l8-v8w.js","/_next/static/chunks/1a28s5qpw8ubo.js","/_next/static/chunks/3nx9-unhdiib7.js","/_next/static/chunks/1l3wx9j9z2da9.js","/_next/static/chunks/10prs2o5tycig.js","/_next/static/chunks/02ocli407gubc.js","/_next/static/chunks/2r--443asccvg.js","/_next/static/chunks/0rjt9gb10d05i.js","/_next/static/chunks/3lr20ikf1jmri.js","/_next/static/chunks/2mak3tsoa7kik.js","/_next/static/chunks/39kizkc9pkpb1.js","/_next/static/chunks/2uicbuh3o1h8o.js","/_next/static/chunks/1e-wsdiuv_p_u.js","/_next/static/chunks/12syjpiwnqlnq.js","/_next/static/chunks/0dnfb3y43dtv8.js"],"ClientPageRoot"] 7:I[481209,["/_next/static/chunks/33t46atd3n2zd.js","/_next/static/chunks/1uoo0zstz-tka.js","/_next/static/chunks/1gmtbr53vpwcw.js","/_next/static/chunks/0hwl0k3t96f8i.js","/_next/static/chunks/07dyy4_l8-v8w.js","/_next/static/chunks/1a28s5qpw8ubo.js","/_next/static/chunks/3nx9-unhdiib7.js","/_next/static/chunks/1l3wx9j9z2da9.js","/_next/static/chunks/10prs2o5tycig.js","/_next/static/chunks/02ocli407gubc.js","/_next/static/chunks/2r--443asccvg.js","/_next/static/chunks/0rjt9gb10d05i.js","/_next/static/chunks/3lr20ikf1jmri.js","/_next/static/chunks/2mak3tsoa7kik.js","/_next/static/chunks/39kizkc9pkpb1.js","/_next/static/chunks/2uicbuh3o1h8o.js","/_next/static/chunks/1e-wsdiuv_p_u.js","/_next/static/chunks/12syjpiwnqlnq.js","/_next/static/chunks/0dnfb3y43dtv8.js","/_next/static/chunks/1x1n57nb1d8gg.js","/_next/static/chunks/1m_1gkpcpblc9.js","/_next/static/chunks/0av03tekodyp8.js","/_next/static/chunks/0c_c2b4z1yc02.js","/_next/static/chunks/3s032g6m10tm2.js"],"default"] 8:I[897367,["/_next/static/chunks/33t46atd3n2zd.js","/_next/static/chunks/1uoo0zstz-tka.js","/_next/static/chunks/1gmtbr53vpwcw.js","/_next/static/chunks/0hwl0k3t96f8i.js","/_next/static/chunks/07dyy4_l8-v8w.js","/_next/static/chunks/1a28s5qpw8ubo.js","/_next/static/chunks/3nx9-unhdiib7.js","/_next/static/chunks/1l3wx9j9z2da9.js","/_next/static/chunks/10prs2o5tycig.js","/_next/static/chunks/02ocli407gubc.js","/_next/static/chunks/2r--443asccvg.js","/_next/static/chunks/0rjt9gb10d05i.js","/_next/static/chunks/3lr20ikf1jmri.js","/_next/static/chunks/2mak3tsoa7kik.js","/_next/static/chunks/39kizkc9pkpb1.js","/_next/static/chunks/2uicbuh3o1h8o.js","/_next/static/chunks/1e-wsdiuv_p_u.js","/_next/static/chunks/12syjpiwnqlnq.js","/_next/static/chunks/0dnfb3y43dtv8.js"],"OutletBoundary"] 9:"$Sreact.suspense" c:I[897367,["/_next/static/chunks/33t46atd3n2zd.js","/_next/static/chunks/1uoo0zstz-tka.js","/_next/static/chunks/1gmtbr53vpwcw.js","/_next/static/chunks/0hwl0k3t96f8i.js","/_next/static/chunks/07dyy4_l8-v8w.js","/_next/static/chunks/1a28s5qpw8ubo.js","/_next/static/chunks/3nx9-unhdiib7.js","/_next/static/chunks/1l3wx9j9z2da9.js","/_next/static/chunks/10prs2o5tycig.js","/_next/static/chunks/02ocli407gubc.js","/_next/static/chunks/2r--443asccvg.js","/_next/static/chunks/0rjt9gb10d05i.js","/_next/static/chunks/3lr20ikf1jmri.js","/_next/static/chunks/2mak3tsoa7kik.js","/_next/static/chunks/39kizkc9pkpb1.js","/_next/static/chunks/2uicbuh3o1h8o.js","/_next/static/chunks/1e-wsdiuv_p_u.js","/_next/static/chunks/12syjpiwnqlnq.js","/_next/static/chunks/0dnfb3y43dtv8.js"],"ViewportBoundary"] e:I[897367,["/_next/static/chunks/33t46atd3n2zd.js","/_next/static/chunks/1uoo0zstz-tka.js","/_next/static/chunks/1gmtbr53vpwcw.js","/_next/static/chunks/0hwl0k3t96f8i.js","/_next/static/chunks/07dyy4_l8-v8w.js","/_next/static/chunks/1a28s5qpw8ubo.js","/_next/static/chunks/3nx9-unhdiib7.js","/_next/static/chunks/1l3wx9j9z2da9.js","/_next/static/chunks/10prs2o5tycig.js","/_next/static/chunks/02ocli407gubc.js","/_next/static/chunks/2r--443asccvg.js","/_next/static/chunks/0rjt9gb10d05i.js","/_next/static/chunks/3lr20ikf1jmri.js","/_next/static/chunks/2mak3tsoa7kik.js","/_next/static/chunks/39kizkc9pkpb1.js","/_next/static/chunks/2uicbuh3o1h8o.js","/_next/static/chunks/1e-wsdiuv_p_u.js","/_next/static/chunks/12syjpiwnqlnq.js","/_next/static/chunks/0dnfb3y43dtv8.js"],"MetadataBoundary"] 10:I[168027,["/_next/static/chunks/33t46atd3n2zd.js","/_next/static/chunks/1uoo0zstz-tka.js","/_next/static/chunks/1gmtbr53vpwcw.js","/_next/static/chunks/0hwl0k3t96f8i.js","/_next/static/chunks/07dyy4_l8-v8w.js","/_next/static/chunks/1a28s5qpw8ubo.js","/_next/static/chunks/3nx9-unhdiib7.js","/_next/static/chunks/1l3wx9j9z2da9.js","/_next/static/chunks/10prs2o5tycig.js","/_next/static/chunks/02ocli407gubc.js","/_next/static/chunks/2r--443asccvg.js","/_next/static/chunks/0rjt9gb10d05i.js","/_next/static/chunks/3lr20ikf1jmri.js","/_next/static/chunks/2mak3tsoa7kik.js","/_next/static/chunks/39kizkc9pkpb1.js","/_next/static/chunks/2uicbuh3o1h8o.js","/_next/static/chunks/1e-wsdiuv_p_u.js","/_next/static/chunks/12syjpiwnqlnq.js","/_next/static/chunks/0dnfb3y43dtv8.js"],"default",1] :HL["/_next/static/chunks/2j4g1xjp6d0vl.css","style"] :HL["/_next/static/chunks/2o6hxm_-g5yv3.css","style"] :HL["/_next/static/chunks/1dfycg2iy82zu.css","style"] :HL["/_next/static/media/1bffadaabf893a1e-s.p.3-6t-g6q0vh0a.woff2","font",{"crossOrigin":"","type":"font/woff2"}] :HL["/_next/static/media/5e5dfa9e86f764b8-s.p.22x-ws2e_cdqh.woff2","font",{"crossOrigin":"","type":"font/woff2"}] :HL["/_next/static/media/748b8b5301d2bdb2-s.p.1zvx97hkh1y6p.woff2","font",{"crossOrigin":"","type":"font/woff2"}] :HL["/_next/static/media/797e433ab948586e-s.p.0r6juujl39pe6.woff2","font",{"crossOrigin":"","type":"font/woff2"}] :HL["/_next/static/media/83afe278b6a6bb3c-s.p.2bn3s6zvc0dyp.woff2","font",{"crossOrigin":"","type":"font/woff2"}] :HL["/_next/static/media/8c2eb9ceedecfc8e-s.p.23aeddxv5enbo.woff2","font",{"crossOrigin":"","type":"font/woff2"}] :HL["/_next/static/media/bbc41e54d2fcbd21-s.p.1rgnod-3esatf.woff2","font",{"crossOrigin":"","type":"font/woff2"}] b:X 0:{"P":null,"c":["","exam","6a1a0c49ca96069990f7e8bc","play"],"q":"","i":false,"f":[[["",{"children":["exam",{"children":[["id","6a1a0c49ca96069990f7e8bc","d",[]],{"children":["play",{"children":["__PAGE__",{},"$undefined","$undefined",4096]},"$undefined","$undefined",4096]},"$undefined","$undefined",4096]},"$undefined","$undefined",4096]},"$undefined","$undefined",4112],[["$","$1","c",{"children":[[["$","link","0",{"rel":"stylesheet","href":"/_next/static/chunks/2j4g1xjp6d0vl.css","precedence":"next","crossOrigin":"$undefined","nonce":"$undefined"}],["$","link","1",{"rel":"stylesheet","href":"/_next/static/chunks/2o6hxm_-g5yv3.css","precedence":"next","crossOrigin":"$undefined","nonce":"$undefined"}],["$","link","2",{"rel":"stylesheet","href":"/_next/static/chunks/1dfycg2iy82zu.css","precedence":"next","crossOrigin":"$undefined","nonce":"$undefined"}],["$","script","script-0",{"src":"/_next/static/chunks/33t46atd3n2zd.js","async":true,"nonce":"$undefined"}],["$","script","script-1",{"src":"/_next/static/chunks/1uoo0zstz-tka.js","async":true,"nonce":"$undefined"}],["$","script","script-2",{"src":"/_next/static/chunks/1gmtbr53vpwcw.js","async":true,"nonce":"$undefined"}],["$","script","script-3",{"src":"/_next/static/chunks/0hwl0k3t96f8i.js","async":true,"nonce":"$undefined"}],["$","script","script-4",{"src":"/_next/static/chunks/07dyy4_l8-v8w.js","async":true,"nonce":"$undefined"}],["$","script","script-5",{"src":"/_next/static/chunks/1a28s5qpw8ubo.js","async":true,"nonce":"$undefined"}],["$","script","script-6",{"src":"/_next/static/chunks/3nx9-unhdiib7.js","async":true,"nonce":"$undefined"}],["$","script","script-7",{"src":"/_next/static/chunks/1l3wx9j9z2da9.js","async":true,"nonce":"$undefined"}],["$","script","script-8",{"src":"/_next/static/chunks/10prs2o5tycig.js","async":true,"nonce":"$undefined"}],["$","script","script-9",{"src":"/_next/static/chunks/02ocli407gubc.js","async":true,"nonce":"$undefined"}],["$","script","script-10",{"src":"/_next/static/chunks/2r--443asccvg.js","async":true,"nonce":"$undefined"}],["$","script","script-11",{"src":"/_next/static/chunks/0rjt9gb10d05i.js","async":true,"nonce":"$undefined"}],["$","script","script-12",{"src":"/_next/static/chunks/3lr20ikf1jmri.js","async":true,"nonce":"$undefined"}],["$","script","script-13",{"src":"/_next/static/chunks/2mak3tsoa7kik.js","async":true,"nonce":"$undefined"}],["$","script","script-14",{"src":"/_next/static/chunks/39kizkc9pkpb1.js","async":true,"nonce":"$undefined"}],["$","script","script-15",{"src":"/_next/static/chunks/2uicbuh3o1h8o.js","async":true,"nonce":"$undefined"}],["$","script","script-16",{"src":"/_next/static/chunks/1e-wsdiuv_p_u.js","async":true,"nonce":"$undefined"}],["$","script","script-17",{"src":"/_next/static/chunks/12syjpiwnqlnq.js","async":true,"nonce":"$undefined"}],["$","script","script-18",{"src":"/_next/static/chunks/0dnfb3y43dtv8.js","async":true,"nonce":"$undefined"}]],"$L2"]}],{"children":[["$","$1","c",{"children":[null,["$","$L3",null,{"parallelRouterKey":"children","error":"$undefined","errorStyles":"$undefined","errorScripts":"$undefined","template":["$","$L4",null,{}],"templateStyles":"$undefined","templateScripts":"$undefined","notFound":"$undefined","forbidden":"$undefined","unauthorized":"$undefined"}]]}],{"children":[["$","$1","c",{"children":[[["$","script","script-0",{"src":"/_next/static/chunks/1x1n57nb1d8gg.js","async":true,"nonce":"$undefined"}]],"$L5"]}],{"children":[["$","$1","c",{"children":[null,["$","$L3",null,{"parallelRouterKey":"children","error":"$undefined","errorStyles":"$undefined","errorScripts":"$undefined","template":["$","$L4",null,{}],"templateStyles":"$undefined","templateScripts":"$undefined","notFound":"$undefined","forbidden":"$undefined","unauthorized":"$undefined"}]]}],{"children":[["$","$1","c",{"children":[["$","$L6",null,{"Component":"$7","serverProvidedParams":{"searchParams":{},"params":{"id":"6a1a0c49ca96069990f7e8bc"},"promises":null}}],[["$","script","script-0",{"src":"/_next/static/chunks/1m_1gkpcpblc9.js","async":true,"nonce":"$undefined"}],["$","script","script-1",{"src":"/_next/static/chunks/0av03tekodyp8.js","async":true,"nonce":"$undefined"}],["$","script","script-2",{"src":"/_next/static/chunks/0c_c2b4z1yc02.js","async":true,"nonce":"$undefined"}],["$","script","script-3",{"src":"/_next/static/chunks/3s032g6m10tm2.js","async":true,"nonce":"$undefined"}]],["$","$L8",null,{"children":["$","$9",null,{"name":"Next.MetadataOutlet","children":"$@a"}]}]]}],{},null,false,null]},null,false,"$b"]},null,false,null]},null,false,"$b"]},null,false,null],["$","$1","h",{"children":[null,["$","$Lc",null,{"children":"$Ld"}],["$","div",null,{"hidden":true,"children":["$","$Le",null,{"children":["$","$9",null,{"name":"Next.Metadata","children":"$Lf"}]}]}],["$","meta",null,{"name":"next-size-adjust","content":""}]]}],false]],"m":"$undefined","G":["$10",[["$","link","0",{"rel":"stylesheet","href":"/_next/static/chunks/2j4g1xjp6d0vl.css","precedence":"next","crossOrigin":"$undefined","nonce":"$undefined"}],["$","link","1",{"rel":"stylesheet","href":"/_next/static/chunks/2o6hxm_-g5yv3.css","precedence":"next","crossOrigin":"$undefined","nonce":"$undefined"}],["$","link","2",{"rel":"stylesheet","href":"/_next/static/chunks/1dfycg2iy82zu.css","precedence":"next","crossOrigin":"$undefined","nonce":"$undefined"}]]],"S":false,"h":null,"r":"$undefined","s":"$undefined","a":"$undefined","l":"$undefined","p":"$undefined","d":"$undefined","b":"oucAYIA6LsY4rpAiOzHHm"} 11:I[479520,["/_next/static/chunks/33t46atd3n2zd.js","/_next/static/chunks/1uoo0zstz-tka.js","/_next/static/chunks/1gmtbr53vpwcw.js","/_next/static/chunks/0hwl0k3t96f8i.js","/_next/static/chunks/07dyy4_l8-v8w.js","/_next/static/chunks/1a28s5qpw8ubo.js","/_next/static/chunks/3nx9-unhdiib7.js","/_next/static/chunks/1l3wx9j9z2da9.js","/_next/static/chunks/10prs2o5tycig.js","/_next/static/chunks/02ocli407gubc.js","/_next/static/chunks/2r--443asccvg.js","/_next/static/chunks/0rjt9gb10d05i.js","/_next/static/chunks/3lr20ikf1jmri.js","/_next/static/chunks/2mak3tsoa7kik.js","/_next/static/chunks/39kizkc9pkpb1.js","/_next/static/chunks/2uicbuh3o1h8o.js","/_next/static/chunks/1e-wsdiuv_p_u.js","/_next/static/chunks/12syjpiwnqlnq.js","/_next/static/chunks/0dnfb3y43dtv8.js"],""] :HL["https://fonts.googleapis.com/css2?family=Esteban&family=Inter:wght@400;600&family=IM+Fell+DW+Pica:ital@0;1&family=Roboto:wght@400;500&family=Raleway:wght@400;600&family=Montserrat:wght@400;600&family=Courier+Prime:ital@0;1&family=Newsreader:ital,wght@0,400;0,600;0,700;1,400;1,600&family=Poppins:wght@400;500;600;700&family=Geist:wght@400;500;600;700&family=DM+Sans:wght@400;500;600;700&family=Roboto+Mono:wght@400;500;600;700&family=Literata:ital,wght@0,400;0,600;0,700;1,400;1,600&family=STIX+Two+Text:ital,wght@0,400;0,600;0,700;1,400;1,600&family=Alegreya:ital,wght@0,400;1,400&family=JetBrains+Mono:wght@400;500;600;700&family=IBM+Plex+Mono:wght@400;500;600;700&family=Lora:ital,wght@0,400;0,600;1,400&family=Merriweather:wght@400;700&family=Nunito:wght@400;600&family=Google+Sans+Flex:opsz,wdth,wght@8..144,25..151,100..900&family=Lexend:wght@300;400;500;600;700&family=Special+Elite&display=swap","style"] b:C 12:Ta0f, (function() { function parseColorToHex(colorStr) { if (!colorStr) return ''; if (colorStr.indexOf('#') === 0) return colorStr; var rgbMatch = colorStr.match(/^rgba?\((\d+),\s*(\d+),\s*(\d+)/i); if (rgbMatch) { var r = parseInt(rgbMatch[1], 10).toString(16); var g = parseInt(rgbMatch[2], 10).toString(16); var b = parseInt(rgbMatch[3], 10).toString(16); return '#' + (r.length === 1 ? '0' + r : r) + (g.length === 1 ? '0' + g : g) + (b.length === 1 ? '0' + b : b); } return colorStr; } function syncThemeColor() { try { var bodyStyle = window.getComputedStyle(document.body); var bgColor = bodyStyle ? bodyStyle.backgroundColor : ''; if (!bgColor || bgColor === 'rgba(0, 0, 0, 0)' || bgColor === 'transparent') { var rootStyle = window.getComputedStyle(document.documentElement); bgColor = rootStyle ? rootStyle.backgroundColor : ''; } if (bgColor && bgColor !== 'rgba(0, 0, 0, 0)' && bgColor !== 'transparent') { var hexColor = parseColorToHex(bgColor); var metaTags = document.querySelectorAll('meta[name="theme-color"]'); if (!metaTags.length) { var meta = document.createElement('meta'); meta.setAttribute('name', 'theme-color'); meta.setAttribute('content', hexColor); document.head.appendChild(meta); } else { for (var i = 0; i < metaTags.length; i++) { if (i === 0) { metaTags[i].setAttribute('content', hexColor); metaTags[i].removeAttribute('media'); } else { metaTags[i].parentNode.removeChild(metaTags[i]); } } } } } catch (e) {} } if (document.readyState === 'loading') { document.addEventListener('DOMContentLoaded', syncThemeColor); } else { syncThemeColor(); } })(); 2:["$","html",null,{"suppressHydrationWarning":true,"data-flavor":"krog","data-font":"inter","data-material-font":"literata","data-font-custom":"false","data-material-font-custom":"false","data-background":"dots","data-navbar-position":"left","data-navbar-collapsed":"false","children":[["$","head",null,{"children":[["$","style",null,{"dangerouslySetInnerHTML":{"__html":"\n :root {\n \n \n }\n html[data-font-custom='true'] body {\n font-family: var(--custom-app-font), system-ui, sans-serif !important;\n }\n html[data-material-font-custom='true'] .study-material {\n font-family: var(--custom-material-font), Georgia, serif !important;\n }\n "}}],["$","meta",null,{"name":"theme-color","content":"#ffffff"}],["$","meta",null,{"name":"mobile-web-app-capable","content":"yes"}],["$","meta",null,{"name":"apple-mobile-web-app-capable","content":"yes"}],["$","meta",null,{"name":"apple-mobile-web-app-status-bar-style","content":"black-translucent"}],["$","meta",null,{"name":"apple-mobile-web-app-title","content":"Krog"}],["$","meta",null,{"charSet":"utf-8"}],["$","meta",null,{"name":"google","content":"notranslate"}],["$","link",null,{"rel":"apple-touch-icon","href":"/web-app-manifest-192x192.png"}],["$","link",null,{"rel":"apple-touch-icon","sizes":"512x512","href":"/web-app-manifest-512x512.png"}],["$","link",null,{"rel":"manifest","href":"/manifest.json"}],["$","link",null,{"rel":"preconnect","href":"https://fonts.googleapis.com"}],["$","link",null,{"rel":"preconnect","href":"https://fonts.gstatic.com","crossOrigin":"anonymous"}],["$","link",null,{"rel":"stylesheet","href":"https://fonts.googleapis.com/css2?family=Esteban&family=Inter:wght@400;600&family=IM+Fell+DW+Pica:ital@0;1&family=Roboto:wght@400;500&family=Raleway:wght@400;600&family=Montserrat:wght@400;600&family=Courier+Prime:ital@0;1&family=Newsreader:ital,wght@0,400;0,600;0,700;1,400;1,600&family=Poppins:wght@400;500;600;700&family=Geist:wght@400;500;600;700&family=DM+Sans:wght@400;500;600;700&family=Roboto+Mono:wght@400;500;600;700&family=Literata:ital,wght@0,400;0,600;0,700;1,400;1,600&family=STIX+Two+Text:ital,wght@0,400;0,600;0,700;1,400;1,600&family=Alegreya:ital,wght@0,400;1,400&family=JetBrains+Mono:wght@400;500;600;700&family=IBM+Plex+Mono:wght@400;500;600;700&family=Lora:ital,wght@0,400;0,600;1,400&family=Merriweather:wght@400;700&family=Nunito:wght@400;600&family=Google+Sans+Flex:opsz,wdth,wght@8..144,25..151,100..900&family=Lexend:wght@300;400;500;600;700&family=Special+Elite&display=swap"}]]}],["$","body",null,{"className":"inter_4b6d6b74-module__Dql8BG__className lora_4279c376-module__CdCtpW__variable special_elite_a27916cf-module___4cmmG__variable","suppressHydrationWarning":true,"children":[["$","$L11",null,{"id":"theme-color-init","strategy":"beforeInteractive","dangerouslySetInnerHTML":{"__html":"$12"}}],"$L13","$L14","$L15","$L16","$L17"]}]]}] 19:I[810717,["/_next/static/chunks/33t46atd3n2zd.js","/_next/static/chunks/1uoo0zstz-tka.js","/_next/static/chunks/1gmtbr53vpwcw.js","/_next/static/chunks/0hwl0k3t96f8i.js","/_next/static/chunks/07dyy4_l8-v8w.js","/_next/static/chunks/1a28s5qpw8ubo.js","/_next/static/chunks/3nx9-unhdiib7.js","/_next/static/chunks/1l3wx9j9z2da9.js","/_next/static/chunks/10prs2o5tycig.js","/_next/static/chunks/02ocli407gubc.js","/_next/static/chunks/2r--443asccvg.js","/_next/static/chunks/0rjt9gb10d05i.js","/_next/static/chunks/3lr20ikf1jmri.js","/_next/static/chunks/2mak3tsoa7kik.js","/_next/static/chunks/39kizkc9pkpb1.js","/_next/static/chunks/2uicbuh3o1h8o.js","/_next/static/chunks/1e-wsdiuv_p_u.js","/_next/static/chunks/12syjpiwnqlnq.js","/_next/static/chunks/0dnfb3y43dtv8.js"],"NextSSRPlugin"] 1a:I[540865,["/_next/static/chunks/33t46atd3n2zd.js","/_next/static/chunks/1uoo0zstz-tka.js","/_next/static/chunks/1gmtbr53vpwcw.js","/_next/static/chunks/0hwl0k3t96f8i.js","/_next/static/chunks/07dyy4_l8-v8w.js","/_next/static/chunks/1a28s5qpw8ubo.js","/_next/static/chunks/3nx9-unhdiib7.js","/_next/static/chunks/1l3wx9j9z2da9.js","/_next/static/chunks/10prs2o5tycig.js","/_next/static/chunks/02ocli407gubc.js","/_next/static/chunks/2r--443asccvg.js","/_next/static/chunks/0rjt9gb10d05i.js","/_next/static/chunks/3lr20ikf1jmri.js","/_next/static/chunks/2mak3tsoa7kik.js","/_next/static/chunks/39kizkc9pkpb1.js","/_next/static/chunks/2uicbuh3o1h8o.js","/_next/static/chunks/1e-wsdiuv_p_u.js","/_next/static/chunks/12syjpiwnqlnq.js","/_next/static/chunks/0dnfb3y43dtv8.js"],"default"] 1b:I[883661,["/_next/static/chunks/33t46atd3n2zd.js","/_next/static/chunks/1uoo0zstz-tka.js","/_next/static/chunks/1gmtbr53vpwcw.js","/_next/static/chunks/0hwl0k3t96f8i.js","/_next/static/chunks/07dyy4_l8-v8w.js","/_next/static/chunks/1a28s5qpw8ubo.js","/_next/static/chunks/3nx9-unhdiib7.js","/_next/static/chunks/1l3wx9j9z2da9.js","/_next/static/chunks/10prs2o5tycig.js","/_next/static/chunks/02ocli407gubc.js","/_next/static/chunks/2r--443asccvg.js","/_next/static/chunks/0rjt9gb10d05i.js","/_next/static/chunks/3lr20ikf1jmri.js","/_next/static/chunks/2mak3tsoa7kik.js","/_next/static/chunks/39kizkc9pkpb1.js","/_next/static/chunks/2uicbuh3o1h8o.js","/_next/static/chunks/1e-wsdiuv_p_u.js","/_next/static/chunks/12syjpiwnqlnq.js","/_next/static/chunks/0dnfb3y43dtv8.js"],"ThemeProvider"] 1c:I[193236,["/_next/static/chunks/33t46atd3n2zd.js","/_next/static/chunks/1uoo0zstz-tka.js","/_next/static/chunks/1gmtbr53vpwcw.js","/_next/static/chunks/0hwl0k3t96f8i.js","/_next/static/chunks/07dyy4_l8-v8w.js","/_next/static/chunks/1a28s5qpw8ubo.js","/_next/static/chunks/3nx9-unhdiib7.js","/_next/static/chunks/1l3wx9j9z2da9.js","/_next/static/chunks/10prs2o5tycig.js","/_next/static/chunks/02ocli407gubc.js","/_next/static/chunks/2r--443asccvg.js","/_next/static/chunks/0rjt9gb10d05i.js","/_next/static/chunks/3lr20ikf1jmri.js","/_next/static/chunks/2mak3tsoa7kik.js","/_next/static/chunks/39kizkc9pkpb1.js","/_next/static/chunks/2uicbuh3o1h8o.js","/_next/static/chunks/1e-wsdiuv_p_u.js","/_next/static/chunks/12syjpiwnqlnq.js","/_next/static/chunks/0dnfb3y43dtv8.js"],"default"] 1d:I[939286,["/_next/static/chunks/33t46atd3n2zd.js","/_next/static/chunks/1uoo0zstz-tka.js","/_next/static/chunks/1gmtbr53vpwcw.js","/_next/static/chunks/0hwl0k3t96f8i.js","/_next/static/chunks/07dyy4_l8-v8w.js","/_next/static/chunks/1a28s5qpw8ubo.js","/_next/static/chunks/3nx9-unhdiib7.js","/_next/static/chunks/1l3wx9j9z2da9.js","/_next/static/chunks/10prs2o5tycig.js","/_next/static/chunks/02ocli407gubc.js","/_next/static/chunks/2r--443asccvg.js","/_next/static/chunks/0rjt9gb10d05i.js","/_next/static/chunks/3lr20ikf1jmri.js","/_next/static/chunks/2mak3tsoa7kik.js","/_next/static/chunks/39kizkc9pkpb1.js","/_next/static/chunks/2uicbuh3o1h8o.js","/_next/static/chunks/1e-wsdiuv_p_u.js","/_next/static/chunks/12syjpiwnqlnq.js","/_next/static/chunks/0dnfb3y43dtv8.js","/_next/static/chunks/1x1n57nb1d8gg.js"],"AtomicState"] 1f:I[351642,["/_next/static/chunks/33t46atd3n2zd.js","/_next/static/chunks/1uoo0zstz-tka.js","/_next/static/chunks/1gmtbr53vpwcw.js","/_next/static/chunks/0hwl0k3t96f8i.js","/_next/static/chunks/07dyy4_l8-v8w.js","/_next/static/chunks/1a28s5qpw8ubo.js","/_next/static/chunks/3nx9-unhdiib7.js","/_next/static/chunks/1l3wx9j9z2da9.js","/_next/static/chunks/10prs2o5tycig.js","/_next/static/chunks/02ocli407gubc.js","/_next/static/chunks/2r--443asccvg.js","/_next/static/chunks/0rjt9gb10d05i.js","/_next/static/chunks/3lr20ikf1jmri.js","/_next/static/chunks/2mak3tsoa7kik.js","/_next/static/chunks/39kizkc9pkpb1.js","/_next/static/chunks/2uicbuh3o1h8o.js","/_next/static/chunks/1e-wsdiuv_p_u.js","/_next/static/chunks/12syjpiwnqlnq.js","/_next/static/chunks/0dnfb3y43dtv8.js","/_next/static/chunks/1x1n57nb1d8gg.js"],"FetchConfig"] 20:I[741965,["/_next/static/chunks/33t46atd3n2zd.js","/_next/static/chunks/1uoo0zstz-tka.js","/_next/static/chunks/1gmtbr53vpwcw.js","/_next/static/chunks/0hwl0k3t96f8i.js","/_next/static/chunks/07dyy4_l8-v8w.js","/_next/static/chunks/1a28s5qpw8ubo.js","/_next/static/chunks/3nx9-unhdiib7.js","/_next/static/chunks/1l3wx9j9z2da9.js","/_next/static/chunks/10prs2o5tycig.js","/_next/static/chunks/02ocli407gubc.js","/_next/static/chunks/2r--443asccvg.js","/_next/static/chunks/0rjt9gb10d05i.js","/_next/static/chunks/3lr20ikf1jmri.js","/_next/static/chunks/2mak3tsoa7kik.js","/_next/static/chunks/39kizkc9pkpb1.js","/_next/static/chunks/2uicbuh3o1h8o.js","/_next/static/chunks/1e-wsdiuv_p_u.js","/_next/static/chunks/12syjpiwnqlnq.js","/_next/static/chunks/0dnfb3y43dtv8.js"],"IDBCacheProvider"] 21:I[746798,["/_next/static/chunks/33t46atd3n2zd.js","/_next/static/chunks/1uoo0zstz-tka.js","/_next/static/chunks/1gmtbr53vpwcw.js","/_next/static/chunks/0hwl0k3t96f8i.js","/_next/static/chunks/07dyy4_l8-v8w.js","/_next/static/chunks/1a28s5qpw8ubo.js","/_next/static/chunks/3nx9-unhdiib7.js","/_next/static/chunks/1l3wx9j9z2da9.js","/_next/static/chunks/10prs2o5tycig.js","/_next/static/chunks/02ocli407gubc.js","/_next/static/chunks/2r--443asccvg.js","/_next/static/chunks/0rjt9gb10d05i.js","/_next/static/chunks/3lr20ikf1jmri.js","/_next/static/chunks/2mak3tsoa7kik.js","/_next/static/chunks/39kizkc9pkpb1.js","/_next/static/chunks/2uicbuh3o1h8o.js","/_next/static/chunks/1e-wsdiuv_p_u.js","/_next/static/chunks/12syjpiwnqlnq.js","/_next/static/chunks/0dnfb3y43dtv8.js"],"TooltipProvider"] 22:I[423599,["/_next/static/chunks/33t46atd3n2zd.js","/_next/static/chunks/1uoo0zstz-tka.js","/_next/static/chunks/1gmtbr53vpwcw.js","/_next/static/chunks/0hwl0k3t96f8i.js","/_next/static/chunks/07dyy4_l8-v8w.js","/_next/static/chunks/1a28s5qpw8ubo.js","/_next/static/chunks/3nx9-unhdiib7.js","/_next/static/chunks/1l3wx9j9z2da9.js","/_next/static/chunks/10prs2o5tycig.js","/_next/static/chunks/02ocli407gubc.js","/_next/static/chunks/2r--443asccvg.js","/_next/static/chunks/0rjt9gb10d05i.js","/_next/static/chunks/3lr20ikf1jmri.js","/_next/static/chunks/2mak3tsoa7kik.js","/_next/static/chunks/39kizkc9pkpb1.js","/_next/static/chunks/2uicbuh3o1h8o.js","/_next/static/chunks/1e-wsdiuv_p_u.js","/_next/static/chunks/12syjpiwnqlnq.js","/_next/static/chunks/0dnfb3y43dtv8.js"],"FontInitializer"] 23:I[254371,["/_next/static/chunks/33t46atd3n2zd.js","/_next/static/chunks/1uoo0zstz-tka.js","/_next/static/chunks/1gmtbr53vpwcw.js","/_next/static/chunks/0hwl0k3t96f8i.js","/_next/static/chunks/07dyy4_l8-v8w.js","/_next/static/chunks/1a28s5qpw8ubo.js","/_next/static/chunks/3nx9-unhdiib7.js","/_next/static/chunks/1l3wx9j9z2da9.js","/_next/static/chunks/10prs2o5tycig.js","/_next/static/chunks/02ocli407gubc.js","/_next/static/chunks/2r--443asccvg.js","/_next/static/chunks/0rjt9gb10d05i.js","/_next/static/chunks/3lr20ikf1jmri.js","/_next/static/chunks/2mak3tsoa7kik.js","/_next/static/chunks/39kizkc9pkpb1.js","/_next/static/chunks/2uicbuh3o1h8o.js","/_next/static/chunks/1e-wsdiuv_p_u.js","/_next/static/chunks/12syjpiwnqlnq.js","/_next/static/chunks/0dnfb3y43dtv8.js"],"ThemeColorInitializer"] 24:I[300523,["/_next/static/chunks/33t46atd3n2zd.js","/_next/static/chunks/1uoo0zstz-tka.js","/_next/static/chunks/1gmtbr53vpwcw.js","/_next/static/chunks/0hwl0k3t96f8i.js","/_next/static/chunks/07dyy4_l8-v8w.js","/_next/static/chunks/1a28s5qpw8ubo.js","/_next/static/chunks/3nx9-unhdiib7.js","/_next/static/chunks/1l3wx9j9z2da9.js","/_next/static/chunks/10prs2o5tycig.js","/_next/static/chunks/02ocli407gubc.js","/_next/static/chunks/2r--443asccvg.js","/_next/static/chunks/0rjt9gb10d05i.js","/_next/static/chunks/3lr20ikf1jmri.js","/_next/static/chunks/2mak3tsoa7kik.js","/_next/static/chunks/39kizkc9pkpb1.js","/_next/static/chunks/2uicbuh3o1h8o.js","/_next/static/chunks/1e-wsdiuv_p_u.js","/_next/static/chunks/12syjpiwnqlnq.js","/_next/static/chunks/0dnfb3y43dtv8.js"],"default"] 25:I[585832,["/_next/static/chunks/33t46atd3n2zd.js","/_next/static/chunks/1uoo0zstz-tka.js","/_next/static/chunks/1gmtbr53vpwcw.js","/_next/static/chunks/0hwl0k3t96f8i.js","/_next/static/chunks/07dyy4_l8-v8w.js","/_next/static/chunks/1a28s5qpw8ubo.js","/_next/static/chunks/3nx9-unhdiib7.js","/_next/static/chunks/1l3wx9j9z2da9.js","/_next/static/chunks/10prs2o5tycig.js","/_next/static/chunks/02ocli407gubc.js","/_next/static/chunks/2r--443asccvg.js","/_next/static/chunks/0rjt9gb10d05i.js","/_next/static/chunks/3lr20ikf1jmri.js","/_next/static/chunks/2mak3tsoa7kik.js","/_next/static/chunks/39kizkc9pkpb1.js","/_next/static/chunks/2uicbuh3o1h8o.js","/_next/static/chunks/1e-wsdiuv_p_u.js","/_next/static/chunks/12syjpiwnqlnq.js","/_next/static/chunks/0dnfb3y43dtv8.js"],"default"] 26:I[437224,["/_next/static/chunks/33t46atd3n2zd.js","/_next/static/chunks/1uoo0zstz-tka.js","/_next/static/chunks/1gmtbr53vpwcw.js","/_next/static/chunks/0hwl0k3t96f8i.js","/_next/static/chunks/07dyy4_l8-v8w.js","/_next/static/chunks/1a28s5qpw8ubo.js","/_next/static/chunks/3nx9-unhdiib7.js","/_next/static/chunks/1l3wx9j9z2da9.js","/_next/static/chunks/10prs2o5tycig.js","/_next/static/chunks/02ocli407gubc.js","/_next/static/chunks/2r--443asccvg.js","/_next/static/chunks/0rjt9gb10d05i.js","/_next/static/chunks/3lr20ikf1jmri.js","/_next/static/chunks/2mak3tsoa7kik.js","/_next/static/chunks/39kizkc9pkpb1.js","/_next/static/chunks/2uicbuh3o1h8o.js","/_next/static/chunks/1e-wsdiuv_p_u.js","/_next/static/chunks/12syjpiwnqlnq.js","/_next/static/chunks/0dnfb3y43dtv8.js"],"default"] 27:I[758298,["/_next/static/chunks/33t46atd3n2zd.js","/_next/static/chunks/1uoo0zstz-tka.js","/_next/static/chunks/1gmtbr53vpwcw.js","/_next/static/chunks/0hwl0k3t96f8i.js","/_next/static/chunks/07dyy4_l8-v8w.js","/_next/static/chunks/1a28s5qpw8ubo.js","/_next/static/chunks/3nx9-unhdiib7.js","/_next/static/chunks/1l3wx9j9z2da9.js","/_next/static/chunks/10prs2o5tycig.js","/_next/static/chunks/02ocli407gubc.js","/_next/static/chunks/2r--443asccvg.js","/_next/static/chunks/0rjt9gb10d05i.js","/_next/static/chunks/3lr20ikf1jmri.js","/_next/static/chunks/2mak3tsoa7kik.js","/_next/static/chunks/39kizkc9pkpb1.js","/_next/static/chunks/2uicbuh3o1h8o.js","/_next/static/chunks/1e-wsdiuv_p_u.js","/_next/static/chunks/12syjpiwnqlnq.js","/_next/static/chunks/0dnfb3y43dtv8.js","/_next/static/chunks/1o5rgnnlr-jeo.js"],"default"] 28:I[139898,["/_next/static/chunks/33t46atd3n2zd.js","/_next/static/chunks/1uoo0zstz-tka.js","/_next/static/chunks/1gmtbr53vpwcw.js","/_next/static/chunks/0hwl0k3t96f8i.js","/_next/static/chunks/07dyy4_l8-v8w.js","/_next/static/chunks/1a28s5qpw8ubo.js","/_next/static/chunks/3nx9-unhdiib7.js","/_next/static/chunks/1l3wx9j9z2da9.js","/_next/static/chunks/10prs2o5tycig.js","/_next/static/chunks/02ocli407gubc.js","/_next/static/chunks/2r--443asccvg.js","/_next/static/chunks/0rjt9gb10d05i.js","/_next/static/chunks/3lr20ikf1jmri.js","/_next/static/chunks/2mak3tsoa7kik.js","/_next/static/chunks/39kizkc9pkpb1.js","/_next/static/chunks/2uicbuh3o1h8o.js","/_next/static/chunks/1e-wsdiuv_p_u.js","/_next/static/chunks/12syjpiwnqlnq.js","/_next/static/chunks/0dnfb3y43dtv8.js","/_next/static/chunks/1x1n57nb1d8gg.js","/_next/static/chunks/0ye_k5ch2a1m0.js","/_next/static/chunks/0vndyjagwio9s.js"],"default"] 29:I[329306,["/_next/static/chunks/33t46atd3n2zd.js","/_next/static/chunks/1uoo0zstz-tka.js","/_next/static/chunks/1gmtbr53vpwcw.js","/_next/static/chunks/0hwl0k3t96f8i.js","/_next/static/chunks/07dyy4_l8-v8w.js","/_next/static/chunks/1a28s5qpw8ubo.js","/_next/static/chunks/3nx9-unhdiib7.js","/_next/static/chunks/1l3wx9j9z2da9.js","/_next/static/chunks/10prs2o5tycig.js","/_next/static/chunks/02ocli407gubc.js","/_next/static/chunks/2r--443asccvg.js","/_next/static/chunks/0rjt9gb10d05i.js","/_next/static/chunks/3lr20ikf1jmri.js","/_next/static/chunks/2mak3tsoa7kik.js","/_next/static/chunks/39kizkc9pkpb1.js","/_next/static/chunks/2uicbuh3o1h8o.js","/_next/static/chunks/1e-wsdiuv_p_u.js","/_next/static/chunks/12syjpiwnqlnq.js","/_next/static/chunks/0dnfb3y43dtv8.js","/_next/static/chunks/0nd8iowpl9fn8.js"],"default"] 2a:I[435311,["/_next/static/chunks/33t46atd3n2zd.js","/_next/static/chunks/1uoo0zstz-tka.js","/_next/static/chunks/1gmtbr53vpwcw.js","/_next/static/chunks/0hwl0k3t96f8i.js","/_next/static/chunks/07dyy4_l8-v8w.js","/_next/static/chunks/1a28s5qpw8ubo.js","/_next/static/chunks/3nx9-unhdiib7.js","/_next/static/chunks/1l3wx9j9z2da9.js","/_next/static/chunks/10prs2o5tycig.js","/_next/static/chunks/02ocli407gubc.js","/_next/static/chunks/2r--443asccvg.js","/_next/static/chunks/0rjt9gb10d05i.js","/_next/static/chunks/3lr20ikf1jmri.js","/_next/static/chunks/2mak3tsoa7kik.js","/_next/static/chunks/39kizkc9pkpb1.js","/_next/static/chunks/2uicbuh3o1h8o.js","/_next/static/chunks/1e-wsdiuv_p_u.js","/_next/static/chunks/12syjpiwnqlnq.js","/_next/static/chunks/0dnfb3y43dtv8.js"],"QuickLinksPanel"] 2b:I[331194,["/_next/static/chunks/33t46atd3n2zd.js","/_next/static/chunks/1uoo0zstz-tka.js","/_next/static/chunks/1gmtbr53vpwcw.js","/_next/static/chunks/0hwl0k3t96f8i.js","/_next/static/chunks/07dyy4_l8-v8w.js","/_next/static/chunks/1a28s5qpw8ubo.js","/_next/static/chunks/3nx9-unhdiib7.js","/_next/static/chunks/1l3wx9j9z2da9.js","/_next/static/chunks/10prs2o5tycig.js","/_next/static/chunks/02ocli407gubc.js","/_next/static/chunks/2r--443asccvg.js","/_next/static/chunks/0rjt9gb10d05i.js","/_next/static/chunks/3lr20ikf1jmri.js","/_next/static/chunks/2mak3tsoa7kik.js","/_next/static/chunks/39kizkc9pkpb1.js","/_next/static/chunks/2uicbuh3o1h8o.js","/_next/static/chunks/1e-wsdiuv_p_u.js","/_next/static/chunks/12syjpiwnqlnq.js","/_next/static/chunks/0dnfb3y43dtv8.js"],"TimerWidget"] 