[{"data":1,"prerenderedAt":-1},["ShallowReactive",2],{"nav-notes-count":3,"feed-notes":4},9,[5,629,1094,1404,1779,2295,2673,2981,3224],{"id":6,"title":7,"author":8,"authorAvatar":9,"authorRole":10,"body":11,"category":597,"date":598,"description":149,"excerpt":599,"extension":600,"image":601,"index":602,"meta":603,"navigation":223,"path":617,"pullQuote":618,"readingTime":619,"seo":620,"slug":621,"stem":622,"takeaways":623,"total":627,"__hash__":628},"notes\u002Fnotes\u002Fa-practical-performance-review-for-a-nuxt-site.md","A Practical Performance Review for a Nuxt Site","Norbert Brett","\u002Fimages\u002Fnorbert-brett.jpg","Full-Stack & AI Systems Engineer",{"type":12,"value":13,"toc":586},"minimark",[14,19,33,103,117,124,128,136,143,373,377,387,445,452,459,463,575,582],[15,16,18],"h2",{"id":17},"why-a-single-lighthouse-score-never-tells-the-whole-story","Why a Single Lighthouse Score Never Tells the Whole Story",[20,21,22,23,27,28,32],"p",{},"Business owners and developers love sharing screenshots of four green ",[24,25,26],"code",{},"100"," circles from Google Lighthouse. Don't get me wrong—this portfolio maintains a ",[29,30,31],"strong",{},"98+ Lighthouse performance score",", and automated speed checks are a great baseline. But if you have ever visited a website that loaded its top banner in one second and then stuttered or lagged every time you scrolled down or clicked a price toggle, you have experienced what a single Lighthouse number misses.",[34,35,36,41],"blockquote",{},[20,37,38],{},[29,39,40],{},"In Plain English — Web Speed Jargon Decoded:",[42,43,44,51,73,85],"ul",{},[45,46,47,50],"li",{},[29,48,49],{},"Lighthouse",": Google's automated lab test that grades a web page from 0 to 100 on speed, accessibility, and SEO. Think of it like checking how fast a restaurant's front door opens—useful, but it stops measuring before you sit down and order.",[45,52,53,60,61,65,66,68,69,72],{},[29,54,55,56,59],{},"SSR (Server-Side Rendering) & Hydration (",[24,57,58],{},"__NUXT_DATA__",")",": To make your page show up immediately on phones and in Google search, the server sends a finished HTML page first (",[62,63,64],"em",{},"SSR","), along with a hidden data packet (",[24,67,58],{},") that wakes up the interactive buttons in the browser (",[62,70,71],{},"Hydration","). If that hidden packet is stuffed with unused data, every visitor downloads a heavy invisible backpack.",[45,74,75,78,79,84],{},[29,76,77],{},"CLS (Cumulative Layout Shift)",": That annoying jump when you are about to tap a button on your phone and a slow-loading font or image pushes the text down by half an inch. Our target is ",[29,80,81],{},[24,82,83],{},"0.00",".",[45,86,87,90,91,94,95,98,99,102],{},[29,88,89],{},"INP (Interaction to Next Paint)",": How many milliseconds pass between tapping a button (like switching ",[24,92,93],{},"HUF"," to ",[24,96,97],{},"EUR",") and seeing the screen update. Anything under ",[24,100,101],{},"50ms"," feels instantaneous to the human eye.",[20,104,105,106,109,110,113,114,116],{},"Months earlier, while reviewing a Nuxt application built in 2025 that boasted a ",[24,107,108],{},"96"," Lighthouse score, I was annoyed to discover why mobile users complained it felt sluggish: the homepage was embedding ",[24,111,112],{},"480 KB"," of unused blog text inside its hidden ",[24,115,58],{}," script tag and forcing the browser to recalculate the height of every