2c:I[377855,["/_next/static/chunks/33t46atd3n2zd.js","/_next/static/chunks/1uoo0zstz-tka.js","/_next/static/chunks/1gmtbr53vpwcw.js","/_next/static/chunks/0hwl0k3t96f8i.js","/_next/static/chunks/07dyy4_l8-v8w.js","/_next/static/chunks/1a28s5qpw8ubo.js","/_next/static/chunks/3nx9-unhdiib7.js","/_next/static/chunks/1l3wx9j9z2da9.js","/_next/static/chunks/10prs2o5tycig.js","/_next/static/chunks/02ocli407gubc.js","/_next/static/chunks/2r--443asccvg.js","/_next/static/chunks/0rjt9gb10d05i.js","/_next/static/chunks/3lr20ikf1jmri.js","/_next/static/chunks/2mak3tsoa7kik.js","/_next/static/chunks/39kizkc9pkpb1.js","/_next/static/chunks/2uicbuh3o1h8o.js","/_next/static/chunks/1e-wsdiuv_p_u.js","/_next/static/chunks/12syjpiwnqlnq.js","/_next/static/chunks/0dnfb3y43dtv8.js"],"Toaster"] 2d:I[797756,["/_next/static/chunks/33t46atd3n2zd.js","/_next/static/chunks/1uoo0zstz-tka.js","/_next/static/chunks/1gmtbr53vpwcw.js","/_next/static/chunks/0hwl0k3t96f8i.js","/_next/static/chunks/07dyy4_l8-v8w.js","/_next/static/chunks/1a28s5qpw8ubo.js","/_next/static/chunks/3nx9-unhdiib7.js","/_next/static/chunks/1l3wx9j9z2da9.js","/_next/static/chunks/10prs2o5tycig.js","/_next/static/chunks/02ocli407gubc.js","/_next/static/chunks/2r--443asccvg.js","/_next/static/chunks/0rjt9gb10d05i.js","/_next/static/chunks/3lr20ikf1jmri.js","/_next/static/chunks/2mak3tsoa7kik.js","/_next/static/chunks/39kizkc9pkpb1.js","/_next/static/chunks/2uicbuh3o1h8o.js","/_next/static/chunks/1e-wsdiuv_p_u.js","/_next/static/chunks/12syjpiwnqlnq.js","/_next/static/chunks/0dnfb3y43dtv8.js"],"default"] 2e:I[118931,["/_next/static/chunks/33t46atd3n2zd.js","/_next/static/chunks/1uoo0zstz-tka.js","/_next/static/chunks/1gmtbr53vpwcw.js","/_next/static/chunks/0hwl0k3t96f8i.js","/_next/static/chunks/07dyy4_l8-v8w.js","/_next/static/chunks/1a28s5qpw8ubo.js","/_next/static/chunks/3nx9-unhdiib7.js","/_next/static/chunks/1l3wx9j9z2da9.js","/_next/static/chunks/10prs2o5tycig.js","/_next/static/chunks/02ocli407gubc.js","/_next/static/chunks/2r--443asccvg.js","/_next/static/chunks/0rjt9gb10d05i.js","/_next/static/chunks/3lr20ikf1jmri.js","/_next/static/chunks/2mak3tsoa7kik.js","/_next/static/chunks/39kizkc9pkpb1.js","/_next/static/chunks/2uicbuh3o1h8o.js","/_next/static/chunks/1e-wsdiuv_p_u.js","/_next/static/chunks/12syjpiwnqlnq.js","/_next/static/chunks/0dnfb3y43dtv8.js"],"default"] 2f:I[437322,["/_next/static/chunks/33t46atd3n2zd.js","/_next/static/chunks/1uoo0zstz-tka.js","/_next/static/chunks/1gmtbr53vpwcw.js","/_next/static/chunks/0hwl0k3t96f8i.js","/_next/static/chunks/07dyy4_l8-v8w.js","/_next/static/chunks/1a28s5qpw8ubo.js","/_next/static/chunks/3nx9-unhdiib7.js","/_next/static/chunks/1l3wx9j9z2da9.js","/_next/static/chunks/10prs2o5tycig.js","/_next/static/chunks/02ocli407gubc.js","/_next/static/chunks/2r--443asccvg.js","/_next/static/chunks/0rjt9gb10d05i.js","/_next/static/chunks/3lr20ikf1jmri.js","/_next/static/chunks/2mak3tsoa7kik.js","/_next/static/chunks/39kizkc9pkpb1.js","/_next/static/chunks/2uicbuh3o1h8o.js","/_next/static/chunks/1e-wsdiuv_p_u.js","/_next/static/chunks/12syjpiwnqlnq.js","/_next/static/chunks/0dnfb3y43dtv8.js"],"GlobalSocketListener"] 30:I[203009,["/_next/static/chunks/33t46atd3n2zd.js","/_next/static/chunks/1uoo0zstz-tka.js","/_next/static/chunks/1gmtbr53vpwcw.js","/_next/static/chunks/0hwl0k3t96f8i.js","/_next/static/chunks/07dyy4_l8-v8w.js","/_next/static/chunks/1a28s5qpw8ubo.js","/_next/static/chunks/3nx9-unhdiib7.js","/_next/static/chunks/1l3wx9j9z2da9.js","/_next/static/chunks/10prs2o5tycig.js","/_next/static/chunks/02ocli407gubc.js","/_next/static/chunks/2r--443asccvg.js","/_next/static/chunks/0rjt9gb10d05i.js","/_next/static/chunks/3lr20ikf1jmri.js","/_next/static/chunks/2mak3tsoa7kik.js","/_next/static/chunks/39kizkc9pkpb1.js","/_next/static/chunks/2uicbuh3o1h8o.js","/_next/static/chunks/1e-wsdiuv_p_u.js","/_next/static/chunks/12syjpiwnqlnq.js","/_next/static/chunks/0dnfb3y43dtv8.js"],"OfflineSyncListener"] 18:T75a, (function() { var presets = ['inter', 'geist', 'dm-sans', 'nunito', 'poppins', 'raleway', 'montserrat', 'lexend', 'lora', 'literata', 'merriweather', 'newsreader', 'stix', 'cmu', 'pica', 'tnr', 'roboto-mono', 'jetbrains-mono', 'plex-mono', 'google-sans-flex', 'special-elite']; try { var appFontRaw = localStorage.getItem('app-font'); if (appFontRaw) { var appFont = appFontRaw; try { appFont = JSON.parse(appFontRaw); } catch(e) {} if (appFont) { document.documentElement.dataset.font = appFont; var isAppCustom = presets.indexOf(appFont) === -1; document.documentElement.dataset.fontCustom = String(isAppCustom); if (isAppCustom) { document.documentElement.style.setProperty('--custom-app-font', '"' + appFont.replace(/['"]/g, '') + '"'); } } } var matFontRaw = localStorage.getItem('material-font'); if (matFontRaw) { var matFont = matFontRaw; try { matFont = JSON.parse(matFontRaw); } catch(e) {} if (matFont) { document.documentElement.dataset.materialFont = matFont; var isMatCustom = presets.indexOf(matFont) === -1; document.documentElement.dataset.materialFontCustom = String(isMatCustom); if (isMatCustom) { document.documentElement.style.setProperty('--custom-material-font', '"' + matFont.replace(/['"]/g, '') + '"'); } } } } catch (e) {} })(); 13:["$","$L11",null,{"id":"font-init","strategy":"beforeInteractive","dangerouslySetInnerHTML":{"__html":"$18"}}] 14:["$","$L11",null,{"id":"navbar-init","strategy":"beforeInteractive","dangerouslySetInnerHTML":{"__html":"\n (function() {\n try {\n if (window.location.pathname === '/') {\n document.documentElement.dataset.navbarPosition = 'hidden';\n document.documentElement.style.setProperty('--navbar-width', '0px');\n }\n } catch (e) {}\n })();\n "}}] 15:["$","$L19",null,{"routerConfig":[{"slug":"questionImage","config":{"image":{"maxFileSize":"1MB","maxFileCount":1,"minFileCount":1,"contentDisposition":"inline"}}},{"slug":"profilePicture","config":{"image":{"maxFileSize":"1MB","maxFileCount":1,"minFileCount":1,"contentDisposition":"inline"}}},{"slug":"coverPicture","config":{"image":{"maxFileSize":"1MB","maxFileCount":1,"minFileCount":1,"contentDisposition":"inline"}}},{"slug":"materialFile","config":{"blob":{"maxFileSize":"32MB","maxFileCount":1,"minFileCount":1,"contentDisposition":"inline"}}}]}] 16:["$","$L1a",null,{}] 17:["$","$L1b",null,{"attribute":"class","defaultTheme":"light","children":["$","main",null,{"className":"min-h-screen","children":["$","$L1c",null,{"session":null,"children":["$","$L1d",null,{"value":{"server-theme":"light","flavor":"krog","app-font":"inter","material-font":"literata","app-background":"dots","desktop-markdown-editor":"codemirror","monaco-theme":"one-dark-pro","assistant-language":"auto","lang":"$@1e","language-prompted":"$undefined","navbar-position":"left","navbar-collapsed":false},"children":["$","$L1f",null,{"baseUrl":"/api","value":{"User":{},"Preferences":{}},"children":["$","$L20",null,{"children":["$","$L21",null,{"children":[["$","$L22",null,{}],["$","$L23",null,{}],["$","$L24",null,{"children":[["$","$L25",null,{}],["$","div",null,{"className":"flex-1 flex flex-col min-w-0","children":[["$","$L26",null,{}],["$","div",null,{"data-layout-container":true,"className":"w-full max-w-screen mx-auto py-8 px-3 md:px-8 overflow-x-auto flex-1","children":["$","$L3",null,{"parallelRouterKey":"children","error":"$27","errorStyles":[],"errorScripts":[["$","script","script-0",{"src":"/_next/static/chunks/1o5rgnnlr-jeo.js","async":true}]],"template":["$","$L28",null,{"children":["$","$L4",null,{}]}],"templateStyles":[],"templateScripts":[["$","script","script-0",{"src":"/_next/static/chunks/0ye_k5ch2a1m0.js","async":true}],["$","script","script-1",{"src":"/_next/static/chunks/0vndyjagwio9s.js","async":true}]],"notFound":[["$","$L29",null,{}],[]],"forbidden":"$undefined","unauthorized":"$undefined"}]}]]}]]}],["$","$L2a",null,{}],["$","$L2b",null,{}],["$","$L2c",null,{}],["$","$L2d",null,{}],["$","$L2e",null,{}],["$","$L2f",null,{}],["$","$L30",null,{}]]}]}]}]}]}]}]}] 1e:"en" d:[["$","meta","0",{"charSet":"utf-8"}],["$","meta","1",{"name":"viewport","content":"width=device-width, initial-scale=1, maximum-scale=1, user-scalable=no"}]] 31:I[27201,["/_next/static/chunks/33t46atd3n2zd.js","/_next/static/chunks/1uoo0zstz-tka.js","/_next/static/chunks/1gmtbr53vpwcw.js","/_next/static/chunks/0hwl0k3t96f8i.js","/_next/static/chunks/07dyy4_l8-v8w.js","/_next/static/chunks/1a28s5qpw8ubo.js","/_next/static/chunks/3nx9-unhdiib7.js","/_next/static/chunks/1l3wx9j9z2da9.js","/_next/static/chunks/10prs2o5tycig.js","/_next/static/chunks/02ocli407gubc.js","/_next/static/chunks/2r--443asccvg.js","/_next/static/chunks/0rjt9gb10d05i.js","/_next/static/chunks/3lr20ikf1jmri.js","/_next/static/chunks/2mak3tsoa7kik.js","/_next/static/chunks/39kizkc9pkpb1.js","/_next/static/chunks/2uicbuh3o1h8o.js","/_next/static/chunks/1e-wsdiuv_p_u.js","/_next/static/chunks/12syjpiwnqlnq.js","/_next/static/chunks/0dnfb3y43dtv8.js"],"IconMark"] a:null f:[["$","title","0",{"children":"Claude Agentic Architecture: Notes + 113 Practice Questions - Krog | Krog"}],["$","meta","1",{"name":"description","content":"Share, have fun, learn."}],["$","meta","2",{"name":"application-name","content":"Krog"}],["$","link","3",{"rel":"manifest","href":"/manifest.json","crossOrigin":"$undefined"}],["$","meta","4",{"name":"mobile-web-app-capable","content":"yes"}],["$","meta","5",{"name":"apple-mobile-web-app-title","content":"Krog"}],["$","meta","6",{"name":"apple-mobile-web-app-status-bar-style","content":"black-translucent"}],["$","meta","7",{"property":"og:title","content":"Claude Agentic Architecture: Notes + 113 Practice Questions - Krog"}],["$","meta","8",{"property":"og:description","content":"Share, have fun, learn."}],["$","meta","9",{"property":"og:image","content":"https://images.unsplash.com/photo-1456513080510-7bf3a84b82f8?w=600&h=360&fit=crop&crop=center&q=80&auto=format"}],["$","meta","10",{"property":"og:image:width","content":"1200"}],["$","meta","11",{"property":"og:image:height","content":"320"}],["$","meta","12",{"property":"og:image:alt","content":"Claude Agentic Architecture: Notes + 113 Practice Questions"}],["$","meta","13",{"name":"twitter:card","content":"summary_large_image"}],["$","meta","14",{"name":"twitter:title","content":"Claude Agentic Architecture: Notes + 113 Practice Questions - Krog"}],["$","meta","15",{"name":"twitter:description","content":"Share, have fun, learn."}],["$","meta","16",{"name":"twitter:image","content":"https://images.unsplash.com/photo-1456513080510-7bf3a84b82f8?w=600&h=360&fit=crop&crop=center&q=80&auto=format"}],["$","link","17",{"rel":"icon","href":"/favicon.svg","type":"image/svg+xml"}],["$","link","18",{"rel":"icon","href":"/web-app-manifest-192x192.png","sizes":"192x192","type":"image/png"}],["$","link","19",{"rel":"icon","href":"/web-app-manifest-512x512.png","sizes":"512x512","type":"image/png"}],["$","link","20",{"rel":"apple-touch-icon","href":"/web-app-manifest-192x192.png","sizes":"192x192","type":"image/png"}],["$","link","21",{"rel":"apple-touch-icon","href":"/web-app-manifest-512x512.png","sizes":"512x512","type":"image/png"}],["$","$L31","22",{}]] 32:T13e55, @h-underlines off # Claude Certified Architect (Foundations): Exam Map and Study Method The **Claude Certified Architect (Foundations)** examination assesses the ability to design, build, and operate production systems built on Claude — agentic loops, multi-agent orchestration, tool and MCP design, Claude Code configuration, prompt engineering, and context management. The notes below synthesise the full five-domain curriculum. The reference material is an independent community study guide and is **not affiliated with or endorsed by Anthropic**; where it states product specifics (SDK version numbers, command flags), those are presented as the guide presents them, and current behaviour should be verified against the official documentation at `docs.claude.com` and `code.claude.com`. ## How the exam is structured The exam is organised into five weighted domains and thirty task statements. The weightings indicate roughly how many questions each domain contributes, and they should drive study time allocation: Domain 1 alone is more than a quarter of the exam. | Domain | Topic | Weight | Task statements | | --- | --- | --- | --- | | 1 | Agentic Architecture & Orchestration | 27% | 7 | | 2 | Tool Design & MCP Integration | 18% | 5 | | 3 | Claude Code Configuration & Workflows | 20% | 6 | | 4 | Prompt Engineering & Structured Output | 20% | 6 | | 5 | Context Management & Reliability | 15% | 6 | The published pass mark is **720 out of 1,000 points**. Experienced Claude developers are advised to budget roughly 15–20 hours of preparation; newcomers 30–40 hours. ### The recurring "meta-patterns" the exam rewards Across all five domains, the same diagnostic instincts are tested repeatedly. Internalising these makes individual questions easier because the right answer usually expresses one of them. - **Determinism beats probability for high-stakes rules.** When a single failure causes money loss, a security breach, or a compliance violation, the answer is a code-level mechanism (hook, gate, settings file), never a stronger prompt. - **Trace failures to their origin.** In multi-agent systems, a bad output usually comes from the *input* a component received (the coordinator's decomposition or context passing), not from the component that produced the visible result. - **Prefer low-effort, high-leverage fixes first.** Better tool descriptions before routing classifiers; scoped access before full access; community MCP servers before custom builds; examples before more prose. - **Structural problems need structural fixes.** Attention dilution, context degradation, and lost-in-the-middle are *not* solved by bigger models or larger context windows. - **Self-reported confidence and sentiment are poorly calibrated** and should never directly drive automated decisions without calibration. Many questions present a tempting but wrong distractor: "add stronger instructions," "use a bigger context window," "increase the iteration cap," "average the conflicting values," "escalate on negative sentiment." Recognising the distractor family is often faster than reasoning each option from scratch. ------ # Domain 1 — Agentic Architecture & Orchestration This is the heaviest domain. It covers the execution loop that drives every agent, how multiple agents coordinate, how subagents are invoked and fed context, how workflows are enforced, how SDK hooks inject deterministic behaviour, how tasks are decomposed, and how sessions persist. ## 1.1 Agentic Loops An **agentic loop** is the core execution cycle of a Claude agent. It is a deterministic control-flow pattern — not a prompt trick, not a generic retry loop, and not a single chatbot turn. The lifecycle has four repeating steps: 1. **Send a request** to Claude via the Messages API, including the full conversation history (system prompt, prior messages, and any tool results from the previous iteration). 2. **Inspect the `stop_reason` field** in the response — the authoritative signal for what happens next. 3. If `stop_reason` is `"tool_use"`, **execute the requested tool(s), append the tool results to the conversation history**, and send the updated conversation back to Claude. The loop continues. 4. If `stop_reason` is `"end_turn"`, **the agent has finished**; present the final response. The make-or-break detail is step 3: tool results **must** be appended to history, or Claude cannot reason about the new information on the next turn. The field that authoritatively controls loop termination is `{{stop_reason}}`. It is deterministic and unambiguous, which is exactly why it exists. ### The three loop-termination anti-patterns **Three wrong ways to detect completion — all tested:** 1. **Parsing natural language** ("did Claude say 'I'm done'?"). Language is ambiguous; Claude may say "I've finished the first file" while intending to continue. 2. **Arbitrary iteration caps as the *primary* stopping mechanism** ("stop after 10 loops"). This cuts off useful work or wastes iterations. A cap is acceptable only as a safety net, never as the main control. 3. **Checking `response.content[0].type == "text"`.** Claude can return explanatory text *alongside* a `tool_use` block, so the presence of text says nothing about completion. A fourth trap is forcing `tool_choice: "any"` to stop the model returning text — this forces a tool call even when the agent is genuinely finished, creating an infinite loop. **Model-driven decision-making** means Claude itself reasons about the task and selects the appropriate tool from those available, in contrast to **pre-configured decision trees** or **fixed tool sequences** that hard-code which tool runs when. The exam favours model-driven flexibility — except where business logic requires deterministic compliance (covered in 1.4). A support agent stops mid-task on complex queries. Its loop checks `response.content[0].type == "text"` to decide it is finished. What is the fix? Claude returns text ("Let me look up your order") alongside the `tool_use` block; the content-type check sees text and wrongly concludes completion. Only `stop_reason` is authoritative. Caps address runaway loops, not premature exits. When is an iteration cap acceptable in an agentic loop?Only as a safety net (a maximum bound to prevent runaway agents) — never as the primary stopping mechanism. Termination is driven by `stop_reason`. **One-line summary:** The agentic loop sends history, reads `stop_reason`, executes-and-appends tool results on `"tool_use"`, and stops on `"end_turn"`. ## 1.2 Multi-Agent Orchestration Multi-agent systems use a **hub-and-spoke** architecture with a **coordinator** at the centre and specialised **subagents** as the spokes. ```mermaid flowchart TD C[Coordinator] -->|task + context| S1[Web Search Subagent] C -->|task + context| S2[Document Analysis Subagent] C -->|findings| S3[Synthesis Subagent] S1 -->|results| C S2 -->|results| C S3 -->|report| C ``` The cardinal rule: **all communication flows through the coordinator. Subagents never communicate directly with one another** — not for efficiency, not for convenience, ever. Centralisation buys three things the exam cares about: observability (one place to log every message), consistent error handling (uniform recovery policy), and controlled information flow. ### The critical isolation principle The single most misunderstood concept in multi-agent systems, and the one the exam exploits most: **Subagents do NOT inherit the coordinator's conversation history**, and they **do NOT share memory between invocations.** A spawned subagent starts with only what the coordinator explicitly places in its prompt. Calling the same subagent twice gives the second call no knowledge of the first. The four coordinator responsibilities tested: **dynamic subagent selection** (not every query routes through the full pipeline), **research scope partitioning** (assign distinct subtopics to minimise duplication), **iterative refinement loops** (re-delegate when synthesis has gaps), and **centralised routing**. ### The narrow decomposition failure A signature exam pattern: a report on "impact of AI on creative industries" covers only visual arts, missing music, writing, and film. Every subagent worked perfectly on what it was assigned; the **coordinator's decomposition was too narrow**. When a multi-agent output is incomplete **in scope** (entire categories missing), the root cause is almost always the coordinator's decomposition — *not* a weak search subagent, *not* a weak synthesis agent, and *not* a shortage of subagents. Adding more subagents does not help, because they receive equally narrow assignments. The fix is broader decomposition. A "renewable energy" report covers only solar and wind. Every subagent thoroughly researched its assignment. What is the root cause? Subagents research only what they are assigned. Coverage gaps trace to the coordinator's decomposition. Trace failures to their origin. **One-line summary:** Hub-and-spoke with all traffic through the coordinator; subagents are isolated; scope gaps trace to decomposition, not to the subagents. ## 1.3 Subagent Invocation and Context Passing Where 1.2 is the architecture, 1.3 is the wiring. The **Agent tool** (renamed from the **Task tool** in Claude Agent SDK v2.1.63, with `Task` retained as a backward-compatible alias) is the mechanism for spawning subagents. There is a hard configuration gate: The coordinator's `allowedTools` **must** include `"Agent"` (or `"Task"`). Without it the coordinator physically cannot spawn any subagent — this is a binary gate, not a soft preference. Each subagent is defined by an **AgentDefinition** specifying three things: a **description** (so the coordinator knows when to invoke it), a **system prompt**, and **tool restrictions** scoped to its role. ### Three rules for context passing 1. **Include complete findings from prior agents** — pass them in full; the downstream agent cannot "look them up." 2. **Use structured formats that separate content from metadata** — every finding carries both the claim and its source URL, document name, page number, confidence, etc. 3. **Specify goals, not procedures** — goal-oriented prompts let subagents adapt; step-by-step instructions constrain them. **The attribution-failure trap:** a synthesis agent produces unsourced claims. The search and analysis subagents work fine. The cause is that the coordinator passed *content without metadata*, so the synthesis agent literally had no source information to cite. The fix is structured metadata, **not** editing the synthesis prompt and **not** giving it direct tool access. **Parallel spawning:** for independent subagent tasks, emit **multiple Agent tool calls in a single coordinator response** rather than one per turn. Sequential spawning adds needless latency. **`fork_session` vs `--resume`:** `fork_session` creates **independent branches from a shared baseline** for divergent exploration (e.g., comparing two testing strategies after a shared codebase analysis). `--resume` continues one specific named session. Fork for divergence; resume for continuation. A synthesis agent's report has claims with no citations, though both research subagents return proper sources. Root cause?The coordinator stripped the structured metadata before passing content to synthesis. Pass claim + source URL + document name + page number intact. ## 1.4 Workflow Enforcement and Handoff This task draws the hard line between **prompt-based guidance** (probabilistic — works ~90–95% of the time, non-zero failure rate) and **programmatic enforcement** (deterministic — works every time). **The exam decision rule:** if a *single* failure would cause financial loss, a security breach, or a compliance violation, use **programmatic enforcement** (hooks, prerequisite gates, code checks). Reserve prompt-based guidance for low-stakes preferences (formatting, style). | Operation type | Mechanism | | --- | --- | | Financial (refunds, transfers, payments) | Programmatic enforcement | | Security (identity verification, access control) | Programmatic enforcement | | Compliance (AML, regulatory checks) | Programmatic enforcement | | Low-stakes (formatting, style, ordering) | Prompt-based guidance acceptable | A **prerequisite gate** is code that blocks a tool until a precondition is met — e.g., `process_refund` cannot run until `get_customer` has returned a verified customer ID in the session. The model cannot bypass it even if it decides to skip verification. For high-stakes scenarios the exam offers "add stronger instructions," "add few-shot examples," and "add a routing classifier" as distractors. All improve probability; none reaches 0% failure. A routing classifier handles *which agent* gets a request, not *within-agent* workflow ordering. Only a gate gives deterministic guarantees. **Subagent lifecycle hooks:** `SubagentStart` fires when a subagent is spawned (log, validate, rate-limit, check context); `SubagentStop` fires when it returns (validate output schema, strip sensitive data, log). Subagents may define their own `PreToolUse`/`PostToolUse` hooks scoped only to their own calls, and `Stop` hooks in a subagent's frontmatter auto-convert to `SubagentStop` events at runtime. **Structured handoff protocols:** when escalating to a human, remember the human **does not have the conversation transcript.** The summary must be self-contained: customer ID, conversation summary, root-cause analysis, refund amount (if any), and recommended action. **Multi-concern requests** ("return my order, update my address, and ask about loyalty points") should be **decomposed**, **investigated in parallel using shared context**, and resolved in a **single unified response** — never handled by addressing only the first item. Production data shows 8% of refunds are processed without verifying account ownership, despite a clear system-prompt instruction to verify first. Best fix? This is a financial operation; an 8% failure rate must go to 0%. Only programmatic enforcement is deterministic. Prompts and examples reduce but never eliminate the failure rate; classifiers route, they do not enforce per-agent ordering. ## 1.5 Agent SDK Hooks Hooks inject deterministic behaviour into a probabilistic system at the boundary between the model's decisions and the real world. **PostToolUse hooks** run *after* a tool executes but *before* the model sees the result — used for **data normalisation**. **Tool-call interception (pre-execution) hooks** run *before* a tool executes — used for **policy enforcement**; they can block, modify, or redirect the call so the tool never runs. **PostToolUse / normalisation:** different MCP tools return heterogeneous formats (Unix timestamps vs ISO 8601 vs `DD/MM/YYYY`; numeric status codes vs strings vs single characters). A PostToolUse hook converts everything to one consistent schema before the model reasons about it, eliminating per-iteration interpretation errors. **Pre-execution interception / policy:** intercept `process_refund` and block amounts over a threshold; intercept `transfer_funds` until an AML check passes; route discounts over 20% to a manager-approval queue. The exam's favourite hook trap: using a **PostToolUse hook to block a policy-violating action.** PostToolUse runs *after* execution — by the time it fires, the non-compliant action has already happened. To block, you must use a **pre-execution interception hook.** Always know the *direction* of the hook. The decision framework: | Requirement | Mechanism | Guarantee | | --- | --- | --- | | Must hold 100% of the time | Hooks | Deterministic | | Preferred, occasional deviation acceptable | Prompts | Probabilistic | International transfers occasionally skip the required AML check; compliance demands 100% enforcement. The current prompt works ~95%. Correct approach? 100% regulatory enforcement requires deterministic, pre-execution blocking. PostToolUse fires after the transfer already executed; prompts and examples stay probabilistic. ## 1.6 Task Decomposition Strategies Two patterns, and one failure mode. Predetermined steps in fixed order (prompt chaining). Best for **predictable, structured tasks** (code reviews, document extraction, compliance checks). Consistent, reliable, easy to debug and monitor — but cannot adapt to unexpected findings mid-run. The plan evolves as the agent learns. Best for **open-ended investigation** of unknown scope (legacy-system exploration, security audits, debugging unfamiliar code). Adapts to discovered complexity — but less predictable in time and harder to debug. Match the pattern to **task characteristics, not to what sounds more sophisticated**. Multi-file code review (steps known) → fixed pipeline. Legacy codebase exploration (dependencies emerge) → dynamic decomposition. ### The attention dilution problem A failure mode where an agent processing too many items in a single pass produces **inconsistent depth**: thorough analysis of early items, shallow analysis of later ones, and contradictory judgements (the same pattern flagged in one file but approved in another). The classic example: a 14-file review gives detailed feedback on files 1–5, moderate on 6–9, and superficial on 10–14 (missing null-pointer bugs and SQL-injection vulnerabilities), while flagging a `forEach` loop in file 3 yet approving identical code in file 11. Attention dilution is an **architectural** problem, not a model-capability problem. A bigger model, a larger context window, or a stronger prompt does **not** fix it. The fix is a **multi-pass architecture**: per-item local passes (each item gets the full attention budget) plus a separate cross-item integration pass (data flow, API consistency, cross-file pattern checks). Batching files into groups *without* a cross-file integration pass still misses cross-batch issues. A 14-file review degrades after the first few files and flags a `forEach` loop in one file but approves identical code in another. Root cause and fix? Inconsistent depth and contradictory pattern evaluation are the signatures of attention dilution, a structural problem. Multi-pass architecture is the fix; bigger models, prompts, and naive batching are not. ## 1.7 Session State and Resumption Three session-management options, each for a distinct purpose: | Scenario | Option | Why | | --- | --- | --- | | Continuing yesterday's work, no files changed | `--resume ` | Prior context still valid; full history useful | | Comparing two refactoring approaches | `fork_session` | Divergent branches from a shared baseline | | Resumed after modifying 3 of 50 files | Fresh start + summary | Stale tool results would cause contradictions | | Long session with cluttered history | Fresh start + summary | Degraded context benefits from a clean baseline | | Resumed after dependency updates | Fresh start + summary | Many files may have changed indirectly | ### The stale context problem When an agent resumes a session after code changes and reasons from **cached tool results that no longer reflect the current files**, producing contradictory advice — recommending fixes already made, or referencing code that no longer exists. The naive fix — resume and ask the agent to re-read the changed files — is insufficient: the **stale tool results remain in the conversation history** and can still influence tangential reasoning. The reliable fix is a **fresh session with a structured summary injected**, naming the specific changed files for **targeted re-analysis** (not full re-exploration of all 50 files). Using `fork_session` does not help, because a fork inherits the existing session's stale results. Best response after modifying 3 files in a 50-file codebase, when the resumed agent gives contradictory advice?Fresh session + injected summary of prior findings, specifying the 3 changed files for targeted re-analysis. Avoid plain `--resume` (stale results persist) and `fork_session` (inherits stale results). ------ # Domain 2 — Tool Design & MCP Integration This domain covers how tools are described, how failures are reported, how tools are distributed and chosen, how MCP servers are configured, and how Claude Code's built-in tools are selected. ## 2.1 Tool Interface Design Tool **descriptions are the PRIMARY mechanism** an LLM uses to select tools — not supplementary metadata. A production-grade description includes five elements: **what the tool does**, **what inputs it expects** (types, formats, constraints), **example queries it handles well**, **edge cases and limitations**, and **explicit boundaries** (when to use this tool versus a similar one). ### The misrouting problem and its fix hierarchy When two tools have minimal, overlapping descriptions (`get_customer`: "Retrieves customer information"; `lookup_order`: "Retrieves order details"), the agent misroutes queries. The fix hierarchy the exam rewards: **Expand the descriptions first.** It is low-effort and high-leverage and addresses the root cause. Reject few-shot examples (token overhead, treats symptoms), routing classifiers (over-engineered, bypasses the LLM), and tool consolidation (valid long-term, but more effort than needed as a first step). Related techniques: **tool splitting** (replace a generic `analyze_document` with `extract_data_points`, `summarize_content`, `verify_claim_against_source`); **tool renaming** for clarity; and reviewing the **system prompt** for keyword-sensitive instructions that can silently override well-written descriptions. An agent calls `get_customer` for "check my order #12345" instead of `lookup_order`; both tools have one-line descriptions. Most effective first step? Descriptions are the primary selection mechanism. Low-effort, high-leverage fixes come first: better descriptions before examples, classifiers, or consolidation. ## 2.2 Structured Error Responses When an MCP tool fails, the error response determines whether the agent recovers intelligently. The MCP protocol provides the **`isError`** flag to mark failures; generic "Operation failed" text is useless to the model. ### The four error categories | Category | Meaning | Retryable? | Recovery | | --- | --- | --- | --- | | Transient | Timeout, rate limit, service down | Yes | Retry after brief delay | | Validation | Bad input format, missing field | Yes | Fix the input and retry | | Business | Policy/limit/rule violation | **No** | Alternative path / escalate | | Permission | Access denied, bad credentials | **No** | Escalate / different credentials | Structured errors carry `errorCategory`, `isRetryable`, and a human-readable `description`. **Business errors are never resolved by retrying** — the same request always fails. ### Access failure vs valid empty result One of the most critical distinctions in the domain. An **access failure** means the tool could not reach the data source (timeout, auth failure) — the data might exist but was not checked; consider retry. A **valid empty result** means the query *executed successfully and found nothing* — that IS the answer; do **not** retry. Tools must make a successful-but-empty response (`isError: false`, `resultCount: 0`) look structurally different from a failed query. In multi-agent systems, error handling follows **local recovery with selective propagation**: subagents retry transient failures locally; only unresolved errors propagate upward, **including partial results and what was attempted**. The two anti-patterns are silent suppression (returning empty as success) and terminating the whole workflow on one failure. A customer lookup returns an empty array; the agent retries 3 times then escalates. The account simply does not exist. Root cause? The query succeeded and correctly found nothing. Without a structural distinction between "could not reach" and "found nothing," the agent wastes retries and escalates incorrectly. ## 2.3 Tool Distribution & Tool Choice The number of tools an agent has directly affects selection reliability. Giving one agent **18 tools degrades reliability**; the optimal range is **4–5 tools per agent**, scoped to its role. A synthesis agent should not hold web-search tools, or it may duplicate work it was already given. ### The `tool_choice` parameter | Mode | Behaviour | Use when | | --- | --- | --- | | `auto` (default) | Model decides whether to call a tool or return text | General conversational flexibility | | `any` | Model must call *some* tool, choosing which | Guaranteed structured output, type unknown | | `{type: tool, name: X}` | Model must call the *named* tool | Enforce a mandatory first step | ### Scoped cross-role tools Routing every minor need through the coordinator adds 2–3 round trips and can raise latency by ~40%. The fix is a **scoped cross-role tool**: give the agent a constrained capability for the common case. A synthesis agent that verifies facts often gets a scoped `verify_fact` for the ~85% simple lookups; only the ~15% complex verifications route to the coordinator. Apply **least privilege to tool design** — replace a generic `fetch_url` (fetches anything) with a constrained `load_document` that validates document URLs only. A synthesis agent returns to the coordinator for every fact check, adding ~40% latency; 85% are simple lookups. Best fix? A scoped cross-role tool handles the 85% common case locally, eliminating round-trip latency while keeping complex cases on the full pipeline. ## 2.4 MCP Server Integration MCP servers connect Claude to external systems. Configuration lives at two scopes: | File | Scope | Version-controlled? | Use for | | --- | --- | --- | --- | | `.mcp.json` (project root) | Project | Yes, shared via git | Team-wide servers (GitHub, Jira) | | `~/.claude.json` (home) | User | No, personal | Experimental/personal servers | All tools from all configured, reachable servers are discovered at connection time and available simultaneously — there is no manual activation step. Use `${VARIABLE_NAME}` **environment-variable expansion** so credentials stay out of version control while configuration is shared. **MCP resources** expose content catalogues (issue summaries, documentation hierarchies, database schemas) upfront, reducing exploratory tool calls. **Resources let agents see data; tools let agents act on it.** ### The build-vs-use decision For standard integrations (Jira, GitHub, Slack, Linear, Notion), **evaluate existing community MCP servers first.** Build a custom server only when the team has workflows community servers cannot handle, needs embedded custom business logic, or must integrate proprietary internal systems. A subtle failure: **sparse MCP tool descriptions** cause the agent to prefer richly described built-in tools (like Grep). Enhance MCP descriptions to explain capabilities, outputs, and when to prefer them over built-ins. A team needs Jira integration in Claude Code. A developer proposes building a custom MCP server. Correct first step? Community servers cover standard integrations and are the low-effort, high-leverage choice. Team-wide servers belong in `.mcp.json`, not personal config. ## 2.5 Built-in Tools Claude Code provides six built-in tools: **Read, Write, Edit, Bash, Grep, Glob**. **Grep searches file *contents*** for patterns (function callers, error strings, imports). **Glob matches file *paths*** by naming pattern (`**/*.test.tsx`, `**/config.*`). In one line: Grep finds what is *inside* files; Glob finds files by their *names*. A frequent trap is using **Glob to find function callers** (it matches paths, not contents — it will fail) or **Grep to find test files by naming pattern** (Glob is the purpose-built tool). **Edit, Read, Write:** Edit performs targeted unique-text replacements and is the **default** — fast and low-context. When Edit reports a non-unique match, the documented recovery is to **widen `old_str` with surrounding context** until it pins one location, or set **`replace_all: true`** for a global change. Read + Write (load whole file, rewrite it) is the **last resort**, not the next step. Two penalised behaviours: (1) defaulting to Read + Write for every modification (burns context tokens), and (2) jumping straight from a non-unique Edit failure to Read + Write instead of widening the anchor or using `replace_all`. **Incremental codebase understanding:** never read all files upfront (a context-budget killer). Instead: Grep to find entry points → Read to follow imports and trace flows → Grep again to trace usage → Read only what is justified by the previous step. **The deprecation scenario:** to find all callers of a deprecated function *and* their tests — **Grep** for the function name (finds callers and tests importing it directly) → **Glob** for sibling test files (e.g., `**/OrderProcessor.test.*`, catching tests that exercise it indirectly) → **Grep again** for wrapper names (transitive coverage). The sequence is Grep, then Glob, then Grep — never Glob first. Grep vs Glob in one sentence.Grep searches file contents; Glob matches file paths by name. ------ # Domain 3 — Claude Code Configuration & Workflows This domain is about configuring Claude Code: the CLAUDE.md hierarchy, skills and slash commands, path-specific rules, plan vs direct execution, iterative refinement, and CI/CD. ## 3.1 CLAUDE.md Hierarchy, Scoping, and Modular Organisation CLAUDE.md is read at three levels: | Level | Location | Shared? | Use for | | --- | --- | --- | --- | | User | `~/.claude/CLAUDE.md` | No (personal) | Personal preferences, verbosity | | Project | `.claude/CLAUDE.md` or root `CLAUDE.md` | Yes (git) | Team-wide standards | | Directory | subdirectory `CLAUDE.md` | Only in that directory | Package-specific conventions | CLAUDE.md is **not** a strict-precedence config. All discovered files are **concatenated** into context (broadest scope first, working-directory-closest last; `CLAUDE.local.md` appended after `CLAUDE.md` at each level). It is delivered as a *user message*, not the system prompt, so there is **no guarantee of strict compliance** — if two rules contradict, Claude may pick one arbitrarily. Watch for distractors claiming "more specific scope wins" or "user-level overrides project-level" — the official docs never say this. For rules that **must** hold every run (a blocked tool, a required formatter, a permission policy), do not rely on CLAUDE.md — encode them in **`settings.json`** (a strict precedence chain: managed policy > local > project > user, enforced by the client) or a **hook**. **Modular organisation:** the `@` directive (no `@import` keyword) inlines a referenced file at load time. Imports load **eagerly**, so the context Claude sees is the same size as if pasted in — splitting a file is for maintainability, not token savings. To actually shrink per-session context, use **`.claude/rules/`** with path-scoped frontmatter (see 3.3). The **`/memory`** command **reveals which memory files are loaded** — it diagnoses, it does not activate loading. ### The signature trap: the new team member Developer A's Claude Code follows API naming conventions perfectly; new Developer B gets inconsistent naming on the same repo and branch. Most likely cause? User-level config is not version-controlled or shared via git, so a fresh clone never receives it. Move the conventions to project-level `.claude/CLAUDE.md`. `/memory` only reports loaded files; it does not trigger loading. ## 3.2 Custom Slash Commands and Skills Custom commands and skills are a single unified **Skills system**. Files at `.claude/skills/` (canonical) or `.claude/commands/` (backward-compatible alias) both create a `/command`. The same project-vs-user scoping pattern applies: `.claude/` is shared via git; `~/.claude/` is personal. Skills support optional YAML frontmatter: - **`context: fork`** — runs the skill in an isolated sub-agent context so verbose output (codebase analysis, brainstorming) stays out of the main conversation. - **`allowed-tools`** — restricts which tools the skill may use (a security boundary; a read-only analysis skill should not have Write or Bash). - **`argument-hint`** — prompts the developer for required parameters when invoked without arguments. **Skills** are on-demand, task-specific workflows: their descriptions stay in context, but the full body loads only on invocation (explicit `/name`, an intent match against the `description`, or a `paths`-scoped match). **CLAUDE.md** (and `.claude/rules/`) is always-loaded, universal standards applied to every session. Do not put task-specific procedures in CLAUDE.md, and do not put always-on reference material in skills. The team wants a `/review` command for everyone who clones the repo; a developer also wants a personal `/brainstorm` skill that produces verbose analysis without cluttering the main conversation. Where and how? Team commands belong in project scope (`.claude/`); personal commands in user scope (`~/.claude/`). `context: fork` isolates verbose output to keep the main context clean. ## 3.3 Path-Specific Rules for Conditional Convention Loading Rule files in `.claude/rules/` carry YAML frontmatter with a `paths` glob field; their content loads **only when editing matching files**. ```yaml --- paths: ["**/*.test.ts", "**/*.test.tsx", "**/*.spec.ts"] --- # Test Conventions ... ``` The decisive advantage is **cross-directory coverage**: a glob like `**/*.test.tsx` matches every test file across 50+ directories. This is more maintainable than a directory-level CLAUDE.md (which would need a copy in every directory and would inevitably drift) and more token-efficient than root CLAUDE.md (which loads for every session regardless of relevance). | Scenario | Best approach | | --- | --- | | Universal standards for all code | Root CLAUDE.md | | Conventions for one package directory | Directory-level CLAUDE.md | | File-type conventions across many directories | Path-specific rules with glob patterns | | On-demand workflows | Skills in `.claude/skills/` | Both skills and `.claude/rules/` can auto-activate via a `paths` field, but they differ: **rules stay in context as background guidance** (loaded when Claude reads a matching file), shaping every edit; **skills load on-demand as task-style workflows**. For always-on convention loading for a file type, the answer is **path-specific rules**, not a skill. Test files are co-located with source files across 50+ directories, and the team wants uniform test conventions everywhere. Most maintainable approach? Glob-scoped rules give one file, one pattern, universal coverage — loaded only when editing matching files, avoiding both the maintenance burden of per-directory files and the token waste of root CLAUDE.md. ## 3.4 Plan Mode vs Direct Execution The deciding factor is **ambiguity, not difficulty.** A difficult but well-defined bug fix (clear stack trace, single function, known cause) is **direct execution**. A simple-sounding feature that could be built three ways and touches multiple modules is **plan mode**. **Plan mode** suits large-scale changes, multiple valid approaches, architectural decisions, multi-file modifications, and codebase exploration. **Direct execution** suits well-scoped, known-approach, limited changes. The common **hybrid** is **plan THEN execute**: plan to design a migration strategy across many files, then switch to direct execution to apply it consistently. The **Explore subagent** isolates verbose discovery output (file listings, dependency graphs) from the main conversation, returning summaries so implementation context stays clean. | Task | Mode | | --- | --- | | Architectural restructuring | Plan mode | | Library migration across many files | Plan mode, then direct execution | | Single-file bug fix with clear stack trace | Direct execution | | Configuration value update | Direct execution | A specific trap: choosing direct execution and switching to plan mode only *if* complexity emerges. When the requirements already state the task is complex ("restructure the monolith into microservices"), choose plan mode **upfront** — the complexity is known, not speculative. ## 3.5 Iterative Refinement Techniques A clear **technique hierarchy** for guiding Claude Code: 1. **Concrete input/output examples** — the first tool to reach for when prose is interpreted inconsistently. 2–3 before/after pairs establish a pattern the model generalises from. More prose does not fix interpretation ambiguity. 2. **Test-driven iteration** — most effective for complex transformations; share unambiguous test failures ("Expected X, got Y"). 3. **Interview pattern** — for unfamiliar domains, ask Claude to pose questions before implementing, surfacing considerations (cache invalidation, TTL, consistency) you would miss. Do not confuse the **interview pattern** (unfamiliar domain — you might miss considerations) with **concrete examples** (you know the exact transformation but the model misinterprets it). They solve different problems. **Batch vs sequential feedback:** give **all interacting fixes in one message** (error format + logging format + type definitions that depend on each other) so the model sees every constraint; fix **independent issues sequentially** to avoid confusing which feedback applies where. A developer describes a transformation in prose and Claude Code interprets it differently each run. What to try first? Inconsistent interpretation of prose is the textbook trigger for concrete examples. More prose still relies on interpretation; examples eliminate the ambiguity. ## 3.6 CI/CD Integration The single most directly tested fact in Domain 3 is the **`-p` (`--print`) flag**, which runs Claude Code in **non-interactive print mode**. Without it, a CI job hangs forever waiting for keyboard input. When a CI job hangs and logs show Claude waiting for input, the answer is **`-p`**. It is **not** `CLAUDE_HEADLESS=true` (does not exist), **not** `--batch` (does not exist), and **not** stdin redirection from `/dev/null`. **Structured output:** `--output-format json` plus `--json-schema` produce machine-parseable findings (file, line, severity, message) for inline PR comments and severity filtering. **Session context isolation:** the same session that generated code is weaker at reviewing it because it retains its own justification. Use an **independent review invocation** with no shared context. **Incremental review context:** include prior findings and instruct Claude to report only *new* or *still-unaddressed* issues, preventing duplicate comments that erode trust. **CLAUDE.md** supplies project context (testing standards, fixtures, review criteria) in CI just as interactively. **Batch vs real-time for CI:** the Message Batches API offers 50% savings but up to a 24-hour window with no latency SLA. Use real-time (synchronous) for **blocking** pre-merge checks; batch only for **latency-tolerant** overnight/weekly work. A CI script runs `claude` with a prompt but the job hangs; logs show it waiting for interactive input. Correct fix? `-p`/`--print` is the documented non-interactive mode. The other options are invented or do not address Claude Code's interactive mode. ------ # Domain 4 — Prompt Engineering & Structured Output This domain covers explicit-criteria prompts, structured output via tool use, validation-retry loops, few-shot prompting, batch processing, and multi-instance review. ## 4.1 System Prompts with Explicit Criteria The biggest production prompting mistake is **vague instructions** ("be conservative," "only report high-confidence findings," "use your best judgement") — they give no actionable decision boundary, which is exactly why they appear as distractors. The correct approach is **explicit categorical criteria**: define precisely what to flag (bugs, security vulnerabilities), what to skip (style preferences, local patterns), and the trigger for comment flags (claimed behaviour contradicts actual behaviour). Calibrate severity with **concrete code examples**, not prose. ### The false-positive trust problem A high false-positive rate in **one** category destroys trust in **all** categories — if documentation-mismatch findings are wrong 40% of the time, developers stop reading the 98%-accurate security findings too. The counterintuitive fix: **temporarily disable the high-false-positive category** to restore system-wide trust, then refine its criteria with code examples before re-enabling. **Why confidence-based filtering fails:** LLM self-reported confidence is poorly calibrated — often confident on wrong findings and uncertain on right ones. The hierarchy is **explicit criteria first, confidence-based routing second**. A review pipeline has a 40% false-positive rate on documentation findings, causing developers to ignore ALL categories, including accurate security findings. Most effective fix? Trust bleeds across categories; disabling the broken one restores trust system-wide while you fix its criteria. Vague confidence language does not improve precision. ## 4.2 Structured Output with Tool Use The reliability hierarchy: **`tool_use` with JSON schemas** (eliminates JSON *syntax* errors) over **prompt-based JSON** (can be malformed). The three `tool_choice` modes mirror 2.3: `auto` (may return text), `any` (guaranteed tool call, model picks which), and forced (`{type: tool, name: X}`). `tool_use` eliminates **syntax** errors but **not semantic** errors: line items that do not sum to the stated total, values in wrong fields, and **fabrication** of values for required fields the source lacks. Schema guarantees structure, not correctness. **Schema design defences:** make fields **optional/nullable** when the source may lack them — the primary defence against fabrication, because a required field pressures the model to invent a value. Add an **"unclear"** enum option for genuinely ambiguous cases, an **"other" + detail string** for extensibility, and put **format-normalisation rules** in the prompt. An extraction system uses `tool_use` with all fields required; testers report invented dates and amounts for documents lacking that information. Best fix? Required fields force fabrication; nullable fields allow an honest `null`. This is the structural root-cause fix. ## 4.3 Prompt Chaining and Validation-Retry Loops The **retry-with-error-feedback** pattern sends three things back to the model: the **original document**, the **failed extraction**, and the **specific validation error**. Naive retries without the specific error reproduce the same mistake. ### The retry effectiveness boundary Retries **fix** format mismatches, structural errors, misplaced values, and arithmetic slips (a missed line item). Retries **cannot** create information genuinely absent from the source. If a document lacks a department name, retrying never helps — flag for human review or return `null` (if the schema allows). The exam presents both fixable and unfixable cases and expects you to tell them apart. **Self-correction in the schema:** extract both `calculated_total` and `stated_total` so a discrepancy is automatically flagged; add `conflict_detected` booleans when a document contradicts itself; add `detected_pattern` fields to findings so dismissal patterns reveal which prompts need refinement. Remember the cross-link to 4.2: **schema syntax errors** are eliminated by `tool_use`; **semantic validation errors** require validation logic and retry loops. Document A's line items sum to £450 but the stated total is £500; Document B is missing the `department` field entirely. Correct retry strategy? A is fixable (a missed line item is a structural error); B is unfixable (the data does not exist in the source). Identify fixability before retrying. ## 4.4 Few-Shot Prompting Few-shot examples are the **most effective technique for consistency** — reached for before more instructions, confidence thresholds, or temperature changes. Three deployment triggers: detailed instructions still produce **inconsistent formatting**; the model makes **inconsistent judgement calls** on ambiguous cases; or extraction returns **empty fields for information that exists** (often embedded in narrative rather than tables). Construction rules: use **2–4 targeted examples** covering the failing scenarios, and **each example must show reasoning**, not just input/output — so the model learns the decision *principle* and generalises, rather than literal pattern-matching. A secondary benefit is **hallucination reduction** in extraction across varied document structures, and reduced false positives by showing both what to flag and what to ignore. | Problem | Correct technique | | --- | --- | | Inconsistent output formatting | Few-shot examples | | Malformed JSON | `tool_use` with JSON schemas | | Fabricated values for missing fields | Optional/nullable schema fields | | Wrong tool selection | Better tool descriptions, then few-shot | | Misses info in narrative text | Few-shot examples of narrative extraction | | Sum does not match total | Validation-retry loop | Extraction works on structured tables but returns empty fields when the same data appears in narrative paragraphs; detailed instructions already specify all fields. Try first? The data exists but in an unexpected structure — the classic trigger for examples covering the failing structure. More instructions will not help. ## 4.5 Batch Processing and Prompt Optimisation The **Message Batches API** facts: **50% cost savings**, **up to a 24-hour processing window**, **no latency SLA**, **no multi-turn tool calling within a single batch request**, and **`custom_id`** fields to correlate requests with responses. **The matching rule (most tested):** synchronous API for **blocking** workflows where someone waits (pre-merge checks, real-time review); batch API for **latency-tolerant** workflows consumed later (overnight debt reports, weekly audits, nightly test generation). Do not switch everything to batch for cost savings. **SLA calculation:** for a 30-hour SLA, subtract the 24-hour maximum window to get 6 hours of buffer; submit batches every 4–6 hours so a fresh batch is always in flight. **Failure handling:** identify failures by `custom_id`, resubmit **only** failures with modifications (chunk oversized docs, simplify prompts, add format-specific examples), and **refine prompts on a 5–10 document sample before** the full batch — a 90% first-pass rate means far fewer resubmissions than 60%. A blocking pre-merge check and an overnight technical-debt report; the manager proposes moving both to batch for 50% savings. How to evaluate? Pre-merge checks are blocking — the 24-hour window with no SLA is unacceptable. Only latency-tolerant work should use batch. ## 4.6 Multi-Instance Review and Output Validation A model reviewing its own output in the **same session** retains its generation reasoning and tends to **confirm rather than challenge** its decisions. Use an **independent instance** with no prior context for review. For large reviews, use a **multi-pass architecture** (the same fix as attention dilution in 1.6): **per-file local passes** for consistent depth, then a **separate cross-file integration pass** for data flow, API consistency, and contradictions. A larger context window does not fix attention *quality*. **Confidence-based routing:** the model can self-report confidence per finding; route high-confidence to developers and low-confidence to human review — but **calibrate thresholds against labelled validation sets** first. Raw, uncalibrated confidence must not drive automated decisions. A 14-file PR gets inconsistent review depth, missed bugs, and contradictory findings (a pattern flagged in one file, approved in another). Restructure how? This is attention dilution; the structural fix is per-file passes plus an integration pass. A larger window holds more text but does not equalise attention. ------ # Domain 5 — Context Management & Reliability The lowest-weighted domain, but it ties the others together: protecting critical facts across long sessions, calibrating escalation, propagating errors, surviving long codebase exploration, allocating human review, and preserving provenance. ## 5.1 Context Window Management ### The progressive summarisation trap Summarising old turns to free tokens systematically destroys the most critical information — **numerical values, dates, percentages, identifiers, and customer-stated expectations**. "I'd like a refund of $247.83 for order #8891 placed on March 3rd" becomes "Customer wants a refund for a recent order," losing exactly the facts needed to act. **The single most important pattern in this domain: the persistent case facts block.** Extract transactional facts (amounts, dates, order numbers, statuses) into a structured block included in **every prompt, outside the summarised history, never summarised.** For multi-issue sessions, give each issue its own structured entry to prevent cross-contamination. **The "lost in the middle" effect:** models attend reliably to the beginning and end of long inputs but may miss the middle. The fix is **structural** — place a key-findings summary at the **start** of aggregated inputs and use explicit section headers — not a prompt reminder to "pay attention." **Tool result trimming:** trim verbose tool outputs (a 40-field order lookup) to the few relevant fields **before** they enter history (ideally in a PostToolUse hook), or they consume tokens on every later turn. The Claude API is **stateless** — every request must include the complete conversation history; selective truncation breaks coherence, so use case-facts blocks and summarisation instead. A multi-issue support session, with history summarised between turns, starts referring to "your recent refund request" instead of the specific $247.83 / order #8891 refund. Best fix? Summarisation reliably destroys specific values; the case-facts block keeps them outside the summarised portion permanently. ## 5.2 Escalation & Ambiguity Resolution **Three valid escalation triggers:** (1) the customer **explicitly requests a human** (honour immediately — no "let me try first"); (2) a **policy gap** (the policy is silent — distinct from a policy *violation*, which has a documented "no"); (3) **inability to make meaningful progress** after actually attempting resolution. **Two unreliable triggers** the exam flags: **sentiment-based escalation** (frustration does not correlate with case complexity — a furious customer with a late delivery is easy; a calm customer with a policy gap needs a human) and **self-reported confidence scores** (poorly calibrated — confident on hard cases, uncertain on easy ones). **The frustration nuance:** a frustrated customer with a *resolvable* issue → acknowledge and resolve; if they then reiterate wanting a human → escalate; if they ask for a human from the start → escalate immediately. **Ambiguous customer matching:** when multiple records match, **ask for additional identifiers** — never select the most recent or most active record (privacy and wrong-account risks). The proportionate first improvement is **explicit escalation criteria with few-shot examples** in the system prompt, before adding classifiers or sentiment analysis. A support agent hits only 55% first-contact resolution: it escalates simple damage replacements while attempting complex policy-exception requests. Most effective improvement? The miscalibration is exactly the self-confidence failure mode. Explicit criteria with examples is the proportionate, prompt-level first fix; sentiment and raw confidence are the two unreliable triggers. ## 5.3 Error Propagation in Multi-Agent Systems A failed subagent must return **structured error context** with four elements: **failure type** (transient/validation/business/permission), **what was attempted** (query, parameters, target), **partial results** gathered before failure, and **potential alternative approaches**. **Two anti-patterns.