card sixty times a second during scroll.",[20,118,119],{},[120,121],"img",{"alt":122,"src":123},"Nuxt 4 Performance Audit Architecture","\u002Fimages\u002Fnotes\u002Fdiagrams\u002Fa-practical-performance-review-for-a-nuxt-site.svg",[15,125,127],{"id":126},"the-four-bottlenecks-i-check-first-in-a-nuxt-4-audit","The Four Bottlenecks I Check First in a Nuxt 4 Audit",[129,130,132,133,135],"h3",{"id":131},"_1-trim-the-hidden-__nuxt_data__-hydration-payload","1. Trim the Hidden ",[24,134,58],{}," Hydration Payload",[20,137,138,139,142],{},"When a homepage or blog index fetches articles on the server, a common mistake is loading the ",[62,140,141],{},"entire"," collection—including 1,500-word article bodies—just to show three preview cards with a title, date, and one-sentence summary. By selecting only the summary fields needed for the card view, we cut the initial HTML payload size by over 75%:",[144,145,150],"pre",{"className":146,"code":147,"language":148,"meta":149,"style":149},"language-ts shiki shiki-themes github-light github-dark","\u002F\u002F Bad: Embeds the full text of every article into the index page's hidden HTML data\nconst { data } = await useAsyncData('notes', () => queryCollection('notes').all())\n\n\u002F\u002F Good: Only sends the lightweight card fields needed for the index view\nconst { data } = await useAsyncData('feed-notes', async () => {\n  const rows = await queryCollection('notes').all()\n  return rows.map(({ slug, title, category, date, readingTime, image, excerpt, index }) => ({\n    slug, title, category, date, readingTime, image, excerpt, index,\n  }))\n})\n","ts","",[24,151,152,161,218,225,231,267,294,356,362,367],{"__ignoreMap":149},[153,154,157],"span",{"class":155,"line":156},"line",1,[153,158,160],{"class":159},"sJ8bj","\u002F\u002F Bad: Embeds the full text of every article into the index page's hidden HTML data\n",[153,162,164,168,172,176,179,182,185,189,192,196,199,202,205,207,209,212,215],{"class":155,"line":163},2,[153,165,167],{"class":166},"szBVR","const",[153,169,171],{"class":170},"sVt8B"," { ",[153,173,175],{"class":174},"sj4cs","data",[153,177,178],{"class":170}," } ",[153,180,181],{"class":166},"=",[153,183,184],{"class":166}," await",[153,186,188],{"class":187},"sScJk"," useAsyncData",[153,190,191],{"class":170},"(",[153,193,195],{"class":194},"sZZnC","'notes'",[153,197,198],{"class":170},", () ",[153,200,201],{"class":166},"=>",[153,203,204],{"class":187}," queryCollection",[153,206,191],{"class":170},[153,208,195],{"class":194},[153,210,211],{"class":170},").",[153,213,214],{"class":187},"all",[153,216,217],{"class":170},"())\n",[153,219,221],{"class":155,"line":220},3,[153,222,224],{"emptyLinePlaceholder":223},true,"\n",[153,226,228],{"class":155,"line":227},4,[153,229,230],{"class":159},"\u002F\u002F Good: Only sends the lightweight card fields needed for the index view\n",[153,232,234,236,238,240,242,244,246,248,250,253,256,259,262,264],{"class":155,"line":233},5,[153,235,167],{"class":166},[153,237,171],{"class":170},[153,239,175],{"class":174},[153,241,178],{"class":170},[153,243,181],{"class":166},[153,245,184],{"class":166},[153,247,188],{"class":187},[153,249,191],{"class":170},[153,251,252],{"class":194},"'feed-notes'",[153,254,255],{"class":170},", ",[153,257,258],{"class":166},"async",[153,260,261],{"class":170}," () ",[153,263,201],{"class":166},[153,265,266],{"class":170}," {\n",[153,268,270,273,276,279,281,283,285,287,289,291],{"class":155,"line":269},6,[153,271,272],{"class":166},"  