** **Silent suppression** — returning empty results marked as success — is the worst: the coordinator believes the search ran and found nothing, never retries, and ships a report with invisible gaps (or tells a customer they have no account when the lookup system was down). **Workflow termination** — killing the whole pipeline on one failure — discards the work of subagents that succeeded. The same **access failure vs valid empty result** distinction from 2.2 applies: an access failure (timeout, connection error) did not execute and may merit retry; a valid empty result executed successfully and IS the answer. **Coverage annotations** make gaps explicit ("Section on geothermal is limited due to unavailable journal access") rather than silently omitting topics. Subagents do **local recovery** for transient failures before propagating, always including what was attempted and any partial results. A web search subagent times out on a complex topic. Which failure-reporting approach best enables intelligent coordinator recovery? Structured context lets the coordinator choose to retry, try an alternative, proceed with partial results, or escalate. Generic errors and the two anti-patterns destroy that ability. ## 5.4 Codebase Exploration & Context Degradation A failure mode in extended exploration where the model loses grip on earlier specific findings as verbose discovery output accumulates — it begins citing "typical patterns" instead of the actual `OrderRepository` class at `src/repos/order.ts` it discovered earlier. It is **not** a token-limit problem; a bigger window still fills with verbose output. Mitigations: - **Scratchpad files** — the agent writes key findings (specific class names, file paths, dependency chains) to a file and reads it for later questions, persisting knowledge outside the conversation context. Use from the start of any extended session. - **Subagent delegation** — primarily for **context isolation** (not just parallelisation): subagents explore verbosely in their own context and return structured summaries, keeping the main context clean. - **Summary injection between phases** — inject Phase 1 findings into Phase 2 subagent prompts to avoid cold-start re-exploration. - **`/compact`** — summarises the conversation to free space; use **proactively**, not only at the context limit. - **Crash recovery via structured state manifests** — export explored paths, key findings, current phase, and next steps to a manifest the coordinator reloads on resume. An exploration agent starts referencing "typical repository patterns" instead of the specific class names and dependency chains it discovered earlier. Most effective mitigation? Context degradation is an attention-quality problem; persisting specific findings outside the context (scratchpad) is the fix. A larger window still degrades. ## 5.5 Human Review & Confidence Calibration ### The aggregate metrics trap A reported **97% aggregate accuracy can hide catastrophic per-segment failure** — 99.5% on standard invoices but 60% on handwritten receipts, 72% on poor-OCR scans, 45% on international formats. The aggregate looks excellent because high-volume segments dominate. **Always validate accuracy by document type AND field segment before automating.** **Stratified random sampling** must include **high-confidence (automated) extractions**, not just low-confidence ones — only sampling automated outputs catches a novel error pattern in them. **Field-level confidence must be calibrated** against labelled ground-truth data (0.90 confidence on dates may mean 94% accuracy; on amounts only 82%) before driving routing. **Reviewer capacity** should be allocated to the **highest-uncertainty items first**, with a **dynamically reordered** queue — never spread evenly. The correct sequence: measure by type and field → calibrate confidence → set calibrated thresholds → add stratified sampling → only then reduce human review on validated segments. An extraction system reports 97% overall accuracy; the team wants to automate all extractions above 95% model confidence. Critical risk? Aggregates hide segment failures, and raw confidence is uncalibrated. Validate by type and field and calibrate before automating. ## 5.6 Information Provenance & Multi-Source Synthesis Every finding must carry a **structured claim-source mapping**: claim, source URL, document name, relevant excerpt, and publication date. **Attribution dies during summarisation.** When the synthesis agent compresses and paraphrases, claim-source mappings vanish unless it is explicitly instructed to preserve and merge them through every step — the most common loss point is the synthesis step. The final output must trace each claim to a specific source. **Conflict handling:** when two credible sources disagree (12% vs 8% growth), do **not** arbitrarily select one, average them, or pick the more authoritative publisher. **Annotate both values with full attribution and dates**, and let the consumer decide. **Temporal awareness:** different publication/data-collection dates often explain different numbers — an apparent contradiction (8% in 2023, 12% in 2024) may be a *trend*. Require dates in all structured outputs. **Content-appropriate rendering:** financial data → **tables**; news and current events → **prose**; technical findings → **structured lists**. Forcing everything into one format degrades comprehension. Document analysis should **complete its work with conflicts intact and annotated**, leaving resolution to the coordinator or consumer. Two credible sources report different growth: Source A 12% (2023 data), Source B 8% (2024 data); the synthesis agent currently picks the more recent. Correct approach? Arbitrary selection or averaging destroys information and creates false certainty; different dates may explain the difference as a trend. Preserve both with attribution. ------ # Synthesis: Cross-Domain Connections and Final Review The domains are deliberately interlocking. The same handful of principles resurface under different labels, and recognising the connections turns scattered facts into a coherent model. ## Concepts that span multiple domains | Concept | Where it appears | | --- | --- | | Deterministic enforcement vs probabilistic prompts | 1.4 (gates), 1.5 (hooks), 3.1 (`settings.json` vs CLAUDE.md) | | Attention dilution → multi-pass architecture | 1.6 (task decomposition), 4.6 (multi-instance review) | | Subagent context isolation / explicit passing | 1.2, 1.3, 5.1 (upstream optimisation), 5.4 (delegation) | | Access failure vs valid empty result | 2.2 (tool errors), 5.3 (error propagation) | | Structured metadata / provenance preservation | 1.3 (context passing), 5.6 (claim-source mappings) | | `tool_choice` (auto / any / forced) | 2.3 (distribution), 4.2 (structured output) | | Independent review instance (no shared context) | 3.6 (CI), 4.6 (review architecture) | | Few-shot examples for consistency | 3.5 (refinement), 4.4 (few-shot), 5.2 (escalation criteria) | | Calibrated vs raw confidence | 4.1, 4.6, 5.2, 5.5 | | Trace failures to their origin | 1.2, 1.3, 5.3 | ## The trap families, consolidated ```mermaid flowchart TD Q[A scenario asks for a fix] --> H{Single failure causes loss?} H -->|Yes| D[Deterministic: gate, hook, settings.json] H -->|No| A{Inconsistent depth across items?} A -->|Yes| M[Multi-pass architecture] A -->|No| C{Inconsistent interpretation or format?} C -->|Yes| F[Concrete few-shot examples] C -->|No| O[Trace to origin: coordinator input or context passing] ``` Memorise the **distractor families** that are almost always wrong: "use a bigger context window / better model" for structural problems; "add stronger instructions / few-shot examples" for high-stakes enforcement; "increase the iteration cap" for premature termination; "average / pick the newest" for conflicting sources; "escalate on sentiment" or "route on raw confidence" for calibration questions; and "build a custom MCP server" before evaluating community servers. ## Final recall flashcards Authoritative signal for agentic loop control.The `stop_reason` field: continue on `"tool_use"`, terminate on `"end_turn"`. The hard gate for spawning subagents.The coordinator's `allowedTools` must include `"Agent"` (or `"Task"`). PostToolUse vs pre-execution interception hooks.PostToolUse normalises data *after* execution; pre-execution interception blocks/redirects *before* execution. Use interception to block policy violations. Optimal number of tools per agent.4–5, scoped to the agent's role. Project vs user scope in Claude Code config.`.claude/` (project) is shared via git; `~/.claude/` (user) is personal and not shared. Applies to CLAUDE.md, skills/commands, and rules. The single most directly tested Domain 3 fact.The `-p` (`--print`) flag for non-interactive Claude Code in CI. What `tool_use` with a JSON schema does NOT prevent.Semantic errors: wrong sums, misplaced values, and fabricated values for missing fields. Make fields nullable and add validation. The single most important context-management pattern.The persistent case-facts block — transactional facts in every prompt, outside summarised history, never summarised. Message Batches API key facts.50% savings, up to 24-hour window, no latency SLA, no multi-turn tool calling, `custom_id` for correlation. Batch only latency-tolerant workflows. Handling two conflicting credible sources in synthesis.Annotate both values with full attribution and publication dates; never average or arbitrarily pick one. Different dates may reveal a trend, not a contradiction. Across the exam, which fix is correct for a multi-agent report that is missing an entire topic category, when every subagent worked correctly? Scope gaps trace to decomposition. More subagents inherit the same narrow assignments; the synthesis agent only synthesises what it received. Trace the failure to its origin — the coordinator. **One-line summary of the whole exam:** prefer deterministic enforcement for high-stakes rules, structural fixes for structural problems, explicit criteria and examples over vague instructions, isolated-and-explicitly-fed subagents over assumed shared state, and always trace a failure to the input that caused it. ------ # Concepts The tool could not reach the data source (timeout, auth error, service down); the query did not execute. May warrant a retry. Contrast with a valid empty result. The mechanism for spawning subagents (renamed from the Task tool in SDK v2.1.63; `Task` remains a backward-compatible alias). Must appear in the coordinator's `allowedTools`, or no subagent can be spawned. The four-step execution cycle that drives a Claude agent: send the request with full history, read `stop_reason`, execute and append tool results on `tool_use`, and finish on `end_turn`. Reporting one overall accuracy figure that hides catastrophic failure rates on specific document types or fields. Validate accuracy by document type AND field segment before automating. Inconsistent analysis depth when too many items are processed in a single pass — thorough early, shallow later, with contradictory judgements. Fixed with multi-pass architecture, not a bigger model or larger context window. A structured block of transactional facts (amounts, dates, order numbers, statuses) included in every prompt and never summarised. The core defence against the progressive summarisation trap. In extended exploration, the model loses grip on specific earlier findings as verbose output accumulates, drifting to generic "typical patterns." An attention-quality problem, not a token-limit one. The central hub in hub-and-spoke orchestration. It decomposes tasks, selects subagents, passes context explicitly, aggregates results, handles errors, and routes all communication. An explicit note in synthesis output marking which areas are well-supported, limited, or unavailable, instead of silently omitting topics that could not be researched. Creates independent branches from a shared baseline for divergent exploration. Not for continuation (use `--resume`) and not for stale context (use a fresh start with summary injection), since a fork inherits existing stale results. The required multi-agent architecture: a coordinator at the centre, specialised subagents as the spokes, and no direct subagent-to-subagent communication — all traffic flows through the coordinator. The MCP flag signalling that a tool execution failed, so the model reasons about recovery rather than treating error text as a successful result. Distinct from a valid empty result. Models attend reliably to the start and end of long inputs but may miss content in the middle. Mitigated structurally — key findings first, explicit section headers — not by prompt reminders. A content catalogue (database schemas, issue lists, documentation hierarchies) exposed upfront to reduce exploratory tool calls. Resources let agents see data; tools let agents act on it. Claude reasoning about which tool to call based on context, in contrast to hard-coded decision trees or fixed tool sequences. Favoured for flexibility, except where business logic requires deterministic compliance. Runs after a tool executes but before the model sees the result. Used for data normalisation (converting heterogeneous formats to one consistent schema). Cannot block actions, since the action has already occurred. Runs before a tool executes and can block, modify, or redirect the call. The correct hook for enforcing policy, because the tool never runs if the hook blocks it. Code that blocks a tool until a precondition is met (e.g., a refund blocked until identity is verified). Deterministic enforcement that the model cannot bypass. Summarising conversation history systematically destroys numerical values, dates, and identifiers. Avoided with a persistent case-facts block held outside the summarised history. A file where an agent records key findings (class names, file paths, dependency chains) and re-reads them, persisting knowledge outside the conversation context and immune to context degradation. Reasoning from cached tool results that no longer reflect current files after edits, producing contradictory advice. Fixed with a fresh session plus an injected summary and targeted re-analysis. The authoritative loop-control signal: `tool_use` means continue the loop, `end_turn` means the agent has finished. The only reliable termination signal. Sampling each segment (document type, confidence band, field) for human verification — including high-confidence automated outputs, the only way to catch novel error patterns in automated extractions. The parameter controlling tool calls: `auto` (model may return text), `any` (must call some tool, model picks which), and forced `{type: tool, name: X}` (must call the named tool). A query that executed successfully and found nothing (`isError: false`, `resultCount: 0`). This is the answer and should not be retried. Killing the entire pipeline when a single subagent fails, discarding completed work from other subagents. An anti-pattern; use structured error propagation and partial results instead. 33:T16de,# Source & Reference Links The lesson links below point to an **independent community study guide** that is not affiliated with or endorsed by Anthropic. Treat the official Anthropic documentation as the authority; verify any product specifics (SDK versions, flags, commands) at the current docs, since they change. ## Official Anthropic documentation These are the primary sources. If a link has moved, start from `https://docs.claude.com` or `https://code.claude.com`. - Claude Agent SDK — Overview: https://platform.claude.com/docs/en/agent-sdk/overview - Claude Agent SDK — Hooks: https://platform.claude.com/docs/en/agent-sdk/hooks - Messages API reference: https://docs.anthropic.com/en/api/messages - Tool use (function calling): https://platform.claude.com/docs/en/build-with-claude/tool-use - Prompt engineering overview: https://platform.claude.com/docs/en/build-with-claude/prompt-engineering/overview - Message Batches API: https://platform.claude.com/docs/en/build-with-claude/batch-processing - Claude Code docs: https://code.claude.com/docs/en - Claude Code — Memory & CLAUDE.md: https://code.claude.com/docs/en/memory - Claude Code — Skills & commands: https://code.claude.com/docs/en/skills - Claude Code — MCP configuration: https://code.claude.com/docs/en/mcp - Claude Code — Built-in tools: https://code.claude.com/docs/en/tools - Claude Code — CLI reference: https://code.claude.com/docs/en/cli-reference - Building effective agents: https://www.anthropic.com/research/building-effective-agents - How we built our multi-agent research system: https://www.anthropic.com/engineering/built-multi-agent-research-system - Model Context Protocol spec: https://modelcontextprotocol.io / https://github.com/modelcontextprotocol ## Community guide — lessons by domain ### Domain 1 — Agentic Architecture & Orchestration (27%) - 1.1 Agentic Loops: https://claudecertificationguide.com/learn/1-agentic-architecture/1-1-agentic-loops - 1.2 Multi-Agent Orchestration: https://claudecertificationguide.com/learn/1-agentic-architecture/1-2-orchestration-patterns - 1.3 Subagent Invocation & Context Passing: https://claudecertificationguide.com/learn/1-agentic-architecture/1-3-guardrails-safety - 1.4 Workflow Enforcement & Handoff: https://claudecertificationguide.com/learn/1-agentic-architecture/1-4-claude-agent-sdk - 1.5 Agent SDK Hooks: https://claudecertificationguide.com/learn/1-agentic-architecture/1-5-multi-agent-systems - 1.6 Task Decomposition Strategies: https://claudecertificationguide.com/learn/1-agentic-architecture/1-6-human-in-the-loop - 1.7 Session State & Resumption: https://claudecertificationguide.com/learn/1-agentic-architecture/1-7-error-recovery-resilience ### Domain 2 — Tool Design & MCP Integration (18%) - 2.1 Tool Interface Design: https://claudecertificationguide.com/learn/2-tool-design-mcp/2-1-tool-schema-design - 2.2 Structured Error Responses: https://claudecertificationguide.com/learn/2-tool-design-mcp/2-2-mcp-server-implementation - 2.3 Tool Distribution & Tool Choice: https://claudecertificationguide.com/learn/2-tool-design-mcp/2-3-mcp-client-integration - 2.4 MCP Server Integration: https://claudecertificationguide.com/learn/2-tool-design-mcp/2-4-tool-error-handling - 2.5 Built-in Tools: https://claudecertificationguide.com/learn/2-tool-design-mcp/2-5-tool-selection-routing ### Domain 3 — Claude Code Configuration & Workflows (20%) - 3.1 CLAUDE.md Hierarchy & Scoping: https://claudecertificationguide.com/learn/3-claude-code-config/3-1-configuration-settings - 3.2 Custom Slash Commands & Skills: https://claudecertificationguide.com/learn/3-claude-code-config/3-2-claude-md-files - 3.3 Path-Specific Rules: https://claudecertificationguide.com/learn/3-claude-code-config/3-3-hooks-automation - 3.4 Plan Mode vs Direct Execution: https://claudecertificationguide.com/learn/3-claude-code-config/3-4-permissions-security - 3.5 Iterative Refinement Techniques: https://claudecertificationguide.com/learn/3-claude-code-config/3-5-slash-commands-mcp - 3.6 CI/CD Integration: https://claudecertificationguide.com/learn/3-claude-code-config/3-6-cicd-integration ### Domain 4 — Prompt Engineering & Structured Output (20%) - 4.1 System Prompts with Explicit Criteria: https://claudecertificationguide.com/learn/4-prompt-engineering/4-1-system-prompts - 4.2 Structured Output with Tool Use: https://claudecertificationguide.com/learn/4-prompt-engineering/4-2-structured-output - 4.3 Prompt Chaining & Validation-Retry: https://claudecertificationguide.com/learn/4-prompt-engineering/4-3-prompt-chaining - 4.4 Few-Shot Prompting: https://claudecertificationguide.com/learn/4-prompt-engineering/4-4-few-shot-examples - 4.5 Batch Processing & Prompt Optimisation: https://claudecertificationguide.com/learn/4-prompt-engineering/4-5-prompt-optimisation - 4.6 Multi-Instance Review & Output Validation: https://claudecertificationguide.com/learn/4-prompt-engineering/4-6-output-validation ### Domain 5 — Context Management & Reliability (15%) - 5.1 Context Window Management: https://claudecertificationguide.com/learn/5-context-management/5-1-context-window-management - 5.2 Escalation & Ambiguity Resolution: https://claudecertificationguide.com/learn/5-context-management/5-2-prompt-caching - 5.3 Error Propagation in Multi-Agent Systems: https://claudecertificationguide.com/learn/5-context-management/5-3-long-conversations - 5.4 Codebase Exploration & Context Degradation: https://claudecertificationguide.com/learn/5-context-management/5-4-rate-limiting-quotas - 5.5 Human Review & Confidence Calibration: https://claudecertificationguide.com/learn/5-context-management/5-5-monitoring-observability - 5.6 Information Provenance & Multi-Source Synthesis: https://claudecertificationguide.com/learn/5-context-management/5-6-production-reliability34:Te87,# Cheat Sheet — Rapid Recall ## Agentic loop - Control signal: **`stop_reason`** — continue on `tool_use`, stop on `end_turn`. - Always **append tool results to history** before the next request. - Iteration cap = safety net only, never primary control. ## Multi-agent orchestration - Architecture: **hub-and-spoke**; all traffic through the **coordinator**. - Subagents are **isolated**: no inherited history, no shared memory. - Scope gaps → coordinator **decomposition**, not the subagents. - Spawn tool: **Agent** (alias `Task`); coordinator `allowedTools` must include it. - Independent tasks → **parallel** (multiple Agent calls in one response). - **fork** = divergent branches; **--resume** = continue one session; **fresh start + summary** = stale context. ## Enforcement - Single failure causes loss → **deterministic** (hook / gate / settings.json). - Low-stakes preference → prompt is fine. - **Pre-execution hook** = block; **PostToolUse hook** = normalise after. ## Tools & MCP | Item | Value | | --- | --- | | Tools per agent | 4–5, scoped to role | | Grep | searches file **contents** | | Glob | matches file **paths** | | Edit fail (non-unique) | widen anchor or `replace_all`; Read+Write last | | `.mcp.json` | project, version-controlled | | `~/.claude.json` | user, personal | | Build vs use | community MCP first | ## Error categories | Category | Retryable | Recovery | | --- | --- | --- | | Transient | Yes | retry after delay | | Validation | Yes | fix input, retry | | Business | **No** | alternative / escalate | | Permission | **No** | escalate / new creds | - **Access failure** (could not reach) ≠ **valid empty result** (`isError:false`, do not retry). ## tool_choice - `auto` = may return text · `any` = must call some tool · `{type:tool,name:X}` = forced specific tool. ## Claude Code config - `.claude/` = project (git) · `~/.claude/` = user (personal). - CLAUDE.md = **concatenated**, not strict precedence; must-honour rules → `settings.json` or hook. - `/memory` **reveals** loaded files (does not load). - `@path` imports load **eagerly**; shrink context with **`.claude/rules/`** (path-scoped). - Skill isolation: **`context: fork`**. - CI non-interactive: **`-p` / `--print`**. ## Plan vs direct - Decide on **ambiguity, not difficulty**. - Library migration = plan **then** execute. ## Refinement hierarchy 1. Concrete examples (inconsistent interpretation) 2. Test-driven iteration (complex transforms) 3. Interview pattern (unfamiliar domain) - Interacting fixes → **one message**; independent → sequential. ## Structured output - `tool_use` kills **syntax** errors, not **semantic** ones. - Missing data → **nullable** fields (stops fabrication). ## Batch API - **50%** savings · up to **24h** · **no SLA** · **no multi-turn tools** · `custom_id` to correlate. - Blocking work → synchronous. Refine prompts on a sample first. ## Context management - **Persistent case-facts block** (never summarised) beats progressive summarisation. - Lost-in-the-middle → key findings **first** + section headers. - **Trim tool results** before they enter history. ## Escalation - Valid: explicit human request · policy **gap** · cannot progress. - Unreliable: **sentiment** · **raw confidence**. - Ambiguous match → ask for an identifier, never pick a record. ## Human review - 97% aggregate can hide 45% per-segment → validate by **type AND field**. - **Calibrate** confidence vs labelled sets; sample **high-confidence** too. ## Provenance - Mapping = claim + URL + doc name + excerpt + **date**. - Conflicting sources → **annotate both**, never average/pick. - Render: financial → tables, news → prose, technical → lists.35:Td1f,# Distractor Families — and Why They're Wrong These options *sound* reasonable and appear repeatedly as wrong answers. Learn to spot the family, then check whether the scenario actually calls for the real fix. ## "Use a bigger model / larger context window" Almost always wrong for **structural** problems. Attention dilution needs multi-pass architecture; context degradation needs scratchpad files and delegation; uneven review needs per-file passes. More capacity does not equalise attention or stop verbose output from accumulating. ## "Add stronger / more detailed instructions" Wrong for **high-stakes enforcement** (financial, security, compliance) — prompts stay probabilistic; use a gate or hook. Also wrong for **inconsistent interpretation** — the fix there is concrete examples, not more prose. ## "Add few-shot examples" (in the wrong place) Right for consistency and ambiguous judgement; **wrong** as the first fix for tool **misrouting** (expand descriptions first) and **wrong** for 100% compliance (still probabilistic). ## "Raise / add an iteration cap" Wrong for **premature termination** — caps address runaway loops, not early exits. The fix is checking `stop_reason` correctly. ## "Implement a routing classifier" Over-engineered as a *first* step for tool selection (better descriptions first) and irrelevant for **within-agent** workflow ordering (that needs a gate, not routing). ## "Build a custom MCP server" Wrong as the first move for **standard integrations** (Jira, GitHub, Slack) — evaluate community servers first. ## "Average the values" / "pick the most recent/authoritative source" Wrong for **conflicting sources** — annotate both with attribution and dates; different dates may reveal a trend, not a contradiction. ## "Escalate on negative sentiment" Wrong escalation trigger — frustration does not correlate with case complexity. ## "Use the model's confidence score directly" Wrong without **calibration** against labelled sets — raw confidence is poorly calibrated, often confident on hard cases. ## "Retry the empty result" / "raise the retry limit" Wrong when the result is a **valid empty result** (query succeeded, found nothing). Retrying repeats the same answer. Also wrong for **business errors**, which never resolve via retry. ## "Use a PostToolUse hook to block it" Wrong direction — PostToolUse runs *after* execution. Blocking needs a **pre-execution** interception hook. ## "Default to Read + Write" Wrong for routine edits and wrong as the immediate step after a non-unique Edit failure — widen the anchor or use `replace_all` first. ## "Put it in `~/.claude/...`" for team-wide config Wrong scope — user-level is personal and not shared via git. Team-wide config goes in `.claude/` (project). ## "Switch everything to the Batch API for savings" Wrong for **blocking** workflows (pre-merge checks) — no latency SLA, up to 24h. Batch is for latency-tolerant work only. ## "Make all schema fields required" Wrong — required fields pressure the model to **fabricate** missing values. Use nullable fields so it can return `null`. **The meta-rule:** trace the failure to its origin, match the fix to whether the problem is structural or probabilistic, and prefer the lowest-effort fix that actually addresses the root cause. 5:["$","$L1f",null,{"value":{"GET /exams/6a1a0c49ca96069990f7e8bc":{"user":{"id":"679481b34625e2e86cd008d3","username":"danybeltran"},"exam":{"questions":[{"id":"d55897cc-eea1-42a8-9a07-1c7e46eca245","original_quiz_id":"6a1a0c49ca96069990f7e8bc","created":"$D2026-06-03T18:39:05.747Z","description":"An engineer builds an agent loop and needs a reliable way to know when Claude has finished its work versus when it wants to call another tool.","questionContent":"Authoritative signal for agentic loop termination","questionImage":null,"type":"single","options":[{"id":"A","text":"The presence of a text block in `response.content`, since a fully finished agent always emits a closing natural-language message to the user"},{"id":"B","text":"The `stop_reason` field, continuing on `tool_use` and stopping on `end_turn`"},{"id":"C","text":"A natural-language check for phrases such as \"task complete\", \"all done\", or \"I have finished\" appearing anywhere in the response text"},{"id":"D","text":"A fixed cap of 10 iterations, after which the loop assumes the work has been completed and exits cleanly"}],"correctOptions":["B"],"matches":[],"match_pairs":[],"explanation":"## Why `stop_reason`\nThe `stop_reason` field is the only deterministic, unambiguous signal for loop control. It returns `tool_use` when Claude wants to call tools (loop continues) and `end_turn` when Claude is finished (loop stops).\nThe other options are the three classic anti-patterns:\n- **Text presence** fails because Claude can return explanatory text alongside a `tool_use` block.\n- **Natural-language parsing** is ambiguous (\"I finished the first file\" may mean more work follows).\n- **Iteration caps** are acceptable only as a safety net, never as the primary control.","helpful_links":[],"tags":[]},{"id":"7cb0bf04-7170-4ac9-87c1-ce4a66d2d1cc","original_quiz_id":"6a1a0c49ca96069990f7e8bc","created":"$D2026-06-03T18:39:05.747Z","description":"A customer support agent works for simple queries but stops mid-task on complex ones. Its loop checks `if response.content[0].type == \"text\"` to decide it is done. Users report incomplete answers.","questionContent":"Premature termination bug in a support agent","questionImage":null,"type":"single","options":[{"id":"A","text":"Add an iteration cap of 15 so complex queries have enough loops to reach a natural completion before the check fires"},{"id":"B","text":"Replace the content check with a `stop_reason` check"},{"id":"C","text":"Parse the assistant text for completion phrases like \"I have finished\" or \"that completes your request\" before terminating the loop"},{"id":"D","text":"Set `tool_choice` to `any` so Claude always calls a tool instead of returning a text block in the first position"}],"correctOptions":["B"],"matches":[],"match_pairs":[],"explanation":"Claude returns text such as \"Let me look up your order\" alongside the `tool_use` block. The code sees text in position `[0]` and wrongly concludes completion. The fix is to continue when `stop_reason == \"tool_use\"` and stop on `end_turn`. Caps address runaway loops, not premature exits; `tool_choice: any` forces a tool call even when genuinely done, causing an infinite loop.","helpful_links":[],"tags":[]},{"id":"066f84a2-a4aa-4917-b49d-b29ddfa4c52d","original_quiz_id":"6a1a0c49ca96069990f7e8bc","created":"$D2026-06-03T18:39:05.747Z","description":"When `stop_reason` is `tool_use`, the developer executes the requested tool. What must happen next for the loop to function correctly?","questionContent":"The critical step inside the agentic loop","questionImage":null,"type":"single","options":[{"id":"A","text":"Discard the tool result and re-send the original prompt so the model re-derives its plan from a clean slate each iteration"},{"id":"B","text":"Append the tool result to the conversation history before the next request"},{"id":"C","text":"Summarise the tool result into a single concise sentence and append that summary instead of the raw output"},{"id":"D","text":"Store the result in an external database and pass only a reference ID back to the model on the next turn"}],"correctOptions":["B"],"matches":[],"match_pairs":[],"explanation":"The tool result must be appended to the conversation history as a new message. Without it, Claude cannot reason about the new information on the next iteration — it would have no idea what the tool returned. Summarising or storing externally breaks the model's ability to reason about the raw result.","helpful_links":[],"tags":[]},{"id":"fafcd56c-b1f5-4562-aaab-2a708b92e4c1","original_quiz_id":"6a1a0c49ca96069990f7e8bc","created":"$D2026-06-03T18:39:05.747Z","description":"","questionContent":"Cardinal communication rule in hub-and-spoke orchestration","questionImage":null,"type":"single","options":[{"id":"A","text":"Subagents communicate directly with one another to reduce coordinator load and cut down on round-trip latency"},{"id":"B","text":"All communication flows through the coordinator"},{"id":"C","text":"Subagents share a global memory store that any agent can read from or write to during execution"},{"id":"D","text":"The coordinator routes only error conditions, while successful data is exchanged peer-to-peer between subagents"}],"correctOptions":["B"],"matches":[],"match_pairs":[],"explanation":"In hub-and-spoke architecture, all inter-subagent communication routes through the coordinator — never directly between subagents, regardless of perceived efficiency. Centralisation provides observability (one place to log every message), consistent error handling, and controlled information flow. Direct subagent communication breaks all three.","helpful_links":[],"tags":[]},{"id":"2e84f570-35c7-4219-a082-254295cc3e41","original_quiz_id":"6a1a0c49ca96069990f7e8bc","created":"$D2026-06-03T18:39:05.747Z","description":"A coordinator spawns a synthesis subagent after running a web-search subagent. The synthesis subagent produces output that ignores the search results.","questionContent":"Subagent context isolation","questionImage":null,"type":"single","options":[{"id":"A","text":"Subagents automatically inherit the coordinator's full conversation history, including every prior tool result and message"},{"id":"B","text":"Subagents start only with what the coordinator explicitly includes in their prompt"},{"id":"C","text":"Subagents share a memory cache across invocations, so later calls can read what earlier calls produced"},{"id":"D","text":"Subagents can query the coordinator's context on demand whenever they need additional background information"}],"correctOptions":["B"],"matches":[],"match_pairs":[],"explanation":"Subagents have completely isolated context. They do not inherit the coordinator's history and do not share memory between invocations. Every piece of information a subagent needs must be explicitly included in its prompt by the coordinator. If the synthesis agent ignored search results, the coordinator failed to pass them.","helpful_links":[],"tags":[]},{"id":"eef2f4e0-689b-4cc5-a2c8-0ef33d0c0ba6","original_quiz_id":"6a1a0c49ca96069990f7e8bc","created":"$D2026-06-03T18:39:05.747Z","description":"A multi-agent system produces a report on \"renewable energy technologies\" covering only solar and wind. The web-search subagent returned relevant results for every query it received, and the synthesis subagent accurately combined everything it was given.","questionContent":"Root cause of a narrow-coverage research report","questionImage":null,"type":"single","options":[{"id":"A","text":"The web-search subagent used queries that were too narrow and therefore never surfaced sources on the other energy types"},{"id":"B","text":"The synthesis subagent failed to detect the missing coverage and did not request additional research to fill the gaps"},{"id":"C","text":"The coordinator decomposed the topic into only solar and wind subtopics"},{"id":"D","text":"The document-analysis subagent lacked access to the source corpus covering geothermal, tidal, biomass, and fusion"}],"correctOptions":["C"],"matches":[],"match_pairs":[],"explanation":"## Trace to the origin\nSubagents research only what they are assigned. If geothermal, tidal, biomass, and fusion were never assigned, no subagent could cover them. The gap is in scope, not depth, so the cause is the coordinator's decomposition. Adding more subagents would not help — they would receive equally narrow assignments. The fix is broader decomposition.","helpful_links":[],"tags":[]},{"id":"7111b9f4-b917-4800-9f61-b52293c17d78","original_quiz_id":"6a1a0c49ca96069990f7e8bc","created":"$D2026-06-03T18:39:05.747Z","description":"A coordinator is defined but cannot spawn any subagents at runtime, no matter how the subagents are configured.","questionContent":"The Agent tool and its configuration gate","questionImage":null,"type":"single","options":[{"id":"A","text":"The subagents are missing a `description` field, so the coordinator cannot determine when each one should be invoked"},{"id":"B","text":"The coordinator's `allowedTools` does not include `Agent` (or `Task`)"},{"id":"C","text":"The coordinator has no system prompt, so it never reasons about delegating work to subagents in the first place"},{"id":"D","text":"The subagents need their own `allowedTools` set to `any` before the coordinator is permitted to call them"}],"correctOptions":["B"],"matches":[],"match_pairs":[],"explanation":"The Agent tool (renamed from Task in SDK v2.1.63, with `Task` kept as a backward-compatible alias) is the mechanism for spawning subagents. The coordinator's `allowedTools` must include `Agent` or `Task`. This is a binary gate — without it the coordinator physically cannot invoke any subagent.","helpful_links":[],"tags":[]},{"id":"42758cb0-fc75-4b80-aed5-5d4cfd12f51b","original_quiz_id":"6a1a0c49ca96069990f7e8bc","created":"$D2026-06-03T18:39:05.747Z","description":"A synthesis agent produces a report where several claims have no source attribution. The web-search subagent correctly returns URLs and titles; the document-analysis subagent correctly returns page references. Both subagents are verified working.","questionContent":"Attribution failure in multi-agent synthesis","questionImage":null,"type":"single","options":[{"id":"A","text":"The synthesis agent's system prompt lacks an explicit instruction to cite the source URL and page number for every claim it makes"},{"id":"B","text":"The coordinator passes content to synthesis without structured metadata (URLs, document names, page numbers)"},{"id":"C","text":"The web-search subagent returns results in a format that the synthesis agent is unable to parse into citations"},{"id":"D","text":"The synthesis agent should be given direct web-search access so it can re-locate and verify each source itself"}],"correctOptions":["B"],"matches":[],"match_pairs":[],"explanation":"The synthesis agent can only cite sources it has been given. If the coordinator strips metadata and passes only claim text, the synthesis agent literally has no source information to include. The fix is to pass structured metadata (claim + source URL + document name + page number) — not to edit the synthesis prompt or give it tool access.","helpful_links":[],"tags":[]},{"id":"3746a3c6-3fa7-4b67-85b9-b7ee71c63a25","original_quiz_id":"6a1a0c49ca96069990f7e8bc","created":"$D2026-06-03T18:39:05.747Z","description":"A coordinator must invoke a web-search subagent and a document-analysis subagent for independent tasks. The current code invokes them one per turn across separate responses.","questionContent":"Parallel vs sequential subagent spawning","questionImage":null,"type":"single","options":[{"id":"A","text":"Keep them sequential for easier debugging, since interleaving their output makes failures harder to trace"},{"id":"B","text":"Emit multiple Agent tool calls in a single coordinator response"},{"id":"C","text":"Merge the two into a single subagent that holds both toolsets and performs search and analysis together"},{"id":"D","text":"Route both through a third broker subagent that coordinates their execution order and timing"}],"correctOptions":["B"],"matches":[],"match_pairs":[],"explanation":"Independent tasks should be spawned in parallel by emitting multiple Agent tool calls in a single coordinator response. Sequential invocation across separate turns adds unnecessary latency, because there is no reason to wait for one independent task to finish before starting the other.","helpful_links":[],"tags":[]},{"id":"4a7a39b6-c072-48c0-8b8c-449ea08ac732","original_quiz_id":"6a1a0c49ca96069990f7e8bc","created":"$D2026-06-03T18:39:05.747Z","description":"","questionContent":"fork_session versus --resume","questionImage":null,"type":"single","options":[{"id":"A","text":"Both continue the same conversation history, differing only in how the session name is supplied"},{"id":"B","text":"fork_session creates independent branches; --resume continues one named session"},{"id":"C","text":"fork_session is meant for crash recovery, while --resume is the mechanism for exploring divergent approaches"},{"id":"D","text":"They are interchangeable aliases for the same underlying session-continuation behaviour"}],"correctOptions":["B"],"matches":[],"match_pairs":[],"explanation":"`fork_session` creates independent branches from a shared baseline for divergent exploration (e.g., comparing two refactoring strategies after a shared analysis). `--resume` continues one specific named session from where it left off. Fork for divergence; resume for continuation.","helpful_links":[],"tags":[]},{"id":"a350eb39-7262-4a53-aaa9-6954ba894455","original_quiz_id":"6a1a0c49ca96069990f7e8bc","created":"$D2026-06-03T18:39:05.747Z","description":"","questionContent":"Prompt-based guidance versus programmatic enforcement","questionImage":null,"type":"single","options":[{"id":"A","text":"Both are deterministic, but prompts are cheaper to implement and therefore preferred for most rules"},{"id":"B","text":"Prompt-based guidance is probabilistic; programmatic enforcement is deterministic"},{"id":"C","text":"Programmatic enforcement is probabilistic because hooks can fail, while prompts are deterministic instructions"},{"id":"D","text":"Both are probabilistic and exhibit roughly similar failure rates in production workloads"}],"correctOptions":["B"],"matches":[],"match_pairs":[],"explanation":"Prompt-based guidance works most of the time (~90–95%) but has a non-zero failure rate because the model is probabilistic. Programmatic enforcement (hooks, prerequisite gates, code checks) works every time regardless of what the model decides. The exam rule: a single failure causing loss → enforcement.","helpful_links":[],"tags":[]},{"id":"1f76c87c-5fc3-4416-80a3-7464aa6001e6","original_quiz_id":"6a1a0c49ca96069990f7e8bc","created":"$D2026-06-03T18:39:05.747Z","description":"Production data shows 8% of refunds are processed without verifying account ownership, occasionally on the wrong account. The system prompt already states \"always verify customer identity before processing refunds.\"","questionContent":"The 8% refund verification failure","questionImage":null,"type":"single","options":[{"id":"A","text":"Add stronger, more emphatic instructions to the system prompt, including warnings about the financial consequences of skipping verification"},{"id":"B","text":"Implement a prerequisite gate blocking `process_refund` until `get_customer` returns a verified ID"},{"id":"C","text":"Add few-shot examples demonstrating the correct verify-then-refund sequence across several customer scenarios"},{"id":"D","text":"Add a routing classifier that detects refund requests and sends them to a dedicated verification-first pipeline"}],"correctOptions":["B"],"matches":[],"match_pairs":[],"explanation":"This is a financial operation; the 8% failure rate must reach 0%. Only programmatic enforcement is deterministic. Stronger prompts and few-shot examples reduce but never eliminate failures. A routing classifier handles which agent gets a request, not within-agent workflow ordering.","helpful_links":[],"tags":[]},{"id":"99ad263f-c9cd-4ea4-b4c6-214002bb3878","original_quiz_id":"6a1a0c49ca96069990f7e8bc","created":"$D2026-06-03T18:39:05.747Z","description":"","questionContent":"Which scenario tolerates prompt-based guidance","questionImage":null,"type":"single","options":[{"id":"A","text":"Anti-money-laundering checks that must run before any international funds transfer is executed"},{"id":"B","text":"Identity verification that must occur before a customer refund is processed"},{"id":"C","text":"Output formatting in markdown"},{"id":"D","text":"Manager approval that must be obtained before applying large promotional discounts"}],"correctOptions":["C"],"matches":[],"match_pairs":[],"explanation":"Prompt-based guidance is acceptable only for low-stakes preferences where occasional deviation is not a business risk — like markdown formatting. AML checks, identity verification, and approval thresholds are compliance/financial/security operations requiring deterministic enforcement.","helpful_links":[],"tags":[]},{"id":"592d9e96-56dd-4420-a1d0-3f642b48ecb8","original_quiz_id":"6a1a0c49ca96069990f7e8bc","created":"$D2026-06-03T18:39:05.747Z","description":"An agent escalates a complex refund to a human. The handoff omits the customer ID and recommended action.","questionContent":"Self-contained handoff summaries","questionImage":null,"type":"single","options":[{"id":"A","text":"The human agent can simply scroll back through the full conversation transcript to recover the missing fields"},{"id":"B","text":"The human agent has no transcript access, so the summary must be self-contained"},{"id":"C","text":"The customer ID is automatically attached to every escalation by the platform, so omitting it is harmless"},{"id":"D","text":"The recommended action is optional for human agents, who are expected to form their own judgement independently"}],"correctOptions":["B"],"matches":[],"match_pairs":[],"explanation":"The human agent does not have the conversation transcript. The handoff summary is the only information they receive, so it must include customer ID, conversation summary, root-cause analysis, refund amount, and recommended action. Omitting fields forces the customer to repeat everything.","helpful_links":[],"tags":[]},{"id":"875d72fd-96ab-493a-87e6-b6613ed69201","original_quiz_id":"6a1a0c49ca96069990f7e8bc","created":"$D2026-06-03T18:39:05.747Z","description":"","questionContent":"Direction of a PostToolUse hook","questionImage":null,"type":"single","options":[{"id":"A","text":"It runs before a tool executes, allowing it to inspect and block policy-violating calls in advance"},{"id":"B","text":"It runs after a tool executes, transforming the result before the model sees it"},{"id":"C","text":"It runs in parallel with the tool so that normalisation and execution complete at the same time"},{"id":"D","text":"It replaces the tool entirely, substituting its own logic for the original tool implementation"}],"correctOptions":["B"],"matches":[],"match_pairs":[],"explanation":"PostToolUse hooks run after a tool executes but before the model processes the result — ideal for data normalisation (e.g., converting Unix timestamps and numeric status codes to consistent formats). To block an action you need a pre-execution interception hook instead.","helpful_links":[],"tags":[]},{"id":"d9dbd247-5407-4dc5-adc9-d2dd4f1de7a8","original_quiz_id":"6a1a0c49ca96069990f7e8bc","created":"$D2026-06-03T18:39:05.747Z","description":"An engineer wants to block international transfers that skip the AML check, and proposes a PostToolUse hook that flags such transfers for review.","questionContent":"Wrong hook for blocking a policy-violating action","questionImage":null,"type":"single","options":[{"id":"A","text":"Use a PostToolUse hook to flag the completed transfer and queue it for a manual compliance review afterward"},{"id":"B","text":"Use a pre-execution interception hook to block `transfer_funds` until AML passes"},{"id":"C","text":"Increase the model temperature so it produces more varied reasoning about when AML checks are required"},{"id":"D","text":"Add more detailed AML instructions to the system prompt so the model is less likely to skip the check"}],"correctOptions":["B"],"matches":[],"match_pairs":[],"explanation":"A PostToolUse hook fires after execution — by then the non-compliant transfer has already happened. To prevent the action, use a pre-execution interception hook that blocks the call before the tool runs. Always know the hook's direction.","helpful_links":[],"tags":[]},{"id":"048466fb-d4da-4d8a-a6f1-55375494fe5a","original_quiz_id":"6a1a0c49ca96069990f7e8bc","created":"$D2026-06-03T18:39:05.747Z","description":"Three MCP tools return dates as Unix timestamps, ISO 8601 strings, and `DD/MM/YYYY`, with status as numeric codes, English strings, and single characters respectively. The model occasionally misreads them.","questionContent":"Heterogeneous tool data formats","questionImage":null,"type":"single","options":[{"id":"A","text":"Tell the model to interpret each format carefully and to double-check ambiguous dates against the surrounding context"},{"id":"B","text":"Add a PostToolUse hook that normalises all results to one consistent schema"},{"id":"C","text":"Use a pre-execution interception hook to validate the format of each request before the tool runs"},{"id":"D","text":"Switch to a larger-context model that can hold more format examples and disambiguate them reliably"}],"correctOptions":["B"],"matches":[],"match_pairs":[],"explanation":"Heterogeneous formats force per-iteration interpretation, which is error-prone. A PostToolUse hook normalises all dates to ISO 8601 and all status codes to human-readable strings before the model sees them, eliminating interpretation errors. Relying on the model to normalise is unreliable.","helpful_links":[],"tags":[]},{"id":"91d27390-7dcd-4bcb-80cc-32a25b00204d","original_quiz_id":"6a1a0c49ca96069990f7e8bc","created":"$D2026-06-03T18:39:05.747Z","description":"","questionContent":"SubagentStop hook behaviour","questionImage":null,"type":"single","options":[{"id":"A","text":"It fires at the moment a subagent is spawned, before that subagent begins any execution"},{"id":"B","text":"It fires when a subagent finishes and returns results, allowing validation or transformation"},{"id":"C","text":"It blocks subagent creation whenever the coordinator has not supplied the required context"},{"id":"D","text":"It only writes log entries and is never permitted to modify the subagent's returned output"}],"correctOptions":["B"],"matches":[],"match_pairs":[],"explanation":"`SubagentStop` fires when a subagent finishes execution and returns its results to the coordinator. It lets you intercept, validate, or transform subagent output before the coordinator processes it (e.g., schema validation, stripping sensitive data). `SubagentStart` is the spawn-time counterpart, and `Stop` hooks in a subagent's frontmatter auto-convert to `SubagentStop` events.","helpful_links":[],"tags":[]},{"id":"f8b2fa07-107f-4261-a2da-8366dd0b54dd","original_quiz_id":"6a1a0c49ca96069990f7e8bc","created":"$D2026-06-03T18:39:05.747Z","description":"An agent must explore an unfamiliar legacy codebase where dependencies and issues emerge only during investigation.","questionContent":"Fixed pipeline versus dynamic decomposition","questionImage":null,"type":"single","options":[{"id":"A","text":"A fixed sequential pipeline, because predetermined steps give the most predictable and repeatable results"},{"id":"B","text":"Dynamic adaptive decomposition because scope is unknown"},{"id":"C","text":"A single-pass review that loads all files at once and analyses them together"},{"id":"D","text":"A routing classifier that selects which predefined pipeline step to run for each file"}],"correctOptions":["B"],"matches":[],"match_pairs":[],"explanation":"Open-ended investigation of unknown scope requires dynamic adaptive decomposition, where the plan evolves as the agent discovers dependencies and issues. Fixed pipelines suit predictable, structured tasks (code reviews, document extraction) but cannot adapt to unexpected findings.","helpful_links":[],"tags":[]},{"id":"f847fcc0-cc84-434b-a288-d65ac774e313","original_quiz_id":"6a1a0c49ca96069990f7e8bc","created":"$D2026-06-03T18:39:05.747Z","description":"A review agent gives detailed feedback for the first 5 files but misses obvious bugs in files 10–14, and flags a `forEach` loop in one file while approving identical code in another.","questionContent":"The 14-file review degradation","questionImage":null,"type":"single","options":[{"id":"A","text":"Upgrade to a model with a larger context window so all 14 files fit comfortably with attention to spare"},{"id":"B","text":"Split into per-file local passes plus a separate cross-file integration pass"},{"id":"C","text":"Add a stronger prompt instructing the model to review every file with equal thoroughness and consistency"},{"id":"D","text":"Process the files in sequential batches of 5 without adding any dedicated cross-file integration pass"}],"correctOptions":["B"],"matches":[],"match_pairs":[],"explanation":"## Attention dilution\nInconsistent depth and contradictory pattern evaluation are the signatures of attention dilution — an architectural problem, not a model-capability one. A larger context window does not equalise attention. The fix is a multi-pass architecture: per-file local passes (each file gets full attention) plus a cross-file integration pass. Batching without an integration pass still misses cross-batch issues.","helpful_links":[],"tags":[]},{"id":"dfd2d02d-776f-4821-8775-65e821a9dc15","original_quiz_id":"6a1a0c49ca96069990f7e8bc","created":"$D2026-06-03T18:39:05.747Z","description":"A developer modifies 3 files in a 50-file codebase and resumes a Claude Code session. The agent recommends fixes already made and references code that no longer exists.","questionContent":"The stale context problem","questionImage":null,"type":"single","options":[{"id":"A","text":"Re-explore the entire 50-file codebase from scratch so the agent rebuilds a fully accurate model of every file"},{"id":"B","text":"Start a fresh session with an injected summary, naming the 3 changed files for targeted re-analysis"},{"id":"C","text":"Resume the existing session and ask the agent to re-read only the 3 modified files to refresh its understanding"},{"id":"D","text":"Use fork_session to branch the existing session and incorporate the file changes on the new branch"}],"correctOptions":["B"],"matches":[],"match_pairs":[],"explanation":"Resuming preserves stale tool results in history, which still influence reasoning even after re-reading files. Full re-exploration is wasteful. `fork_session` inherits the stale results. The reliable fix is a fresh session with a structured summary, specifying the changed files for targeted re-analysis.","helpful_links":[],"tags":[]},{"id":"3d00663b-09cf-4abb-90e2-edfe5ad03fbe","original_quiz_id":"6a1a0c49ca96069990f7e8bc","created":"$D2026-06-03T18:39:05.747Z","description":"A customer writes: \"I want to return my order, update my shipping address, and ask about my loyalty points.\"","questionContent":"Multi-concern request handling","questionImage":null,"type":"single","options":[{"id":"A","text":"Handle only the first concern thoroughly and ask the customer to resubmit the remaining two separately"},{"id":"B","text":"Decompose into distinct items, investigate in parallel with shared context, and respond once with a unified resolution"},{"id":"C","text":"Open three separate conversations so each concern is tracked and resolved in its own isolated thread"},{"id":"D","text":"Escalate the request to a human immediately because compound multi-concern messages exceed agent scope"}],"correctOptions":["B"],"matches":[],"match_pairs":[],"explanation":"Compound requests should be decomposed into distinct items (return, address update, loyalty inquiry), investigated in parallel using shared account context, and resolved in a single unified response. Handling only the first item or splitting into sequential conversations is the wrong approach.","helpful_links":[],"tags":[]},{"id":"027523c6-22ad-4697-a387-f27e5c5c4dae","original_quiz_id":"6a1a0c49ca96069990f7e8bc","created":"$D2026-06-03T18:39:05.747Z","description":"","questionContent":"When to fork a session","questionImage":null,"type":"single","options":[{"id":"A","text":"To continue yesterday's work in a codebase where no files have changed since the last session"},{"id":"B","text":"To compare two independent refactoring approaches from a shared baseline"},{"id":"C","text":"To recover from stale context after several files were modified outside the session"},{"id":"D","text":"To free up the context window when a long session has accumulated too much verbose history"}],"correctOptions":["B"],"matches":[],"match_pairs":[],"explanation":"`fork_session` creates independent branches from a shared analysis baseline for divergent exploration — comparing alternatives that should not see each other's results. Continuation uses `--resume`; stale context after file changes needs a fresh start with summary injection; freeing context uses `/compact`.","helpful_links":[],"tags":[]},{"id":"3de67a43-4a57-4b76-80a4-6e2bfed10725","original_quiz_id":"6a1a0c49ca96069990f7e8bc","created":"$D2026-06-03T18:39:05.747Z","description":"","questionContent":"Iterative refinement loops in a coordinator","questionImage":null,"type":"single","options":[{"id":"A","text":"Single-shot delegation is sufficient for all research tasks, since subagents return complete results on the first pass"},{"id":"B","text":"The coordinator evaluates synthesis output for gaps and re-delegates with targeted queries until coverage is sufficient"},{"id":"C","text":"The synthesis agent re-delegates work to itself whenever it notices its own output is incomplete"},{"id":"D","text":"Subagents decide independently when to re-run and re-submit their results without coordinator involvement"}],"correctOptions":["B"],"matches":[],"match_pairs":[],"explanation":"A coordinator is not a one-shot dispatcher. It evaluates aggregated output for coverage gaps and, if incomplete, re-delegates to search and analysis subagents with targeted queries, re-invoking synthesis until coverage is sufficient. This iterative cycle distinguishes a coordinator from a simple dispatcher.","helpful_links":[],"tags":[]},{"id":"0b094b30-540c-4243-a18a-058aafc5cb34","original_quiz_id":"6a1a0c49ca96069990f7e8bc","created":"$D2026-06-03T18:39:05.747Z","description":"","questionContent":"Primary mechanism for LLM tool selection","questionImage":null,"type":"single","options":[{"id":"A","text":"The order in which the tools appear in the tools array passed to the model"},{"id":"B","text":"The tool descriptions"},{"id":"C","text":"The number of parameters each tool declares in its input schema"},{"id":"D","text":"The tool names alone, interpreted by the model without any accompanying description"}],"correctOptions":["B"],"matches":[],"match_pairs":[],"explanation":"Tool descriptions are the primary mechanism an LLM uses to select tools — not metadata, ordering, or names alone. When descriptions are minimal and overlapping, the model lacks context to differentiate tools, causing misrouting. A production-grade description states purpose, inputs, example queries, edge cases, and explicit boundaries.","helpful_links":[],"tags":[]},{"id":"a4e74e46-3d5c-4383-8fae-b3794f20eab5","original_quiz_id":"6a1a0c49ca96069990f7e8bc","created":"$D2026-06-03T18:39:05.747Z","description":"An agent calls `get_customer` for \"check my order #12345\" instead of `lookup_order`. Both have one-line descriptions (\"Retrieves customer information\" / \"Retrieves order details\").","questionContent":"First fix for tool misrouting","questionImage":null,"type":"single","options":[{"id":"A","text":"Add 5–8 few-shot examples to the system prompt demonstrating correct tool selection for order-related queries"},{"id":"B","text":"Expand each description with input formats, example queries, edge cases, and boundaries"},{"id":"C","text":"Implement a keyword-based routing classifier that inspects the user message and pre-selects the tool before the model runs"},{"id":"D","text":"Consolidate both tools into a single `lookup_entity` tool that internally decides which backend to query"}],"correctOptions":["B"],"matches":[],"match_pairs":[],"explanation":"Descriptions are the root cause and the lowest-effort, highest-leverage fix. Few-shot examples add token overhead and treat symptoms; a routing classifier is over-engineered and bypasses the LLM; consolidation is more effort than needed as a first step. Expand descriptions first.","helpful_links":[],"tags":[]},{"id":"9f81d744-a024-4f9a-8c40-1ea360654e9f","original_quiz_id":"6a1a0c49ca96069990f7e8bc","created":"$D2026-06-03T18:39:05.747Z","description":"A generic `analyze_document` tool (\"Analyses a document and returns results\") causes ambiguous selection.","questionContent":"Tool splitting","questionImage":null,"type":"single","options":[{"id":"A","text":"Add a confidence threshold to its output so the model only acts on high-confidence analyses"},{"id":"B","text":"Split it into purpose-specific tools like `extract_data_points`, `summarize_content`, and `verify_claim_against_source`"},{"id":"C","text":"Rename it to `process_document` so its purpose reads more clearly in the tools array"},{"id":"D","text":"Give it more parameters so the caller can specify exactly which kind of analysis is wanted"}],"correctOptions":["B"],"matches":[],"match_pairs":[],"explanation":"Generic tools with broad responsibilities create ambiguity. Tool splitting replaces them with narrow, purpose-specific tools that each have a clear input/output contract, so the model can select the right one for what the user actually needs.","helpful_links":[],"tags":[]},{"id":"3d1b5060-5acc-4336-9bcc-55cfd33a81ac","original_quiz_id":"6a1a0c49ca96069990f7e8bc","created":"$D2026-06-03T18:39:05.747Z","description":"","questionContent":"The `isError` flag in MCP","questionImage":null,"type":"single","options":[{"id":"A","text":"It cancels the entire agent loop so no further tool calls can be made after a failure"},{"id":"B","text":"It signals a tool failure so the model can reason about recovery"},{"id":"C","text":"It indicates that the query ran successfully but returned an empty result set"},{"id":"D","text":"It forces an automatic retry of the failed tool call without involving the model"}],"correctOptions":["B"],"matches":[],"match_pairs":[],"explanation":"MCP's `isError` flag tells the model the tool execution failed, enabling it to reason about recovery rather than treating error text as a successful result. It does not auto-retry or cancel the loop, and it is distinct from a valid empty result.","helpful_links":[],"tags":[]},{"id":"f8d3d66b-397d-403b-b710-b1af010dcedd","original_quiz_id":"6a1a0c49ca96069990f7e8bc","created":"$D2026-06-03T18:39:05.747Z","description":"A refund tool returns: refund of £750 exceeds the £500 automatic limit.","questionContent":"Are business errors retryable?","questionImage":null,"type":"single","options":[{"id":"A","text":"Yes — retry with the same parameters, since transient conditions sometimes clear on a second attempt"},{"id":"B","text":"No — the same policy violation applies every time; take an alternative path"},{"id":"C","text":"Yes — retry after a short delay to allow any temporary rate limiting to reset"},{"id":"D","text":"Only retry once, then accept the result if the second attempt also fails"}],"correctOptions":["B"],"matches":[],"match_pairs":[],"explanation":"Business errors (policy violations, limit exceedances) are never resolved by retrying — the same request always fails. The correct `isRetryable` value is `false`, and the agent must take a different path, typically escalation or an alternative workflow.","helpful_links":[],"tags":[]},{"id":"b44652e8-846f-4ae2-901b-a30737293bc5","original_quiz_id":"6a1a0c49ca96069990f7e8bc","created":"$D2026-06-03T18:39:05.747Z","description":"A customer-lookup tool returns an empty array; the agent retries 3 times then escalates. Analysis shows the account simply does not