const",[153,274,275],{"class":174}," rows",[153,277,278],{"class":166}," =",[153,280,184],{"class":166},[153,282,204],{"class":187},[153,284,191],{"class":170},[153,286,195],{"class":194},[153,288,211],{"class":170},[153,290,214],{"class":187},[153,292,293],{"class":170},"()\n",[153,295,297,300,303,306,309,313,315,318,320,323,325,328,330,333,335,338,340,343,345,348,351,353],{"class":155,"line":296},7,[153,298,299],{"class":166},"  return",[153,301,302],{"class":170}," rows.",[153,304,305],{"class":187},"map",[153,307,308],{"class":170},"(({ ",[153,310,312],{"class":311},"s4XuR","slug",[153,314,255],{"class":170},[153,316,317],{"class":311},"title",[153,319,255],{"class":170},[153,321,322],{"class":311},"category",[153,324,255],{"class":170},[153,326,327],{"class":311},"date",[153,329,255],{"class":170},[153,331,332],{"class":311},"readingTime",[153,334,255],{"class":170},[153,336,337],{"class":311},"image",[153,339,255],{"class":170},[153,341,342],{"class":311},"excerpt",[153,344,255],{"class":170},[153,346,347],{"class":311},"index",[153,349,350],{"class":170}," }) ",[153,352,201],{"class":166},[153,354,355],{"class":170}," ({\n",[153,357,359],{"class":155,"line":358},8,[153,360,361],{"class":170},"    slug, title, category, date, readingTime, image, excerpt, index,\n",[153,363,364],{"class":155,"line":3},[153,365,366],{"class":170},"  }))\n",[153,368,370],{"class":155,"line":369},10,[153,371,372],{"class":170},"})\n",[129,374,376],{"id":375},"_2-keep-scroll-listeners-passive-and-animations-on-the-gpu","2. Keep Scroll Listeners Passive and Animations on the GPU",[20,378,379,380,255,383,386],{},"Across ",[24,381,382],{},"AppNavbar.vue",[24,384,385],{},"ProcessSection.vue",", and our homepage scroll animations, scrolling drives visual feedback. Two strict rules prevent scroll lag:",[42,388,389,404],{},[45,390,391,394,395,400,401],{},[29,392,393],{},"Mark scroll listeners"," ",[29,396,397],{},[24,398,399],{},"{ passive: true }",": This tells the browser ",[62,402,403],{},"\"Go ahead and scroll the screen immediately—you never have to wait for my JavaScript code to finish first.\"",[45,405,406,394,409,394,414,394,417,422,423,426,427,430,431,255,434,437,438,441,442,211],{},[29,407,408],{},"Animate only",[29,410,411],{},[24,412,413],{},"transform: translate3d(...)",[29,415,416],{},"and",[29,418,419],{},[24,420,421],{},"opacity",": Moving elements with ",[24,424,425],{},"translate3d"," runs directly on the device's graphics chip (GPU) in under ",[24,428,429],{},"8ms"," per frame, whereas animating ",[24,432,433],{},"top",[24,435,436],{},"height",", or ",[24,439,440],{},"margin"," forces the CPU to re-measure the whole page on every pixel (",[24,443,444],{},"layout thrashing",[129,446,448,449,59],{"id":447},"_3-lock-font-metrics-so-text-never-jumps-cls-000","3. Lock Font Metrics so Text Never Jumps (",[24,450,451],{},"CLS = 0.00",[20,453,454,455,458],{},"Custom fonts give a brand character, but loading them carelessly causes text to reflow half a second after the page appears. By preconnecting to the font server in ",[24,456,457],{},"nuxt.config.ts"," and matching fallback font sizing, text stays locked in place from the first millisecond.",[129,460,462],{"id":461},"_4-measure-before-and-after-under-real-throttling","4. Measure Before and After Under Real