exist.","questionContent":"Access failure versus valid empty result","questionImage":null,"type":"single","options":[{"id":"A","text":"Raise the retry limit to 5 so the lookup has more chances to return the missing record"},{"id":"B","text":"The tool fails to distinguish access failures from valid empty results, so \"no matches\" is treated as retryable"},{"id":"C","text":"The escalation threshold is too aggressive and triggers human involvement before retries are exhausted"},{"id":"D","text":"The system prompt should explicitly forbid the agent from retrying customer-lookup operations"}],"correctOptions":["B"],"matches":[],"match_pairs":[],"explanation":"The query succeeded and correctly found nothing — a valid empty result, not a failure. Because the response does not look structurally different from an access failure, the agent wastes retries and escalates wrongly. Mark valid empties `isError: false`, `resultCount: 0`, `shouldRetry: false`.","helpful_links":[],"tags":[]},{"id":"4b271a31-bffc-4f73-9e7b-050af5ed3b6c","original_quiz_id":"6a1a0c49ca96069990f7e8bc","created":"$D2026-06-03T18:39:05.747Z","description":"A single agent is given 18 tools and tool-selection reliability is poor.","questionContent":"Optimal number of tools per agent","questionImage":null,"type":"single","options":[{"id":"A","text":"8–10 tools, which balances coverage against the agent's ability to choose correctly"},{"id":"B","text":"4–5 tools scoped to the agent's role"},{"id":"C","text":"As many tools as possible, so the agent always has an option available for any request"},{"id":"D","text":"Exactly 2 tools, to make every selection decision a simple binary choice"}],"correctOptions":["B"],"matches":[],"match_pairs":[],"explanation":"Tool-selection reliability degrades as the toolkit grows. The optimal range is 4–5 tools per agent, each scoped to that agent's role. More tools means more decision complexity and more selection errors. A synthesis agent should not even hold web-search tools.","helpful_links":[],"tags":[]},{"id":"aa55303c-0498-4bee-84b0-a5170e581db8","original_quiz_id":"6a1a0c49ca96069990f7e8bc","created":"$D2026-06-03T18:39:05.747Z","description":"A document's type is unknown and you have separate `extract_invoice`, `extract_receipt`, and `extract_contract` schemas. You must guarantee structured output.","questionContent":"tool_choice for guaranteed structured output of unknown type","questionImage":null,"type":"single","options":[{"id":"A","text":"`tool_choice: auto`, letting the model decide whether a tool call or a text reply is more appropriate"},{"id":"B","text":"`tool_choice: any`"},{"id":"C","text":"`tool_choice: {type: tool, name: extract_invoice}`, forcing the invoice schema as the default extractor"},{"id":"D","text":"No tool_choice setting — rely on a detailed prompt instructing the model to always return JSON"}],"correctOptions":["B"],"matches":[],"match_pairs":[],"explanation":"`any` forces the model to call some tool (guaranteed structured output) while letting it choose which schema fits. `auto` may return conversational text. Forcing a specific named tool would be wrong because the type is unknown.","helpful_links":[],"tags":[]},{"id":"ad7ee886-10e6-4d32-a30d-fd05991fb2b6","original_quiz_id":"6a1a0c49ca96069990f7e8bc","created":"$D2026-06-03T18:39:05.747Z","description":"A synthesis agent returns to the coordinator for every fact check, adding 2–3 round trips and ~40% latency; analysis shows 85% are simple lookups.","questionContent":"Scoped cross-role tool","questionImage":null,"type":"single","options":[{"id":"A","text":"Increase coordinator parallelism so it can service the synthesis agent's verification requests faster"},{"id":"B","text":"Give the synthesis agent a scoped `verify_fact` tool for simple lookups, routing only complex ones to the coordinator"},{"id":"C","text":"Cache all verification results at the coordinator level so repeated lookups return instantly"},{"id":"D","text":"Remove fact verification from the synthesis workflow entirely to eliminate the round-trip latency"}],"correctOptions":["B"],"matches":[],"match_pairs":[],"explanation":"A scoped cross-role tool handles the high-frequency simple case (85%) locally, eliminating coordinator round-trip latency, while the 15% complex verifications still use the full pipeline. This is least privilege applied to tool distribution.","helpful_links":[],"tags":[]},{"id":"8813ec43-8dab-4571-976f-36a4b84aea54","original_quiz_id":"6a1a0c49ca96069990f7e8bc","created":"$D2026-06-03T18:39:05.747Z","description":"","questionContent":"Least privilege in tool design","questionImage":null,"type":"single","options":[{"id":"A","text":"Give every agent a general-purpose `fetch_url` tool so any of them can retrieve whatever resource it needs"},{"id":"B","text":"Replace `fetch_url` with a constrained `load_document` that validates document URLs only"},{"id":"C","text":"Give the coordinator every domain tool so it can step in and act directly whenever a subagent stalls"},{"id":"D","text":"Allow any agent to call any tool, relying on prompt instructions to keep usage within scope"}],"correctOptions":["B"],"matches":[],"match_pairs":[],"explanation":"A generic `fetch_url` can fetch anything from anywhere, enabling misuse. A constrained `load_document` validates document URLs only — preventing arbitrary fetches, clarifying purpose, and reducing side effects. This is the principle of least privilege applied to tool design.","helpful_links":[],"tags":[]},{"id":"d2452d2e-54e5-4039-b8a8-a6d2e999d4fb","original_quiz_id":"6a1a0c49ca96069990f7e8bc","created":"$D2026-06-03T18:39:05.747Z","description":"A team needs a Jira MCP server available to everyone who clones the repository.","questionContent":"MCP configuration scoping","questionImage":null,"type":"single","options":[{"id":"A","text":"Put it in `~/.claude.json` so each developer's personal configuration picks it up automatically"},{"id":"B","text":"Put it in project-level `.mcp.json`, version-controlled"},{"id":"C","text":"Have each developer configure the Jira server manually after cloning, following a setup document"},{"id":"D","text":"Hard-code the Jira server connection details directly into the project's system prompt"}],"correctOptions":["B"],"matches":[],"match_pairs":[],"explanation":"Project-level `.mcp.json` lives at the repo root, is version-controlled, and is shared with everyone who clones. User-level `~/.claude.json` is personal and not shared. Team-wide servers belong in `.mcp.json`.","helpful_links":[],"tags":[]},{"id":"168b575d-617a-41a6-8b05-faa987839cda","original_quiz_id":"6a1a0c49ca96069990f7e8bc","created":"$D2026-06-03T18:39:05.747Z","description":"","questionContent":"Keeping credentials out of version control","questionImage":null,"type":"single","options":[{"id":"A","text":"Commit the token directly in `.mcp.json` so every developer has working access immediately on clone"},{"id":"B","text":"Use `${GITHUB_TOKEN}` environment-variable expansion in `.mcp.json`"},{"id":"C","text":"Store the token inside the system prompt where it is loaded into context but not in the config file"},{"id":"D","text":"Disable version control for `.mcp.json` so the file with the token is never committed to the repository"}],"correctOptions":["B"],"matches":[],"match_pairs":[],"explanation":"`.mcp.json` supports `${VARIABLE_NAME}` expansion. The file references variable names, not values, so it is safe to commit. Each developer sets their own token locally, token rotation needs no config change, and no secrets leak through repository history.","helpful_links":[],"tags":[]},{"id":"1a8d9f9d-636c-4fa6-9eed-f908afb8b9d6","original_quiz_id":"6a1a0c49ca96069990f7e8bc","created":"$D2026-06-03T18:39:05.747Z","description":"A team needs to integrate with Jira; a developer proposes building a custom MCP server.","questionContent":"Build versus use an MCP server","questionImage":null,"type":"single","options":[{"id":"A","text":"Build the custom server immediately with exactly the API endpoints the team's workflow requires"},{"id":"B","text":"Evaluate existing community Jira MCP servers first; build custom only if they cannot handle team-specific workflows"},{"id":"C","text":"Call the Jira REST API directly from Bash commands instead of going through MCP at all"},{"id":"D","text":"Add the Jira integration to `~/.claude.json` so each developer configures and maintains it independently"}],"correctOptions":["B"],"matches":[],"match_pairs":[],"explanation":"For standard integrations (Jira, GitHub, Slack, Linear, Notion), evaluate community servers first — they cover standard cases, are tested, and save maintenance. Custom builds are justified only for team-specific workflows, embedded business logic, or proprietary systems.","helpful_links":[],"tags":[]},{"id":"3c69e3d9-6d70-47e0-95f0-2ec43cf38bee","original_quiz_id":"6a1a0c49ca96069990f7e8bc","created":"$D2026-06-03T18:39:05.747Z","description":"","questionContent":"MCP resources","questionImage":null,"type":"single","options":[{"id":"A","text":"They replace tools entirely, so an agent with resources no longer needs any tool definitions"},{"id":"B","text":"They expose content catalogues upfront, reducing exploratory tool calls"},{"id":"C","text":"They store agent state between sessions so a later session can resume where an earlier one stopped"},{"id":"D","text":"They enforce rate limits on how frequently an agent may call each connected tool"}],"correctOptions":["B"],"matches":[],"match_pairs":[],"explanation":"MCP resources present content catalogues (issue summaries, documentation hierarchies, database schemas) upfront so the agent does not need exploratory tool calls (e.g., `list_tables` then `describe_table` for each). Resources let agents see data; tools let agents act on it.","helpful_links":[],"tags":[]},{"id":"b7eb2158-9484-4ac7-8947-9872f1a23d52","original_quiz_id":"6a1a0c49ca96069990f7e8bc","created":"$D2026-06-03T18:39:05.747Z","description":"","questionContent":"Grep versus Glob","questionImage":null,"type":"single","options":[{"id":"A","text":"Both search the contents of files, differing only in the pattern syntax they accept"},{"id":"B","text":"Grep searches file contents; Glob matches file paths by naming pattern"},{"id":"C","text":"Grep matches file paths by name, while Glob searches inside files for matching text"},{"id":"D","text":"Both match file paths, with Grep handling regex and Glob handling wildcards"}],"correctOptions":["B"],"matches":[],"match_pairs":[],"explanation":"Grep finds what is inside files (function callers, error strings, imports). Glob finds files by their names (`**/*.test.tsx`, `**/config.*`). Using Glob to find function callers fails because Glob matches paths, not contents.","helpful_links":[],"tags":[]},{"id":"0a49fed0-e77c-4c75-afb0-83ecd1e0fa5a","original_quiz_id":"6a1a0c49ca96069990f7e8bc","created":"$D2026-06-03T18:39:05.747Z","description":"A developer must find all callers of `processLegacyOrder()` and all test files for those callers.","questionContent":"Deprecated-function trace sequence","questionImage":null,"type":"single","options":[{"id":"A","text":"Glob for `*processLegacyOrder*` to locate the callers, then Grep across the codebase for the test files"},{"id":"B","text":"Grep for `processLegacyOrder` to find callers, then Glob for the sibling test files of each caller"},{"id":"C","text":"Read every source file manually to find the callers, then read every test file to find their tests"},{"id":"D","text":"Use a Bash combination of `find` and `xargs grep` to handle both the caller search and the test search"}],"correctOptions":["B"],"matches":[],"match_pairs":[],"explanation":"Grep (content search) finds every file referencing the function, including tests importing it directly. Glob (path matching) then catches sibling test files like `**/OrderProcessor.test.*` that exercise it indirectly. The order is Grep then Glob — never Glob first, since Glob cannot search contents.","helpful_links":[],"tags":[]},{"id":"4ffc9da1-35fd-4ad6-b6e3-7dcabc62cf48","original_quiz_id":"6a1a0c49ca96069990f7e8bc","created":"$D2026-06-03T18:39:05.747Z","description":"An Edit fails with \"old_string matches 3 locations.\"","questionContent":"Edit non-unique match recovery","questionImage":null,"type":"single","options":[{"id":"A","text":"Immediately fall back to Read + Write, loading the whole file and rewriting it with the change applied"},{"id":"B","text":"Widen `old_string` with surrounding context until it pins one location, or set `replace_all: true`"},{"id":"C","text":"Delete the file entirely and recreate it from scratch with the intended modification included"},{"id":"D","text":"Use a Bash `sed` command to perform the substitution outside of the Edit tool"}],"correctOptions":["B"],"matches":[],"match_pairs":[],"explanation":"The documented recovery is to widen the anchor with more surrounding context until it matches one location, or set `replace_all: true` for a global change. Both stay on Edit and cost almost nothing. Read + Write is the last resort, penalised when used as the immediate next step.","helpful_links":[],"tags":[]},{"id":"91f4c38f-520d-48ff-b441-383456192388","original_quiz_id":"6a1a0c49ca96069990f7e8bc","created":"$D2026-06-03T18:39:05.747Z","description":"An agent needs to understand how a feature works in a 200-file repository.","questionContent":"Incremental codebase understanding","questionImage":null,"type":"single","options":[{"id":"A","text":"Read all 200 files into context upfront so the agent has complete information before reasoning"},{"id":"B","text":"Grep to find entry points, Read to trace flows, Grep again to trace usage, Read only what is justified"},{"id":"C","text":"Glob for every file in the repository and Read them one by one in directory order"},{"id":"D","text":"Ask the user to write a summary of the codebase so the agent can skip exploration entirely"}],"correctOptions":["B"],"matches":[],"match_pairs":[],"explanation":"Reading all files upfront is a context-budget killer. Incremental discovery starts narrow: Grep finds entry points, Read follows imports and flows, Grep traces usage, and each Read is justified by the previous step — minimal context for maximum understanding.","helpful_links":[],"tags":[]},{"id":"0ed964bf-c567-481b-9611-7cbd368e8044","original_quiz_id":"6a1a0c49ca96069990f7e8bc","created":"$D2026-06-03T18:39:05.747Z","description":"An agent prefers the built-in Grep over a more capable MCP `search_codebase` tool with a one-line description.","questionContent":"Enhancing MCP tool descriptions","questionImage":null,"type":"single","options":[{"id":"A","text":"Disable Grep entirely so the agent is forced to fall back to the MCP search tool instead"},{"id":"B","text":"Enhance the MCP tool's description to explain its capabilities, outputs, and when to prefer it over built-ins"},{"id":"C","text":"Rename the MCP tool to match Grep so the model treats them as equivalent options"},{"id":"D","text":"Add the MCP tool to `~/.claude.json` so it loads with higher priority than the built-in tools"}],"correctOptions":["B"],"matches":[],"match_pairs":[],"explanation":"The model favours tools it understands best, and built-ins have rich descriptions. A sparse MCP description loses out. Enhancing it to explain capabilities, outputs, and when to prefer it over Grep gives the model enough context to select it when it is genuinely more capable.","helpful_links":[],"tags":[]},{"id":"5dccd65a-cea1-4617-adae-6fa4f90a3bfa","original_quiz_id":"6a1a0c49ca96069990f7e8bc","created":"$D2026-06-03T18:39:05.747Z","description":"A tool returns: \"Order ID must be in format #NNNNN. Received 'order-abc'.\"","questionContent":"Validation versus transient error recovery","questionImage":null,"type":"single","options":[{"id":"A","text":"Retry the identical request after a short delay, in case the formatting check was a transient glitch"},{"id":"B","text":"Reformat the input to the required format and retry"},{"id":"C","text":"Escalate the request to a human immediately, since the tool has rejected the agent's input"},{"id":"D","text":"Mark the error as non-retryable and stop, treating the malformed input as an unrecoverable condition"}],"correctOptions":["B"],"matches":[],"match_pairs":[],"explanation":"This is a validation error — the request is malformed but the operation is otherwise valid. Recovery is to fix the input and retry. A transient error would warrant a delayed retry of the same request; a business error would warrant an alternative path.","helpful_links":[],"tags":[]},{"id":"7827d2c9-2e29-42a8-86d9-760cf4da7b04","original_quiz_id":"6a1a0c49ca96069990f7e8bc","created":"$D2026-06-03T18:39:05.747Z","description":"A question asks \"which CLAUDE.md wins when two rules conflict?\"","questionContent":"CLAUDE.md is not strict precedence","questionImage":null,"type":"single","options":[{"id":"A","text":"The most specific scope always wins, so a directory-level file overrides the project root file"},{"id":"B","text":"Files are concatenated; on a true conflict Claude may pick one arbitrarily — move must-honour rules to settings.json or a hook"},{"id":"C","text":"User-level configuration always overrides project-level configuration on a conflict"},{"id":"D","text":"The first file loaded in the hierarchy wins, and later files cannot override its rules"}],"correctOptions":["B"],"matches":[],"match_pairs":[],"explanation":"CLAUDE.md files are concatenated into context, not resolved by precedence. Delivered as a user message, there is no guarantee of strict compliance, so conflicts may resolve arbitrarily. Rules that must always hold belong in `settings.json` (client-enforced, strict precedence) or a hook. Distractors claiming \"more specific wins\" or \"user overrides project\" are paraphrases the docs never make.","helpful_links":[],"tags":[]},{"id":"7800e682-3cdd-47e5-afc6-f0460508c1b8","original_quiz_id":"6a1a0c49ca96069990f7e8bc","created":"$D2026-06-03T18:39:05.747Z","description":"Developer A's Claude Code follows the team's API naming perfectly. New Developer B, on the same repo and branch, gets inconsistent naming.","questionContent":"New team member gets inconsistent conventions","questionImage":null,"type":"single","options":[{"id":"A","text":"Developer B has not yet run `/memory`, which is the command that loads the configuration files into the session"},{"id":"B","text":"The conventions live in Developer A's user-level `~/.claude/CLAUDE.md`, which is not shared via git"},{"id":"C","text":"Developer B is missing an MCP server that supplies the naming-convention rules to the session"},{"id":"D","text":"The rules live in a `.claude/rules/` file that Developer B's version of the tool does not support"}],"correctOptions":["B"],"matches":[],"match_pairs":[],"explanation":"User-level config is not version-controlled, so a fresh clone never receives it. Move the conventions to project-level `.claude/CLAUDE.md`. `/memory` only reports loaded files — it does not trigger loading.","helpful_links":[],"tags":[]},{"id":"e2c8505c-3dd9-4928-ac6c-373910275e2d","original_quiz_id":"6a1a0c49ca96069990f7e8bc","created":"$D2026-06-03T18:39:05.747Z","description":"","questionContent":"What the /memory command does","questionImage":null,"type":"single","options":[{"id":"A","text":"It loads the configuration files for the current directory into the active session"},{"id":"B","text":"It reveals which memory files are currently loaded"},{"id":"C","text":"It clears the context window to free space during a long session"},{"id":"D","text":"It compacts the conversation by summarising earlier turns"}],"correctOptions":["B"],"matches":[],"match_pairs":[],"explanation":"`/memory` is a diagnostic command that shows which memory files are loaded. Configuration loads automatically based on location in the hierarchy. `/memory` helps debug inconsistent behaviour across developers — it does not activate anything. `/compact` is the command that compacts.","helpful_links":[],"tags":[]},{"id":"a1a8b8eb-1f57-4a3f-94b4-11935a7a8888","original_quiz_id":"6a1a0c49ca96069990f7e8bc","created":"$D2026-06-03T18:39:05.747Z","description":"An engineer splits a 600-line CLAUDE.md into six 100-line `@`-imported files to reduce per-session context.","questionContent":"The @ import directive and eager loading","questionImage":null,"type":"single","options":[{"id":"A","text":"This reduces the context Claude sees, because only the imported files actually referenced get loaded"},{"id":"B","text":"Imports load eagerly, so the context size is unchanged; use path-scoped `.claude/rules/` to shrink context"},{"id":"C","text":"The `@import` keyword is required for each line so the loader knows to inline the referenced file"},{"id":"D","text":"Imports load lazily and only when the conversation explicitly references the imported file by path"}],"correctOptions":["B"],"matches":[],"match_pairs":[],"explanation":"`@` imports inline the referenced file eagerly at load time, so the context Claude sees is the same size as if pasted in — splitting helps maintainability, not tokens. To actually shrink per-session context, use `.claude/rules/` with path-scoped frontmatter. There is no `@import` keyword.","helpful_links":[],"tags":[]},{"id":"d1c81d14-b98e-47ef-a470-9483e0b14466","original_quiz_id":"6a1a0c49ca96069990f7e8bc","created":"$D2026-06-03T18:39:05.747Z","description":"Test files are co-located with source files across 50+ directories; the team wants uniform test conventions everywhere.","questionContent":"Cross-directory test conventions","questionImage":null,"type":"single","options":[{"id":"A","text":"Place a CLAUDE.md file in every directory that contains test files so the rules are always nearby"},{"id":"B","text":"A `.claude/rules/` file with `paths: [\"**/*.test.tsx\", \"**/*.test.ts\"]`"},{"id":"C","text":"Put all of the test conventions in the root CLAUDE.md so they load for every session automatically"},{"id":"D","text":"Create a skill containing the test conventions and instruct developers to invoke it before writing tests"}],"correctOptions":["B"],"matches":[],"match_pairs":[],"explanation":"Glob-scoped rules give one file, one pattern, universal coverage — loaded only when editing matching files. Per-directory CLAUDE.md means 50+ copies and inevitable drift; root CLAUDE.md wastes tokens on every session; a skill is on-demand, not always-on for a file type.","helpful_links":[],"tags":[]},{"id":"d10efa77-22c1-4049-b578-ab435caddead","original_quiz_id":"6a1a0c49ca96069990f7e8bc","created":"$D2026-06-03T18:39:05.747Z","description":"","questionContent":"Why path-scoped rules beat root CLAUDE.md","questionImage":null,"type":"single","options":[{"id":"A","text":"They support a richer set of markdown features than the root CLAUDE.md file does"},{"id":"B","text":"They load only when editing matching files, saving tokens on unrelated work"},{"id":"C","text":"They take precedence over and can override the client-enforced settings.json file"},{"id":"D","text":"They are a required prerequisite for connecting and using any MCP server"}],"correctOptions":["B"],"matches":[],"match_pairs":[],"explanation":"Root CLAUDE.md loads for every session regardless of which files you edit, so Terraform conventions consume tokens while editing React. Path-scoped rules load only when editing matching files, keeping the model focused and the token budget lean.","helpful_links":[],"tags":[]},{"id":"abf1441d-a49b-4695-b811-c61936c11127","original_quiz_id":"6a1a0c49ca96069990f7e8bc","created":"$D2026-06-03T18:39:05.747Z","description":"","questionContent":"Unified skills system locations","questionImage":null,"type":"single","options":[{"id":"A","text":"Only `.claude/commands/` creates slash commands; the skills directory serves a different purpose"},{"id":"B","text":"`.claude/skills/` (canonical) and `.claude/commands/` (alias) both create equivalent `/commands`"},{"id":"C","text":"Skills must live exclusively in `~/.claude/`, since the project directory is reserved for commands"},{"id":"D","text":"Skills and commands are entirely separate systems that cannot create the same `/command`"}],"correctOptions":["B"],"matches":[],"match_pairs":[],"explanation":"Commands and skills are unified. Files at `.claude/skills/` (canonical) or `.claude/commands/` (backward-compatible alias) both create a `/command`. `.claude/skills/` is preferred for configured skills (frontmatter, supporting files, auto-discovery).","helpful_links":[],"tags":[]},{"id":"a85203d3-c7fa-45aa-a83e-d5d7d31cd30c","original_quiz_id":"6a1a0c49ca96069990f7e8bc","created":"$D2026-06-03T18:39:05.747Z","description":"A personal `/brainstorm` skill produces verbose codebase analysis that should not clutter the main conversation.","questionContent":"The context: fork frontmatter option","questionImage":null,"type":"single","options":[{"id":"A","text":"`allowed-tools`, which limits the skill to the specific tools it is permitted to use"},{"id":"B","text":"`context: fork`"},{"id":"C","text":"`argument-hint`, which prompts the developer for the parameters the skill expects"},{"id":"D","text":"`disable-model-invocation: true`, which prevents the model from auto-invoking the skill"}],"correctOptions":["B"],"matches":[],"match_pairs":[],"explanation":"`context: fork` runs the skill in an isolated sub-agent context so verbose output stays contained and the main conversation remains clean. Without it, exploratory output fills the context window and degrades subsequent responses.","helpful_links":[],"tags":[]},{"id":"d41745d9-c6c5-4973-bf4d-6a716597b4c6","original_quiz_id":"6a1a0c49ca96069990f7e8bc","created":"$D2026-06-03T18:39:05.747Z","description":"The team needs API naming conventions applied to every code-generation task, plus an occasional multi-step codebase-analysis workflow.","questionContent":"Skills versus CLAUDE.md placement","questionImage":null,"type":"single","options":[{"id":"A","text":"Put both the conventions and the analysis workflow in skills so they are invoked explicitly"},{"id":"B","text":"Conventions in CLAUDE.md (or rules); the analysis workflow as a skill"},{"id":"C","text":"Put both the conventions and the analysis workflow in CLAUDE.md so they are always loaded"},{"id":"D","text":"Make the conventions a skill and document the analysis workflow inside CLAUDE.md"}],"correctOptions":["B"],"matches":[],"match_pairs":[],"explanation":"CLAUDE.md is for always-loaded universal standards (apply to every task). Skills are on-demand, task-specific workflows. Do not put always-on conventions in skills, and do not put task-specific procedures in CLAUDE.md.","helpful_links":[],"tags":[]},{"id":"8e185d2f-9ccf-43ae-bae2-fdec72d1c446","original_quiz_id":"6a1a0c49ca96069990f7e8bc","created":"$D2026-06-03T18:39:05.747Z","description":"","questionContent":"allowed-tools as a security boundary","questionImage":null,"type":"single","options":[{"id":"A","text":"It speeds up the skill by limiting how many tools the model has to evaluate before acting"},{"id":"B","text":"It restricts which tools the skill may use, e.g., read-only Read/Grep/Glob"},{"id":"C","text":"It is a required field that every skill must declare before it can be invoked"},{"id":"D","text":"It changes the skill's scope from user-level to project-level so the whole team can use it"}],"correctOptions":["B"],"matches":[],"match_pairs":[],"explanation":"`allowed-tools` restricts the tools a skill can access during execution — a security boundary. A read-only analysis skill should list only Read, Grep, and Glob, never Write or Bash, so it cannot modify source code.","helpful_links":[],"tags":[]},{"id":"31904063-ba5f-42c9-9b97-e49c295e0d9f","original_quiz_id":"6a1a0c49ca96069990f7e8bc","created":"$D2026-06-03T18:39:05.747Z","description":"","questionContent":"Plan mode versus direct execution decision factor","questionImage":null,"type":"single","options":[{"id":"A","text":"The difficulty of the task, since harder problems always benefit from an upfront planning phase"},{"id":"B","text":"Ambiguity of the task"},{"id":"C","text":"The number of lines of code the change is expected to add or modify"},{"id":"D","text":"The programming language the codebase is written in and its tooling maturity"}],"correctOptions":["B"],"matches":[],"match_pairs":[],"explanation":"The deciding factor is ambiguity, not difficulty. A difficult but well-defined bug fix (clear stack trace, single function) is direct execution. A simple-sounding feature implementable three ways and touching multiple modules is plan mode.","helpful_links":[],"tags":[]},{"id":"373418c8-fb79-4d7c-9df7-f5bcc365166e","original_quiz_id":"6a1a0c49ca96069990f7e8bc","created":"$D2026-06-03T18:39:05.747Z","description":"(1) Restructure a monolith into microservices; (2) Fix a null pointer in a single function with a clear stack trace; (3) Migrate a logging library across 30 files.","questionContent":"Three tasks, which modes","questionImage":null,"type":"single","options":[{"id":"A","text":"Plan mode for all three, because each one involves modifying production source code in some way"},{"id":"B","text":"Plan mode for (1) and (3), direct execution for (2)"},{"id":"C","text":"Plan mode for (1) only, with direct execution for both (2) and (3) since they have known fixes"},{"id":"D","text":"Direct execution for all three, supported by comprehensive upfront instructions in each prompt"}],"correctOptions":["B"],"matches":[],"match_pairs":[],"explanation":"(1) is architectural restructuring → plan mode. (3) is a multi-file migration → plan then execute. (2) is well-scoped with a known cause → direct execution. The migration benefits from planning a consistent strategy before applying it file by file.","helpful_links":[],"tags":[]},{"id":"2d7610a1-910f-44aa-a8e9-f5bba328c944","original_quiz_id":"6a1a0c49ca96069990f7e8bc","created":"$D2026-06-03T18:39:05.747Z","description":"A task is stated as \"restructure the monolith into microservices.\"","questionContent":"Recognising complexity upfront","questionImage":null,"type":"single","options":[{"id":"A","text":"Start with direct execution and switch to plan mode later if unexpected complexity actually emerges"},{"id":"B","text":"Choose plan mode upfront because the complexity is already stated"},{"id":"C","text":"Use the Explore subagent on its own to scope the work before deciding on a mode"},{"id":"D","text":"Use direct execution but supply comprehensive, detailed instructions to cover the complexity"}],"correctOptions":["B"],"matches":[],"match_pairs":[],"explanation":"When complexity is stated in the requirements, plan mode is chosen immediately. The complexity is known, not speculative. Waiting for surprises to emerge before planning is the wrong approach and risks costly rework.","helpful_links":[],"tags":[]},{"id":"11c29cd0-dda3-4eec-8c44-4130e5e666ca","original_quiz_id":"6a1a0c49ca96069990f7e8bc","created":"$D2026-06-03T18:39:05.747Z","description":"","questionContent":"The Explore subagent","questionImage":null,"type":"single","options":[{"id":"A","text":"It executes the planned changes once the design phase has produced an implementation strategy"},{"id":"B","text":"It isolates verbose discovery output, returning summaries to keep the main context clean"},{"id":"C","text":"It enforces permission boundaries on which files and tools the main agent may touch"},{"id":"D","text":"It runs only inside CI pipelines, where interactive exploration is not available"}],"correctOptions":["B"],"matches":[],"match_pairs":[],"explanation":"The Explore subagent runs verbose discovery (file listings, dependency graphs) in isolation and returns summaries to the main conversation, keeping the main context clean for implementation work.","helpful_links":[],"tags":[]},{"id":"18f52763-8bfc-48d7-82e5-082e13f2ffb4","original_quiz_id":"6a1a0c49ca96069990f7e8bc","created":"$D2026-06-03T18:39:05.747Z","description":"A developer describes a transformation in prose and Claude Code interprets it differently each run.","questionContent":"First refinement technique for inconsistent interpretation","questionImage":null,"type":"single","options":[{"id":"A","text":"Rewrite the prose with more precise, technically detailed language to remove room for interpretation"},{"id":"B","text":"Provide 2–3 concrete input/output examples"},{"id":"C","text":"Use the interview pattern, having Claude ask clarifying questions before it implements anything"},{"id":"D","text":"Write a comprehensive test suite and iterate by feeding the failures back to the model"}],"correctOptions":["B"],"matches":[],"match_pairs":[],"explanation":"Inconsistent interpretation of prose is the textbook trigger for concrete examples. The model generalises from 2–3 before/after pairs far more reliably than from more prose, which still relies on interpretation.","helpful_links":[],"tags":[]},{"id":"c9f25ff6-2088-45bf-8926-1908c43c8fcf","original_quiz_id":"6a1a0c49ca96069990f7e8bc","created":"$D2026-06-03T18:39:05.747Z","description":"","questionContent":"Interview pattern versus examples","questionImage":null,"type":"single","options":[{"id":"A","text":"They solve the same underlying problem and can be used interchangeably as needed"},{"id":"B","text":"The interview pattern suits unfamiliar domains; examples suit known transformations the model misinterprets"},{"id":"C","text":"Examples are the right tool for unfamiliar domains where you might overlook key considerations"},{"id":"D","text":"The interview pattern is best for fixing inconsistent output formatting across runs"}],"correctOptions":["B"],"matches":[],"match_pairs":[],"explanation":"The interview pattern has Claude ask questions before implementing, surfacing considerations you would miss in an unfamiliar domain. Examples are for when you know the exact transformation but the model misinterprets it. They solve different problems.","helpful_links":[],"tags":[]},{"id":"9f6fe1c4-f12e-482b-840f-baaceab2f38f","original_quiz_id":"6a1a0c49ca96069990f7e8bc","created":"$D2026-06-03T18:39:05.747Z","description":"Three fixes are needed: the error-response shape, the logging format, and the client SDK type definitions, and they depend on each other.","questionContent":"Batch versus sequential feedback","questionImage":null,"type":"single","options":[{"id":"A","text":"Fix them one at a time in sequence so each change can be verified before the next begins"},{"id":"B","text":"Provide all three in one message so the model sees all interacting constraints"},{"id":"C","text":"Fix only the most important of the three and leave the others for a later iteration"},{"id":"D","text":"Escalate the three interdependent fixes to a human engineer for manual coordination"}],"correctOptions":["B"],"matches":[],"match_pairs":[],"explanation":"When fixes interact, batch them in one message so the model sees all constraints at once and produces a coherent fix. Independent issues are fixed sequentially to avoid confusing which feedback applies where.","helpful_links":[],"tags":[]},{"id":"a2c75780-43f4-49a2-8863-3e3e4c338169","original_quiz_id":"6a1a0c49ca96069990f7e8bc","created":"$D2026-06-03T18:39:05.747Z","description":"A CI script runs `claude` with a prompt but the job hangs; logs show Claude waiting for interactive input.","questionContent":"The most directly testable CI/CD fact","questionImage":null,"type":"single","options":[{"id":"A","text":"Set the environment variable `CLAUDE_HEADLESS=true` before invoking the command in the pipeline"},{"id":"B","text":"Add the `-p` (`--print`) flag"},{"id":"C","text":"Redirect standard input from `/dev/null` so the process has no interactive input to wait for"},{"id":"D","text":"Add a `--batch` flag to switch Claude Code into non-interactive batch execution"}],"correctOptions":["B"],"matches":[],"match_pairs":[],"explanation":"The `-p`/`--print` flag runs Claude Code non-interactively: it processes the prompt, prints to stdout, and exits. `CLAUDE_HEADLESS=true` and `--batch` do not exist, and stdin redirection does not properly address interactive mode.","helpful_links":[],"tags":[]},{"id":"c74b2b54-1039-468a-b03f-c6e7f54bace7","original_quiz_id":"6a1a0c49ca96069990f7e8bc","created":"$D2026-06-03T18:39:05.747Z","description":"","questionContent":"Structured CI output","questionImage":null,"type":"single","options":[{"id":"A","text":"Use `--verbose` text mode and parse the richer human-readable output downstream"},{"id":"B","text":"`--output-format json` with `--json-schema`"},{"id":"C","text":"Parse the plain stdout with regular expressions to extract the findings programmatically"},{"id":"D","text":"Use `-p` alone, since print mode already emits machine-parseable structured output"}],"correctOptions":["B"],"matches":[],"match_pairs":[],"explanation":"CI output must be machine-parseable. `--output-format json` forces JSON and `--json-schema` enforces a structure (file, line, severity, message), enabling automated systems to post inline PR comments, filter by severity, and track findings across runs.","helpful_links":[],"tags":[]},{"id":"58417f87-e2a2-4590-84f9-e71b522864c2","original_quiz_id":"6a1a0c49ca96069990f7e8bc","created":"$D2026-06-03T18:39:05.747Z","description":"","questionContent":"Self-review in the same session","questionImage":null,"type":"single","options":[{"id":"A","text":"It is as effective as independent review, since the same model has all the relevant context loaded"},{"id":"B","text":"It is weaker because the session retains generation reasoning, biasing it toward its own decisions"},{"id":"C","text":"It is more thorough than independent review because the model remembers why it made each choice"},{"id":"D","text":"It is only weaker for large files, where the generation reasoning crowds out review attention"}],"correctOptions":["B"],"matches":[],"match_pairs":[],"explanation":"A session that generated code retains its reasoning (\"why I chose this\") and is less likely to question its decisions. An independent instance with no prior context evaluates the code on its own merits and catches more subtle issues.","helpful_links":[],"tags":[]},{"id":"bddb28e7-5b6f-40aa-881e-2a7eceed27f9","original_quiz_id":"6a1a0c49ca96069990f7e8bc","created":"$D2026-06-03T18:39:05.747Z","description":"A manager proposes moving a blocking pre-merge check and an overnight technical-debt report to the Message Batches API for 50% savings.","questionContent":"Batch API for a blocking pre-merge check","questionImage":null,"type":"single","options":[{"id":"A","text":"Move both workflows to batch and poll the batch status to know when each result is ready"},{"id":"B","text":"Batch the overnight reports; keep real-time for the pre-merge check"},{"id":"C","text":"Keep both workflows on the real-time API to avoid any result-ordering complications"},{"id":"D","text":"Move both to batch but add a real-time fallback that fires if the batch exceeds a timeout"}],"correctOptions":["B"],"matches":[],"match_pairs":[],"explanation":"Pre-merge checks are blocking — developers wait for results, and the batch API's up-to-24-hour window with no SLA is unacceptable. Only latency-tolerant work (overnight reports) should use batch.","helpful_links":[],"tags":[]},{"id":"cffb223f-16b2-4a0a-96d3-87f4d0c77fe9","original_quiz_id":"6a1a0c49ca96069990f7e8bc","created":"$D2026-06-03T18:39:05.747Z","description":"","questionContent":"Incremental review context","questionImage":null,"type":"single","options":[{"id":"A","text":"Re-analyse the whole PR from scratch on every push so each review reflects the complete current state"},{"id":"B","text":"Include prior findings and report only new or still-unaddressed issues"},{"id":"C","text":"Disable automated review on every push after the first to avoid duplicate comments entirely"},{"id":"D","text":"Comment only on the files changed in the latest commit, ignoring the rest of the PR"}],"correctOptions":["B"],"matches":[],"match_pairs":[],"explanation":"Without prior context, each run re-analyses from scratch and produces duplicate comments, eroding developer trust. Including prior findings and instructing Claude to report only new or unaddressed issues preserves the signal-to-noise ratio.","helpful_links":[],"tags":[]},{"id":"73a56b7a-1da8-42ae-9934-268a36e877ae","original_quiz_id":"6a1a0c49ca96069990f7e8bc","created":"$D2026-06-03T18:39:05.747Z","description":"A code-review prompt says \"Be conservative. Only report high-confidence findings.\"","questionContent":"Vague instructions versus explicit criteria","questionImage":null,"type":"single","options":[{"id":"A","text":"This is the recommended production approach, since conservative settings minimise false positives"},{"id":"B","text":"Replace it with explicit categorical criteria (what to flag, what to skip) and code examples"},{"id":"C","text":"Add \"use your best judgement\" so the model can weigh each finding on a case-by-case basis"},{"id":"D","text":"Raise the internal confidence threshold so only the strongest findings are surfaced"}],"correctOptions":["B"],"matches":[],"match_pairs":[],"explanation":"Vague instructions give no actionable decision boundary; \"conservative\" and \"high-confidence\" are uncalibrated. Explicit categorical criteria define exactly what to report (bugs, security) and skip (style, local patterns), with concrete code examples per severity. Explicit criteria first, confidence routing second.","helpful_links":[],"tags":[]},{"id":"87d59213-ac0c-4573-b88c-2216aa7333cd","original_quiz_id":"6a1a0c49ca96069990f7e8bc","created":"$D2026-06-03T18:39:05.747Z","description":"A pipeline has a 40% false-positive rate on documentation findings, causing developers to ignore ALL categories, including accurate security findings.","questionContent":"The false-positive trust problem","questionImage":null,"type":"single","options":[{"id":"A","text":"Add \"only report high-confidence documentation issues\" so the noisy category surfaces fewer findings"},{"id":"B","text":"Temporarily disable the documentation category while refining its criteria with code examples"},{"id":"C","text":"Raise the model temperature and filter out the statistical outliers from the documentation findings"},{"id":"D","text":"Add a second model pass that re-verifies every documentation finding before it is reported"}],"correctOptions":["B"],"matches":[],"match_pairs":[],"explanation":"Trust is not category-specific — it bleeds across all output. Temporarily disabling the high-false-positive category restores system-wide trust while you refine its criteria with concrete code examples, re-enabling once precision improves.","helpful_links":[],"tags":[]},{"id":"9605a753-af7c-4235-8b81-d060a67c7e5b","original_quiz_id":"6a1a0c49ca96069990f7e8bc","created":"$D2026-06-03T18:39:05.747Z","description":"","questionContent":"Why confidence-based filtering fails","questionImage":null,"type":"single","options":[{"id":"A","text":"It is too slow to compute a confidence score for every finding in a large review"},{"id":"B","text":"LLM self-reported confidence is poorly calibrated"},{"id":"C","text":"It requires running a second, separate model purely to produce the confidence estimates"},{"id":"D","text":"It cannot be expressed in a JSON schema alongside the rest of the finding"}],"correctOptions":["B"],"matches":[],"match_pairs":[],"explanation":"LLM self-reported confidence is poorly calibrated — often confident on wrong findings and uncertain on right ones. Confidence is useful for routing after calibration, but it is no substitute for explicit criteria defining what a valid finding is.","helpful_links":[],"tags":[]},{"id":"3a59cd24-2ce5-42f7-a60b-b4b5fd540d24","original_quiz_id":"6a1a0c49ca96069990f7e8bc","created":"$D2026-06-03T18:39:05.747Z","description":"","questionContent":"What tool_use with a JSON schema does NOT prevent","questionImage":null,"type":"single","options":[{"id":"A","text":"Malformed JSON output such as missing brackets, trailing commas, or unquoted keys"},{"id":"B","text":"Semantic errors like wrong sums, misplaced values, and fabricated data"},{"id":"C","text":"Missing required fields that the schema declares as mandatory in the output object"},{"id":"D","text":"Incorrect data types, such as a string supplied where the schema expects an integer"}],"correctOptions":["B"],"matches":[],"match_pairs":[],"explanation":"`tool_use` eliminates syntax errors (malformed JSON, missing fields, wrong types) but not semantic errors — line items that do not sum, values in the wrong fields, or fabricated values for fields the source lacks. Semantic validation needs additional logic.","helpful_links":[],"tags":[]},{"id":"34671bf6-0428-4cf9-9953-36a0afc43921","original_quiz_id":"6a1a0c49ca96069990f7e8bc","created":"$D2026-06-03T18:39:05.747Z","description":"An extraction system with all-required fields invents plausible dates and amounts when documents lack that information.","questionContent":"Preventing fabrication of missing values","questionImage":null,"type":"single","options":[{"id":"A","text":"Add an explicit instruction telling the model never to hallucinate values for fields it cannot find"},{"id":"B","text":"Make fields optional or nullable when the source may lack them"},{"id":"C","text":"Switch from tool_use to prompt-based JSON extraction for more flexibility in handling missing data"},{"id":"D","text":"Add a post-extraction validation step that checks each value against the original source document"}],"correctOptions":["B"],"matches":[],"match_pairs":[],"explanation":"Required fields pressure the model to invent a value when information is absent. Making fields optional/nullable lets it honestly return `null`, which is always preferable to plausible fabrication. This is the structural root-cause fix.","helpful_links":[],"tags":[]},{"id":"c33809e3-702b-484c-9cac-61d6959214e1","original_quiz_id":"6a1a0c49ca96069990f7e8bc","created":"$D2026-06-03T18:39:05.747Z","description":"","questionContent":"tool_choice auto versus any","questionImage":null,"type":"single","options":[{"id":"A","text":"`auto` guarantees a tool call, removing the possibility of a plain text response"},{"id":"B","text":"`any` guarantees a tool call but lets the model choose which"},{"id":"C","text":"They are identical in behaviour and differ only in the name used in the request"},{"id":"D","text":"`auto` forces the model to call one specific named tool that you designate in the request"}],"correctOptions":["B"],"matches":[],"match_pairs":[],"explanation":"`auto` lets the model return conversational text instead of calling a tool — no guarantee of structured output. `any` guarantees a tool call while letting the model choose which schema. For guaranteed structured output with unknown type, use `any`.","helpful_links":[],"tags":[]},{"id":"c2a3a2b2-0e7f-4582-ad4f-f5861aedef03","original_quiz_id":"6a1a0c49ca96069990f7e8bc","created":"$D2026-06-03T18:39:05.747Z","description":"","questionContent":"Retry-with-error-feedback contents","questionImage":null,"type":"single","options":[{"id":"A","text":"Just the original prompt again, re-sent unchanged so the model gets a fresh attempt"},{"id":"B","text":"The original document, the failed extraction, and the specific validation error"},{"id":"C","text":"Only the specific validation error, so the model focuses narrowly on what went wrong"},{"id":"D","text":"A higher temperature setting so the retry produces a different result than the first attempt"}],"correctOptions":["B"],"matches":[],"match_pairs":[],"explanation":"Effective retries send three things: the original document (to re-examine), the failed extraction (to see what was produced), and the specific validation error (to target the fix). Naive retries without the specific error reproduce the same mistake.","helpful_links":[],"tags":[]},{"id":"4213619a-42ac-4fd7-a1f4-f3e3caa30144","original_quiz_id":"6a1a0c49ca96069990f7e8bc","created":"$D2026-06-03T18:39:05.747Z","description":"Document A's line items sum to £450 but the stated total is £500; Document B is missing the `department` field entirely from the source.","questionContent":"The retry effectiveness boundary","questionImage":null,"type":"single","options":[{"id":"A","text":"Retry both documents with their respective validation errors and let the model self-correct each"},{"id":"B","text":"Retry A with the discrepancy error; flag B for human review since the info is absent"},{"id":"C","text":"Retry both, since extraction is non-deterministic and a second attempt may succeed where the first failed"},{"id":"D","text":"Skip retries for both and flag them together for human review to be safe"}],"correctOptions":["B"],"matches":[],"match_pairs":[],"explanation":"Retries fix format, structural, and arithmetic errors (A's missed line item is fixable). Retries cannot create information absent from the source (B's department does not exist). Identify fixability before retrying; flag unfixable cases for human review or return `null`.","helpful_links":[],"tags":[]},{"id":"e28b782b-8b20-4c2f-9c93-564ba107de6c","original_quiz_id":"6a1a0c49ca96069990f7e8bc","created":"$D2026-06-03T18:39:05.747Z","description":"","questionContent":"Self-correction schema fields","questionImage":null,"type":"single","options":[{"id":"A","text":"Only `stated_total`, taken directly from the figure printed on the document"},{"id":"B","text":"Both `calculated_total` and `stated_total` plus a discrepancy flag"},{"id":"C","text":"A single combined `total` field that the model fills in after reconciling the line items"},{"id":"D","text":"A confidence score on the total, so low-confidence totals can be routed to review"}],"correctOptions":["B"],"matches":[],"match_pairs":[],"explanation":"Extracting both the model-computed `calculated_total` and the document's `stated_total` makes a mismatch an automatic discrepancy flag without external logic. `conflict_detected` booleans similarly surface contradictory source information.","helpful_links":[],"tags":[]},{"id":"5fa52291-5c01-41cf-bd5e-ee4c9170b372","original_quiz_id":"6a1a0c49ca96069990f7e8bc","created":"$D2026-06-03T18:39:05.747Z","description":"Extraction works on structured tables but returns empty fields when the same data appears in narrative paragraphs; detailed instructions already specify all fields.","questionContent":"Few-shot examples for narrative extraction","questionImage":null,"type":"single","options":[{"id":"A","text":"Increase the context window so the model can hold more of each document while extracting"},{"id":"B","text":"Add few-shot examples showing extraction from both tables and narrative paragraphs"},{"id":"C","text":"Add a pre-processing step that converts all narrative text into structured tables before extraction"},{"id":"D","text":"Add a post-processing retry that re-extracts any field that came back empty on the first pass"}],"correctOptions":["B"],"matches":[],"match_pairs":[],"explanation":"The data exists but in an unexpected structure — the classic trigger for few-shot examples covering the failing structure. More instructions will not help; examples teach the model to handle structural variety and reduce empty fields.","helpful_links":[],"tags":[]},{"id":"6acdc555-ce28-4a13-89ee-2737c05325cb","original_quiz_id":"6a1a0c49ca96069990f7e8bc","created":"$D2026-06-03T18:39:05.747Z","description":"","questionContent":"Constructing effective few-shot examples","questionImage":null,"type":"single","options":[{"id":"A","text":"Use ten or more examples showing only the input and the expected output for each"},{"id":"B","text":"Use 2–4 targeted examples that include the reasoning for each decision"},{"id":"C","text":"Use a single representative example with no reasoning, to keep the prompt short"},{"id":"D","text":"Use examples that cover only the happy path, since edge cases are rare in practice"}],"correctOptions":["B"],"matches":[],"match_pairs":[],"explanation":"Use 2–4 targeted examples covering the failing scenarios, and include reasoning for each decision — not just input/output. Reasoning teaches the model the decision principle so it generalises to novel cases, rather than literal pattern-matching.","helpful_links":[],"tags":[]},{"id":"565c550a-372b-4e40-b0f2-53153695c304","original_quiz_id":"6a1a0c49ca96069990f7e8bc","created":"$D2026-06-03T18:39:05.747Z","description":"","questionContent":"Message Batches API facts","questionImage":null,"type":"single","options":[{"id":"A","text":"25% cost savings with a guaranteed one-hour processing SLA for every submitted batch"},{"id":"B","text":"50% savings, up to 24-hour window, no latency SLA, no multi-turn tool calling"},{"id":"C","text":"50% cost savings with a guaranteed four-hour delivery window for all batch results"},{"id":"D","text":"No cost savings, but noticeably faster processing than the synchronous API offers"}],"correctOptions":["B"],"matches":[],"match_pairs":[],"explanation":"The Message Batches API offers 50% savings, an up-to-24-hour window, no latency SLA, and no multi-turn tool calling within a single request. It uses `custom_id` fields to correlate requests with responses. Use it only for latency-tolerant workflows.","helpful_links":[],"tags":[]},{"id":"60727b97-d34a-4b3a-9898-1f2046ad19f5","original_quiz_id":"6a1a0c49ca96069990f7e8bc","created":"$D2026-06-03T18:39:05.747Z","description":"An organisation requires a 30-hour SLA for a report, given the 24-hour maximum batch window.","questionContent":"Batch SLA scheduling","questionImage":null,"type":"single","options":[{"id":"A","text":"Submit the batch right at the deadline and rely on the batch usually completing faster than 24 hours"},{"id":"B","text":"Submit at least 30 hours before the deadline, with a 6-hour buffer, batching every 4–6 hours"},{"id":"C","text":"Submit the batch 24 hours after collecting all the requests for the reporting period"},{"id":"D","text":"Use only the real-time API, since the 24-hour window cannot fit inside a 30-hour SLA"}],"correctOptions":["B"],"matches":[],"match_pairs":[],"explanation":"30-hour SLA minus the 24-hour maximum window leaves a 6-hour buffer. Submit the final batch no later than 24 hours before the deadline, and submit every 4–6 hours within the buffer so a fresh batch is always in flight.","helpful_links":[],"tags":[]},{"id":"6e46dfcf-7588-403c-a38b-5310d16b5052","original_quiz_id":"6a1a0c49ca96069990f7e8bc","created":"$D2026-06-03T18:39:05.747Z","description":"","questionContent":"Batch failure handling","questionImage":null,"type":"single","options":[{"id":"A","text":"Resubmit the entire batch so that every document is processed again under the corrected configuration"},{"id":"B","text":"Identify failures by `custom_id` and resubmit only failures with modifications"},{"id":"C","text":"Discard the failed documents and accept the partial results from those that succeeded"},{"id":"D","text":"Switch the whole batch to the real-time API and reprocess every document synchronously"}],"correctOptions":["B"],"matches":[],"match_pairs":[],"explanation":"Parse results to find failed `custom_id` values, then resubmit only the failures with targeted modifications (chunk oversized docs, simplify prompts, add format-specific examples). Resubmitting the whole batch wastes cost on already-successful documents.","helpful_links":[],"tags":[]},{"id":"5b0f0e9f-1098-4666-a9ea-15d657dcea42","original_quiz_id":"6a1a0c49ca96069990f7e8bc","created":"$D2026-06-03T18:39:05.747Z","description":"A 14-file PR gets inconsistent depth, missed bugs, and contradictory findings (a pattern flagged in one file, approved in another).","questionContent":"Multi-pass review for a large PR","questionImage":null,"type":"single","options":[{"id":"A","text":"Switch to a larger-context model and review all 14 files together in one comprehensive pass"},{"id":"B","text":"Per-file local passes plus a separate cross-file integration pass"},{"id":"C","text":"Run three full-PR review passes and keep only the issues that appear in at least two of them"},{"id":"D","text":"Require developers to split large PRs into submissions of 3–4 files before review runs"}],"correctOptions":["B"],"matches":[],"match_pairs":[],"explanation":"This is attention dilution. Per-file local passes give each file full attention for consistent depth, and a cross-file integration pass catches data flow, API consistency, and contradictions. A larger window holds more text but does not equalise attention.","helpful_links":[],"tags":[]},{"id":"d5874f1e-37f8-477a-adb3-67fbb120afdc","original_quiz_id":"6a1a0c49ca96069990f7e8bc","created":"$D2026-06-03T18:39:05.747Z","description":"","questionContent":"Calibrated versus raw confidence for routing","questionImage":null,"type":"single","options":[{"id":"A","text":"Use the model's raw confidence scores directly to decide which findings get automated"},{"id":"B","text":"Calibrate confidence against labelled validation sets before using it for routing"},{"id":"C","text":"Do not use confidence at all, since it can never be made reliable enough for routing"},{"id":"D","text":"Treat confidence purely as developer-facing information and never wire it into routing logic"}],"correctOptions":["B"],"matches":[],"match_pairs":[],"explanation":"Raw self-reported confidence is poorly calibrated — 0.90 on dates may mean 94% accuracy while 0.90 on amounts means 82%. Calibrate against labelled ground-truth data first, then route high-confidence to automation and low-confidence to human review.","helpful_links":[],"tags":[]},{"id":"c6658b8d-8670-4fb2-b28e-56b40aa85ebb","original_quiz_id":"6a1a0c49ca96069990f7e8bc","created":"$D2026-06-03T18:39:05.747Z","description":"In a support session, history is summarised between turns. \"I'd like a refund of $247.83 for order #8891 placed on March 3rd\" becomes \"Customer wants a refund for a recent order.\"","questionContent":"Progressive summarisation trap","questionImage":null,"type":"single","options":[{"id":"A","text":"Increase the context window so summarisation between turns is no longer necessary at all"},{"id":"B","text":"Extract transactional facts into a persistent case-facts block included in every prompt, outside summarised history"},{"id":"C","text":"Instruct the model to always preserve numerical values, dates, and identifiers when it summarises"},{"id":"D","text":"Store the full conversation in an external store and retrieve the relevant turns on demand"}],"correctOptions":["B"],"matches":[],"match_pairs":[],"explanation":"Summarisation systematically destroys amounts, dates, and identifiers. The persistent case-facts block holds transactional facts outside the summarised history and is included in every prompt, never summarised — the single most important context-management pattern.","helpful_links":[],"tags":[]},{"id":"af1ed851-5900-4e2c-9916-ab50b227c289","original_quiz_id":"6a1a0c49ca96069990f7e8bc","created":"$D2026-06-03T18:39:05.747Z","description":"","questionContent":"The lost-in-the-middle effect","questionImage":null,"type":"single","options":[{"id":"A","text":"Add an instruction telling the model to pay equal attention to every part of the input"},{"id":"B","text":"Place key findings at the beginning of aggregated inputs and use explicit section headers"},{"id":"C","text":"Move all of the important content to the very end of the input where attention is strongest"},{"id":"D","text":"Switch to a larger model that can attend to the middle of long inputs more reliably"}],"correctOptions":["B"],"matches":[],"match_pairs":[],"explanation":"Models attend reliably to the beginning and end of long inputs, missing the middle. The fix is structural — put a key-findings summary at the start and use explicit section headers — not a prompt reminder, which is unreliable for position effects.","helpful_links":[],"tags":[]},{"id":"039e12c9-282f-4ea2-a578-51e377accd3a","original_quiz_id":"6a1a0c49ca96069990f7e8bc","created":"$D2026-06-03T18:39:05.747Z","description":"An order lookup returns 40+ fields, but only 5 are relevant to the refund task.","questionContent":"Tool result trimming","questionImage":null,"type":"single","options":[{"id":"A","text":"Keep all 40 fields in context in case some of them turn out to be needed later in the conversation"},{"id":"B","text":"Trim to the relevant fields before the result enters conversation history"},{"id":"C","text":"Summarise the full 40-field result into a prose paragraph and store that summary instead"},{"id":"D","text":"Store all 40 fields in a database and reference them by key when any field is needed"}],"correctOptions":["B"],"matches":[],"match_pairs":[],"explanation":"Untrimmed results consume tokens on every subsequent turn as history grows. Trim to relevant fields before the result enters history (ideally in a PostToolUse hook). Once verbose data is in context, it stays there for the rest of the conversation.","helpful_links":[],"tags":[]},{"id":"d029df61-9d77-4a27-8d27-34d631b345a6","original_quiz_id":"6a1a0c49ca96069990f7e8bc","created":"$D2026-06-03T18:39:05.747Z","description":"","questionContent":"The Claude API is stateless","questionImage":null,"type":"single","options":[{"id":"A","text":"The server stores session state between calls, so older messages can safely be omitted"},{"id":"B","text":"Each request must include the complete conversation history"},{"id":"C","text":"Conversation history becomes optional once the first turn has established the session context"},{"id":"D","text":"Only the most recent message needs to be sent, since the server reconstructs the rest"}],"correctOptions":["B"],"matches":[],"match_pairs":[],"explanation":"The Claude API is stateless — each request must include the full conversation history for coherence. Selectively truncating breaks continuity. Manage growth with persistent case-facts blocks and summarisation, not truncation.","helpful_links":[],"tags":[]},{"id":"c3f23c21-a617-4c19-8a5c-e14393311c92","original_quiz_id":"6a1a0c49ca96069990f7e8bc","created":"$D2026-06-03T18:39:05.747Z","description":"","questionContent":"Three valid escalation triggers","questionImage":null,"type":"single","options":[{"id":"A","text":"Negative customer sentiment, a low self-reported confidence score, and an unusually long conversation"},{"id":"B","text":"Explicit human request, policy gaps, and inability to make progress"},{"id":"C","text":"Any refund above $500, requests received on weekends, and customers flagged as VIP"},{"id":"D","text":"Tool errors, slow response times, and the customer repeating the same question"}],"correctOptions":["B"],"matches":[],"match_pairs":[],"explanation":"The three valid triggers are: (1) the customer explicitly requests a human (honour immediately), (2) a policy gap (the policy is silent — distinct from a violation, which has a documented \"no\"), and (3) inability to make progress after actually attempting resolution.","helpful_links":[],"tags":[]},{"id":"b182bc94-ebca-4fca-9779-6ee8b01ec964","original_quiz_id":"6a1a0c49ca96069990f7e8bc","created":"$D2026-06-03T18:39:05.747Z","description":"A support agent hits 55% first-contact resolution: it escalates simple damage replacements while attempting complex policy exceptions.","questionContent":"Unreliable escalation triggers","questionImage":null,"type":"single","options":[{"id":"A","text":"Add sentiment analysis that automatically escalates whenever detected negative sentiment crosses a threshold"},{"id":"B","text":"Add explicit escalation criteria with few-shot examples to the system prompt"},{"id":"C","text":"Have the agent emit a self-reported confidence score and route to a human when it falls below a threshold"},{"id":"D","text":"Train a separate classifier model on historical tickets to predict which requests need escalation"}],"correctOptions":["B"],"matches":[],"match_pairs":[],"explanation":"The miscalibration is exactly the self-confidence failure mode. Explicit criteria with examples is the proportionate, prompt-level first fix. Sentiment (frustration does not correlate with complexity) and raw confidence (poorly calibrated) are the two unreliable triggers.","helpful_links":[],"tags":[]},{"id":"f0f306c3-9b75-4913-923e-c53748b54615","original_quiz_id":"6a1a0c49ca96069990f7e8bc","created":"$D2026-06-03T18:39:05.747Z","description":"A frustrated customer has a straightforward, resolvable issue and has not asked for a human.","questionContent":"The frustration nuance","questionImage":null,"type":"single","options":[{"id":"A","text":"Escalate to a human immediately because the customer is visibly frustrated with the situation"},{"id":"B","text":"Acknowledge the frustration and offer the resolution directly"},{"id":"C","text":"Run a sentiment analysis pass first and let the detected sentiment score decide whether to escalate"},{"id":"D","text":"Ask the customer to calm down before continuing to work on resolving the issue"}],"correctOptions":["B"],"matches":[],"match_pairs":[],"explanation":"A frustrated customer with a resolvable issue should be helped: acknowledge the frustration and offer the resolution. Escalate immediately only when the customer explicitly asks for a human, or reiterates that preference after being offered help.","helpful_links":[],"tags":[]},{"id":"c905a395-c5d7-4dfa-9b85-133a01d6a5ab","original_quiz_id":"6a1a0c49ca96069990f7e8bc","created":"$D2026-06-03T18:39:05.747Z","description":"A name search returns three \"John Smith\" records.","questionContent":"Ambiguous customer matching","questionImage":null,"type":"single","options":[{"id":"A","text":"Select the most recently created record, on the assumption it reflects the active customer"},{"id":"B","text":"Ask for an additional identifier (email, phone, order number)"},{"id":"C","text":"Select the most active record, since it is the most likely to be the one the customer means"},{"id":"D","text":"Pick one of the three records at random and proceed, correcting later if it turns out wrong"}],"correctOptions":["B"],"matches":[],"match_pairs":[],"explanation":"Heuristic selection (most recent/active) risks privacy violations and wrong-account actions. The only safe response to multiple matches is to ask for additional identifiers to disambiguate.","helpful_links":[],"tags":[]},{"id":"e5049ec9-6968-4ce4-97c8-e84eeb764b23","original_quiz_id":"6a1a0c49ca96069990f7e8bc","created":"$D2026-06-03T18:39:05.747Z","description":"","questionContent":"Policy gap versus policy violation","questionImage":null,"type":"single","options":[{"id":"A","text":"They are effectively the same situation and both should be escalated to a human"},{"id":"B","text":"A gap means the policy is silent (escalate); a violation has a documented answer (do not escalate)"},{"id":"C","text":"A violation means the policy is silent on the situation, while a gap has a clear documented answer"},{"id":"D","text":"Both a gap and a violation have documented answers, so neither requires escalation"}],"correctOptions":["B"],"matches":[],"match_pairs":[],"explanation":"A policy violation (e.g., refund outside the return window) has a documented answer (\"no\") and does not require escalation. A policy gap means the policy is silent on the situation, requiring human judgement — so it warrants escalation.","helpful_links":[],"tags":[]},{"id":"93fa88f9-ccbc-4bcc-ba2d-443188d79783","original_quiz_id":"6a1a0c49ca96069990f7e8bc","created":"$D2026-06-03T18:39:05.747Z","description":"","questionContent":"Structured error context elements","questionImage":null,"type":"single","options":[{"id":"A","text":"Just the error message text, so the coordinator can display it for debugging"},{"id":"B","text":"Failure type, what was attempted, partial results, and potential alternative approaches"},{"id":"C","text":"Only a boolean retry flag indicating whether the operation should be attempted again"},{"id":"D","text":"The full conversation history of the subagent so the coordinator can replay its reasoning"}],"correctOptions":["B"],"matches":[],"match_pairs":[],"explanation":"A failed subagent must return structured error context with four elements: failure type, what was attempted (query, parameters, target), partial results gathered before failure, and potential alternatives. This lets the coordinator choose to retry, try an alternative, proceed with partials, or escalate.","helpful_links":[],"tags":[]},{"id":"1a0742a4-5f05-4c35-b9f0-7e4cf09d22ea","original_quiz_id":"6a1a0c49ca96069990f7e8bc","created":"$D2026-06-03T18:39:05.747Z","description":"A subagent times out but returns `{ \"results\": [], \"status\": \"success\" }`.","questionContent":"Silent suppression anti-pattern","questionImage":null,"type":"single","options":[{"id":"A","text":"This is acceptable, because returning a success keeps the overall pipeline running smoothly"},{"id":"B","text":"This is silent suppression — the coordinator believes the search found nothing and never recovers"},{"id":"C","text":"This is the correct way to handle a timeout, since an empty result is a safe default to return"},{"id":"D","text":"This automatically triggers a retry, because an empty success is treated as a recoverable condition"}],"correctOptions":["B"],"matches":[],"match_pairs":[],"explanation":"Silent suppression is the worst anti-pattern: marking a failure as success makes the coordinator believe the search ran and found nothing, so it never retries or tries alternatives and ships a report with invisible gaps. Return structured error context instead.","helpful_links":[],"tags":[]},{"id":"3977e59c-f7f9-4af2-bfa5-4fef11345a84","original_quiz_id":"6a1a0c49ca96069990f7e8bc","created":"$D2026-06-03T18:39:05.747Z","description":"One of five subagents times out and the entire research pipeline crashes, discarding the four completed results.","questionContent":"Workflow termination anti-pattern","questionImage":null,"type":"single","options":[{"id":"A","text":"This is the safest possible failure mode, because it guarantees no partial output is ever shipped"},{"id":"B","text":"This is workflow termination — it wastes completed work; the coordinator should assess and recover with partial results"},{"id":"C","text":"This prevents inconsistent output, since a partial result set would mislead the consumer"},{"id":"D","text":"This is required for data integrity, as mixing completed and failed results corrupts the report"}],"correctOptions":["B"],"matches":[],"match_pairs":[],"explanation":"Workflow termination kills the whole pipeline on one failure, discarding successful subagents' work. The correct middle ground is structured error propagation: the failing subagent reports, the coordinator assesses, and the system continues with partial results or targeted recovery.","helpful_links":[],"tags":[]},{"id":"8555fab0-5eea-401d-ab93-1a5f5b481071","original_quiz_id":"6a1a0c49ca96069990f7e8bc","created":"$D2026-06-03T18:39:05.747Z","description":"","questionContent":"Coverage annotations","questionImage":null,"type":"single","options":[{"id":"A","text":"Silently omit any topic that could not be researched, so the report reads as complete"},{"id":"B","text":"Note explicitly which areas are limited due to unavailable sources"},{"id":"C","text":"Average the available topics together to fill in the gaps left by the missing one"},{"id":"D","text":"Flag the entire report as failed whenever any single topic could not be researched"}],"correctOptions":["B"],"matches":[],"match_pairs":[],"explanation":"Coverage annotations make gaps explicit (\"Section on geothermal is limited due to unavailable journal access\"). Without them, a gap looks like the topic was irrelevant rather than the source being unavailable. Transparency beats silent omission.","helpful_links":[],"tags":[]},{"id":"618faeb9-0600-4fc6-91d9-8524dba48e3b","original_quiz_id":"6a1a0c49ca96069990f7e8bc","created":"$D2026-06-03T18:39:05.747Z","description":"After investigating several modules, an agent starts referencing \"typical repository patterns\" instead of the specific classes and dependency chains it discovered earlier.","questionContent":"Context degradation in long exploration","questionImage":null,"type":"single","options":[{"id":"A","text":"Increase the context window so the verbose discovery output no longer crowds out earlier findings"},{"id":"B","text":"Maintain scratchpad files of key findings and reference them for later questions"},{"id":"C","text":"Restart the session from scratch and instruct the agent to explore more efficiently this time"},{"id":"D","text":"Pre-load the entire codebase structure into context before the exploration begins"}],"correctOptions":["B"],"matches":[],"match_pairs":[],"explanation":"Context degradation is an attention-quality problem, not a token-limit one — a larger window still fills with verbose output. Scratchpad files persist specific findings (class names, file paths, dependency chains) outside the conversation context, making them immune to degradation.","helpful_links":[],"tags":[]},{"id":"cbacbc24-16fd-4fe2-972f-1faf9bc0ba8b","original_quiz_id":"6a1a0c49ca96069990f7e8bc","created":"$D2026-06-03T18:39:05.747Z","description":"","questionContent":"Subagent delegation for codebase exploration","questionImage":null,"type":"single","options":[{"id":"A","text":"Its only real benefit is parallelisation, letting several explorations run at the same time"},{"id":"B","text":"Its primary benefit is context isolation, keeping the main context clean"},{"id":"C","text":"It increases the main agent's context size by merging each subagent's full output back in"},{"id":"D","text":"It is useful only for crash recovery, not for ordinary exploration of a codebase"}],"correctOptions":["B"],"matches":[],"match_pairs":[],"explanation":"For codebase exploration, the primary benefit of subagent delegation is context isolation — subagents explore verbosely in their own context and return structured summaries, keeping the main agent's context clean for high-level coordination.","helpful_links":[],"tags":[]},{"id":"b23004e5-e77e-471f-8331-5dfcc43d746e","original_quiz_id":"6a1a0c49ca96069990f7e8bc","created":"$D2026-06-03T18:39:05.747Z","description":"","questionContent":"The /compact command","questionImage":null,"type":"single","options":[{"id":"A","text":"Use it only as a last resort once the session has actually run out of context space"},{"id":"B","text":"Use it proactively during extended sessions to maintain context quality"},{"id":"C","text":"It permanently deletes the conversation history to reclaim the full context window"},{"id":"D","text":"It loads the project's configuration files into the current session's context"}],"correctOptions":["B"],"matches":[],"match_pairs":[],"explanation":"`/compact` summarises the conversation to free space while preserving key information. Use it proactively during extended exploration to maintain context quality, not just as a last resort when context is exhausted.","helpful_links":[],"tags":[]},{"id":"a7b3affd-8235-46fb-9acb-61f7d255ba05","original_quiz_id":"6a1a0c49ca96069990f7e8bc","created":"$D2026-06-03T18:39:05.747Z","description":"","questionContent":"Crash recovery via state manifests","questionImage":null,"type":"single","options":[{"id":"A","text":"Restart the exploration from scratch after a crash and re-derive all the earlier findings"},{"id":"B","text":"Export a structured manifest (explored paths, key findings, next steps) the coordinator reloads on resume"},{"id":"C","text":"Rely on the API server to restore the session state automatically after a crash"},{"id":"D","text":"Keep all of the exploration progress only in the context window and accept the loss on a crash"}],"correctOptions":["B"],"matches":[],"match_pairs":[],"explanation":"Extended sessions can crash or exhaust context. A structured state manifest records explored paths, key findings, the current phase, and next steps. On resume, the coordinator loads it and injects it into prompts so exploration continues from the last checkpoint.","helpful_links":[],"tags":[]},{"id":"823ac1b7-1ba7-438d-8fb9-d6f8d54fae72","original_quiz_id":"6a1a0c49ca96069990f7e8bc","created":"$D2026-06-03T18:39:05.747Z","description":"An extraction system reports 97% overall accuracy. The team proposes automating all extractions above 95% confidence.","questionContent":"The aggregate metrics trap","questionImage":null,"type":"single","options":[{"id":"A","text":"Raise the automation threshold to 99% confidence so only the strongest extractions are automated"},{"id":"B","text":"Aggregate accuracy can mask poor per-type/field performance, and confidence needs calibration against labelled sets first"},{"id":"C","text":"The model will drift toward overconfidence over time and will eventually need to be retrained"},{"id":"D","text":"All extractions should be human-reviewed regardless of confidence, making the proposal unworkable"}],"correctOptions":["B"],"matches":[],"match_pairs":[],"explanation":"A 97% aggregate can hide catastrophic per-segment rates (60% on handwritten receipts, 45% on international formats) because high-volume segments dominate. Always validate by document type AND field segment, and calibrate confidence before automating.","helpful_links":[],"tags":[]},{"id":"76ee7f2e-aa86-4acc-b33e-40b1b58493d1","original_quiz_id":"6a1a0c49ca96069990f7e8bc","created":"$D2026-06-03T18:39:05.747Z","description":"","questionContent":"Stratified random sampling scope","questionImage":null,"type":"single","options":[{"id":"A","text":"Sample only the low-confidence extractions, since those are the ones most likely to be wrong"},{"id":"B","text":"Sample high-confidence (automated) extractions too, to catch novel error patterns"},{"id":"C","text":"Sample only the first 100 documents processed, as a representative slice of the workload"},{"id":"D","text":"Skip sampling entirely whenever the aggregate accuracy figure is already high"}],"correctOptions":["B"],"matches":[],"match_pairs":[],"explanation":"Low-confidence items are already human-reviewed. If a novel error pattern affects automated high-confidence extractions, only stratified random sampling of high-confidence items will detect it. Sampling only low-confidence items leaves automated outputs blind.","helpful_links":[],"tags":[]},{"id":"cc041c9b-23a8-40f6-8b05-0c35e02e9a08","original_quiz_id":"6a1a0c49ca96069990f7e8bc","created":"$D2026-06-03T18:39:05.747Z","description":"","questionContent":"Reviewer capacity prioritisation","questionImage":null,"type":"single","options":[{"id":"A","text":"Spread reviewer capacity evenly across all extractions so every item gets equal attention"},{"id":"B","text":"Prioritise the highest-uncertainty items, with a dynamically reordered queue"},{"id":"C","text":"Review the items strictly in chronological order of when they were extracted"},{"id":"D","text":"Review only the highest-confidence items, since those are the ones being automated"}],"correctOptions":["B"],"matches":[],"match_pairs":[],"explanation":"Human reviewers are limited. Even distribution wastes time on high-confidence items the model handles well. Prioritise the highest-uncertainty items, ordering the queue dynamically so each freed reviewer gets the most uncertain remaining item — not the next chronological one.","helpful_links":[],"tags":[]},{"id":"637a960c-49f4-4709-9061-987d3e227a34","original_quiz_id":"6a1a0c49ca96069990f7e8bc","created":"$D2026-06-03T18:39:05.747Z","description":"","questionContent":"Structured claim-source mapping fields","questionImage":null,"type":"single","options":[{"id":"A","text":"Just the claim text, on the assumption the source can be looked up later if needed"},{"id":"B","text":"Claim, source URL, document name, relevant excerpt, and publication date"},{"id":"C","text":"The claim together with a confidence score expressing how strongly it is supported"},{"id":"D","text":"The claim together with the name of the subagent that originally retrieved it"}],"correctOptions":["B"],"matches":[],"match_pairs":[],"explanation":"Every finding must carry full provenance: the claim, source URL, document name, relevant excerpt, and publication date. This is the structural guarantee that the final output traces back to specific sources.","helpful_links":[],"tags":[]},{"id":"a829ab57-2342-48ca-a87c-3240ccae9a44","original_quiz_id":"6a1a0c49ca96069990f7e8bc","created":"$D2026-06-03T18:39:05.747Z","description":"","questionContent":"Where attribution most often dies","questionImage":null,"type":"single","options":[{"id":"A","text":"During initial retrieval, when subagents fetch sources but fail to record their identifiers"},{"id":"B","text":"During the synthesis step, when paraphrasing and compressing"},{"id":"C","text":"During tool execution, when the tool returns results without attaching source metadata"},{"id":"D","text":"During schema validation, when invalid attribution fields are stripped from the output"}],"correctOptions":["B"],"matches":[],"match_pairs":[],"explanation":"Attribution dies during summarisation/synthesis: as the synthesis agent compresses and paraphrases, claim-source mappings vanish unless it is explicitly instructed to preserve and merge them. The synthesis step is the most common loss point.","helpful_links":[],"tags":[]},{"id":"3f402c4a-dcd3-44e0-a786-1708aac2c751","original_quiz_id":"6a1a0c49ca96069990f7e8bc","created":"$D2026-06-03T18:39:05.747Z","description":"Source A reports 12% growth (2023 data); Source B reports 8% growth (2024 data). The synthesis agent currently picks the more recent value.","questionContent":"Handling two conflicting credible sources","questionImage":null,"type":"single","options":[{"id":"A","text":"Always use the most recent source, since newer data should supersede older figures"},{"id":"B","text":"Annotate both values with source attribution and publication dates, letting the consumer decide"},{"id":"C","text":"Average the two values and report 10% growth together with a note about the variance"},{"id":"D","text":"Escalate the conflict to a human reviewer before either value is included in the report"}],"correctOptions":["B"],"matches":[],"match_pairs":[],"explanation":"Arbitrarily selecting or averaging destroys information and creates false certainty. Annotate both values with full attribution and dates. Different publication dates may explain the difference as a trend rather than a contradiction.","helpful_links":[],"tags":[]},{"id":"4e24dccd-3d8b-4cc2-9b8c-a950894b4a41","original_quiz_id":"6a1a0c49ca96069990f7e8bc","created":"$D2026-06-03T18:39:05.747Z","description":"Source A (2023) reports 8% growth; Source B (2024) reports 12% growth.","questionContent":"Temporal awareness in synthesis","questionImage":null,"type":"single","options":[{"id":"A","text":"This is a data-quality contradiction that should be suppressed so the report stays consistent"},{"id":"B","text":"With dates, it likely reflects a trend, not a contradiction; require dates in all outputs"},{"id":"C","text":"Average the two figures to 10% so the report presents a single reconciled growth rate"},{"id":"D","text":"Pick the higher figure, since the more optimistic estimate is usually the more recent one"}],"correctOptions":["B"],"matches":[],"match_pairs":[],"explanation":"Without dates these look contradictory; with dates they tell a story of growth accelerating from 8% to 12%. Require publication/data-collection dates in structured outputs so valid trends are not misread as data-quality issues.","helpful_links":[],"tags":[]},{"id":"aca2f3f1-7e75-4105-acaf-1eda2f8d991f","original_quiz_id":"6a1a0c49ca96069990f7e8bc","created":"$D2026-06-03T18:39:05.747Z","description":"","questionContent":"Content-appropriate rendering","questionImage":null,"type":"single","options":[{"id":"A","text":"Render everything as prose for consistency, so every section of the report reads the same way"},{"id":"B","text":"Financial data as tables, news as prose, technical findings as structured lists"},{"id":"C","text":"Render everything as tables, since tabular layout is the most compact and scannable format"},{"id":"D","text":"Render everything as bullet lists so each finding is broken into discrete, digestible points"}],"correctOptions":["B"],"matches":[],"match_pairs":[],"explanation":"Different content demands different formats: financial data → tables (compare values, spot patterns), news → prose (narrative, cause-and-effect), technical findings → lists (clear hierarchy). Forcing one uniform format degrades readability.","helpful_links":[],"tags":[]},{"id":"a915f269-a0a2-4589-a345-ae2bc6f78077","original_quiz_id":"6a1a0c49ca96069990f7e8bc","created":"$D2026-06-03T18:39:05.747Z","description":"A document analysis agent finds annual revenue stated as $4.2M (audited annual report) and $3.8M (preliminary SEC filing).","questionContent":"Completing analysis with conflicts intact","questionImage":null,"type":"single","options":[{"id":"A","text":"Pick the audited annual-report figure and discard the preliminary one as less reliable"},{"id":"B","text":"Complete the analysis with both values annotated and a possible explanation, leaving resolution to the coordinator"},{"id":"C","text":"Average the two figures to $4.0M and report that as the reconciled annual revenue"},{"id":"D","text":"Stop and escalate to a human before completing the rest of the analysis"}],"correctOptions":["B"],"matches":[],"match_pairs":[],"explanation":"The analysis agent should complete its work with conflicts intact and annotated (both values, sources, context, possible explanation — audited vs preliminary, fiscal vs calendar). Resolution belongs to the coordinator or consumer, not the analysis agent.","helpful_links":[],"tags":[]},{"id":"e186a5bc-c764-4136-b47f-d5b691dd61e2","original_quiz_id":"6a1a0c49ca96069990f7e8bc","created":"$D2026-06-03T18:39:05.747Z","description":"","questionContent":"Local recovery before propagation","questionImage":null,"type":"single","options":[{"id":"A","text":"Propagate every transient failure straight to the coordinator so it has full visibility into all errors"},{"id":"B","text":"Subagents implement local retry for transient failures and propagate only what they cannot resolve"},{"id":"C","text":"Route all retry logic through the coordinator so retry policy stays consistent across subagents"},{"id":"D","text":"Discard transient failures silently and continue, since they usually clear on their own"}],"correctOptions":["B"],"matches":[],"match_pairs":[],"explanation":"Subagents should implement local recovery (retry, fallback sources) for transient failures before escalating. Only persistent failures propagate, always with what was attempted and any partial results. This keeps the coordinator from managing every transient failure.","helpful_links":[],"tags":[]},{"id":"80f33a07-dfc1-4092-a0a6-069609457445","original_quiz_id":"6a1a0c49ca96069990f7e8bc","created":"$D2026-06-03T18:39:05.747Z","description":"A research subagent sends its full reasoning chain and raw content to a synthesis agent with a limited context budget.","questionContent":"Upstream agent output optimisation","questionImage":null,"type":"single","options":[{"id":"A","text":"Keep the verbose reasoning so the synthesis agent has the fullest possible context to work from"},{"id":"B","text":"Modify upstream agents to return structured data (key facts, citations, relevance scores) instead of verbose reasoning"},{"id":"C","text":"Increase the synthesis agent's context window so it can absorb the full reasoning chain without trimming"},{"id":"D","text":"Run a summarisation pass after synthesis to compress the output once everything is combined"}],"correctOptions":["B"],"matches":[],"match_pairs":[],"explanation":"Upstream agents should return structured data — key facts, citations, relevance scores, metadata — not verbose reasoning chains the downstream agent cannot use. This saves tokens and lets downstream agents process findings without re-parsing prose.","helpful_links":[],"tags":[]},{"id":"7e1258ca-68ce-47d8-aafa-28507346d1cd","original_quiz_id":"6a1a0c49ca96069990f7e8bc","created":"$D2026-06-03T18:39:05.747Z","description":"","questionContent":"Domain weightings and study focus","questionImage":null,"type":"single","options":[{"id":"A","text":"All five domains are weighted equally at 20% each, so study time should be split evenly"},{"id":"B","text":"Domain 1 (Agentic Architecture) is the heaviest at 27%"},{"id":"C","text":"Domain 5 (Context Management) is the heaviest and should receive the most study time"},{"id":"D","text":"Domain 2 (Tool Design & MCP) is the heaviest at 27% of the exam"}],"correctOptions":["B"],"matches":[],"match_pairs":[],"explanation":"The weightings are: Domain 1 Agentic Architecture 27%, Domain 3 Claude Code 20%, Domain 4 Prompt Engineering 20%, Domain 2 Tool Design & MCP 18%, Domain 5 Context Management 15%. Domain 1 deserves the most study time.","helpful_links":[],"tags":[]},{"id":"e4d5e839-89c3-4114-b581-0b8a8911fde4","original_quiz_id":"6a1a0c49ca96069990f7e8bc","created":"$D2026-06-03T18:39:05.747Z","description":"","questionContent":"The exam's recurring enforcement principle","questionImage":null,"type":"single","options":[{"id":"A","text":"Always prefer the cheapest solution that plausibly addresses the symptom being reported"},{"id":"B","text":"For high-stakes rules where a single failure causes loss, use deterministic enforcement, not stronger prompts"},{"id":"C","text":"Always reach for the largest and most capable model available when reliability matters"},{"id":"D","text":"Always escalate high-stakes decisions to a human rather than enforcing them programmatically"}],"correctOptions":["B"],"matches":[],"match_pairs":[],"explanation":"A cross-domain principle: when a single failure causes financial loss, a security breach, or a compliance violation, the answer is a deterministic mechanism (gate, hook, settings.json) — never a stronger prompt, more examples, or a bigger model. This recurs in 1.4, 1.5, and 3.1.","helpful_links":[],"tags":[]},{"id":"24ba6389-fe83-4f15-899b-26851105adcc","original_quiz_id":"6a1a0c49ca96069990f7e8bc","created":"$D2026-06-03T18:39:05.747Z","description":"Several questions offer \"switch to a larger context window / a more capable model\" as a fix.","questionContent":"Common distractor family to reject","questionImage":null,"type":"single","options":[{"id":"A","text":"It is the correct fix for attention dilution and context degradation, which stem from limited capacity"},{"id":"B","text":"It is usually wrong for structural problems like attention dilution and context degradation"},{"id":"C","text":"It is the recommended remedy for fabrication of values that are missing from the source document"},{"id":"D","text":"It reliably improves tool selection by giving the model more room to consider each option"}],"correctOptions":["B"],"matches":[],"match_pairs":[],"explanation":"\"Use a bigger model/context window\" is a recurring wrong distractor for structural problems. Attention dilution needs multi-pass architecture; context degradation needs scratchpad files and delegation; fabrication needs nullable schemas; tool selection needs better descriptions. None is solved by more capacity.","helpful_links":[],"tags":[]}],"id":"6a1a0c49ca96069990f7e8bc","name":"Claude Agentic Architecture: Notes + 113 Practice Questions","description":"","owner_id":"679481b34625e2e86cd008d3","organizationId":null,"created":"$D2026-05-29T21:59:37.816Z","cover_image":"/cover.jpg","topics":[],"content_use":null,"editors":[],"private":true,"likedBy":[],"studyMaterial":"$32","studyMaterialSummary":null,"materialFiles":[{"id":"4d286ad4-7952-4d7e-affd-0c03b712e792","name":"Source & Reference Links","url":"","type":"markdown","source":"markdown","data":"$33"},{"id":"65b658c5-9386-4b76-8d41-917def7e23cb","name":"Cheat Sheet","url":"","type":"markdown","source":"markdown","data":"$34"},{"id":"cf59c427-bac2-460c-87b8-ea7d669c68b6","name":"Distractor Families (Wrong Answers That Recur)","url":"","type":"markdown","source":"markdown","data":"$35"}],"materialLocked":true,"materialOnly":null,"allowedEmails":[],"password":null,"startDate":null,"endDate":null,"examDuration":null,"maxAttempts":null,"allowGoBack":null,"viewGradeBehavior":null,"showAttempts":null,"allowStudyMaterial":null,"last_updated":"$D2026-09-03T00:40:35.944Z","forSale":false,"post":null,"price":199,"transferTarget":null,"transferSenderId":null,"organization":null,"exam_editors":[]}}},"children":["$L36","$L37"]}] 38:I[525940,["/_next/static/chunks/33t46atd3n2zd.js","/_next/static/chunks/1uoo0zstz-tka.js","/_next/static/chunks/1gmtbr53vpwcw.js","/_next/static/chunks/0hwl0k3t96f8i.js","/_next/static/chunks/07dyy4_l8-v8w.js","/_next/static/chunks/1a28s5qpw8ubo.js","/_next/static/chunks/3nx9-unhdiib7.js","/_next/static/chunks/1l3wx9j9z2da9.js","/_next/static/chunks/10prs2o5tycig.js","/_next/static/chunks/02ocli407gubc.js","/_next/static/chunks/2r--443asccvg.js","/_next/static/chunks/0rjt9gb10d05i.js","/_next/static/chunks/3lr20ikf1jmri.js","/_next/static/chunks/2mak3tsoa7kik.js","/_next/static/chunks/39kizkc9pkpb1.js","/_next/static/chunks/2uicbuh3o1h8o.js","/_next/static/chunks/1e-wsdiuv_p_u.js","/_next/static/chunks/12syjpiwnqlnq.js","/_next/static/chunks/0dnfb3y43dtv8.js","/_next/static/chunks/1x1n57nb1d8gg.js"],"EmbedThemeApplier"] 39:I[182210,["/_next/static/chunks/33t46atd3n2zd.js","/_next/static/chunks/1uoo0zstz-tka.js","/_next/static/chunks/1gmtbr53vpwcw.js","/_next/static/chunks/0hwl0k3t96f8i.js","/_next/static/chunks/07dyy4_l8-v8w.js","/_next/static/chunks/1a28s5qpw8ubo.js","/_next/static/chunks/3nx9-unhdiib7.js","/_next/static/chunks/1l3wx9j9z2da9.js","/_next/static/chunks/10prs2o5tycig.js","/_next/static/chunks/02ocli407gubc.js","/_next/static/chunks/2r--443asccvg.js","/_next/static/chunks/0rjt9gb10d05i.js","/_next/static/chunks/3lr20ikf1jmri.js","/_next/static/chunks/2mak3tsoa7kik.js","/_next/static/chunks/39kizkc9pkpb1.js","/_next/static/chunks/2uicbuh3o1h8o.js","/_next/static/chunks/1e-wsdiuv_p_u.js","/_next/static/chunks/12syjpiwnqlnq.js","/_next/static/chunks/0dnfb3y43dtv8.js","/_next/static/chunks/1x1n57nb1d8gg.js"],"ExamWidthWrapper"] 3a:I[27105,["/_next/static/chunks/33t46atd3n2zd.js","/_next/static/chunks/1uoo0zstz-tka.js","/_next/static/chunks/1gmtbr53vpwcw.js","/_next/static/chunks/0hwl0k3t96f8i.js","/_next/static/chunks/07dyy4_l8-v8w.js","/_next/static/chunks/1a28s5qpw8ubo.js","/_next/static/chunks/3nx9-unhdiib7.js","/_next/static/chunks/1l3wx9j9z2da9.js","/_next/static/chunks/10prs2o5tycig.js","/_next/static/chunks/02ocli407gubc.js","/_next/static/chunks/2r--443asccvg.js","/_next/static/chunks/0rjt9gb10d05i.js","/_next/static/chunks/3lr20ikf1jmri.js","/_next/static/chunks/2mak3tsoa7kik.js","/_next/static/chunks/39kizkc9pkpb1.js","/_next/static/chunks/2uicbuh3o1h8o.js","/_next/static/chunks/1e-wsdiuv_p_u.js","/_next/static/chunks/12syjpiwnqlnq.js","/_next/static/chunks/0dnfb3y43dtv8.js","/_next/static/chunks/1x1n57nb1d8gg.js"],"default"] 36:["$","$L38",null,{}] 37:["$","$L39",null,{"children":["$","$L3a",null,{"children":["$","$L3",null,{"parallelRouterKey":"children","error":"$undefined","errorStyles":"$undefined","errorScripts":"$undefined","template":["$","$L28",null,{"children":["$","$L4",null,{}]}],"templateStyles":[],"templateScripts":[["$","script","script-0",{"src":"/_next/static/chunks/0ye_k5ch2a1m0.js","async":true}],["$","script","script-1",{"src":"/_next/static/chunks/0vndyjagwio9s.js","async":true}]],"notFound":"$undefined","forbidden":"$undefined","unauthorized":"$undefined"}]}]}]