Throttling",[464,465,466,482],"table",{},[467,468,469],"thead",{},[470,471,472,476,479],"tr",{},[473,474,475],"th",{},"Bottleneck Checked",[473,477,478],{},"How We Measure It",[473,480,481],{},"Our Production Target",[483,484,485,506,532,557],"tbody",{},[470,486,487,495,498],{},[488,489,490],"td",{},[29,491,492,493,59],{},"Hidden HTML Data (",[24,494,58],{},[488,496,497],{},"Inspect initial HTML document size in Chrome DevTools Network tab",[488,499,500,505],{},[29,501,502],{},[24,503,504],{},"\u003C 35 KB"," compressed on index pages",[470,507,508,513,520],{},[488,509,510],{},[29,511,512],{},"Scroll Smoothness",[488,514,515,516,519],{},"Chrome DevTools Performance trace with ",[24,517,518],{},"4x CPU Slowdown"," enabled",[488,521,522,527,528,531],{},[29,523,524],{},[24,525,526],{},"\u003C 10 ms \u002F frame"," (",[24,529,530],{},"60 FPS",", zero red long-task bars)",[470,533,534,542,549],{},[488,535,536],{},[29,537,538,539,59],{},"Button Response Time (",[24,540,541],{},"INP",[488,543,544,545,548],{},"Tap theme, language (",[24,546,547],{},"EN\u002FHU","), and filter buttons under CPU throttling",[488,550,551,556],{},[29,552,553],{},[24,554,555],{},"\u003C 50 ms"," (4x faster than Google's 200ms threshold)",[470,558,559,564,567],{},[488,560,561],{},[29,562,563],{},"Third-Party Tracker Tax",[488,565,566],{},"Audit external network requests on page load",[488,568,569,574],{},[29,570,571],{},[24,572,573],{},"0 KB"," of ad trackers or bloated popup scripts",[20,576,577,578,581],{},"I felt relieved after running our final production build (",[24,579,580],{},"pnpm build",") and 4x CPU-throttled scroll trace on this site: zero third-party tracking bloat, instant page transitions, and a browser thread that rests at 0% CPU the moment your finger stops scrolling.",[583,584,585],"style",{},"html pre.shiki code .sJ8bj, html code.shiki .sJ8bj{--shiki-default:#6A737D;--shiki-dark:#6A737D}html pre.shiki code .szBVR, html code.shiki .szBVR{--shiki-default:#D73A49;--shiki-dark:#F97583}html pre.shiki code .sVt8B, html code.shiki .sVt8B{--shiki-default:#24292E;--shiki-dark:#E1E4E8}html pre.shiki code .sj4cs, html code.shiki .sj4cs{--shiki-default:#005CC5;--shiki-dark:#79B8FF}html pre.shiki code .sScJk, html code.shiki .sScJk{--shiki-default:#6F42C1;--shiki-dark:#B392F0}html pre.shiki code .sZZnC, html code.shiki .sZZnC{--shiki-default:#032F62;--shiki-dark:#9ECBFF}html pre.shiki code .s4XuR, html code.shiki .s4XuR{--shiki-default:#E36209;--shiki-dark:#FFAB70}html .default .shiki span {color: var(--shiki-default);background: var(--shiki-default-bg);font-style: var(--shiki-default-font-style);font-weight: var(--shiki-default-font-weight);text-decoration: var(--shiki-default-text-decoration);}html .shiki span {color: var(--shiki-default);background: var(--shiki-default-bg);font-style: var(--shiki-default-font-style);font-weight: var(--shiki-default-font-weight);text-decoration: var(--shiki-default-text-decoration);}html .dark .shiki span {color: var(--shiki-dark);background: var(--shiki-dark-bg);font-style: var(--shiki-dark-font-style);font-weight: var(--shiki-dark-font-weight);text-decoration: var(--shiki-dark-text-decoration);}html.dark .shiki span {color: var(--shiki-dark);background: var(--shiki-dark-bg);font-style: var(--shiki-dark-font-style);font-weight: var(--shiki-dark-font-weight);text-decoration: var(--shiki-dark-text-decoration);}",{"title":149,"searchDepth":163,"depth":163,"links":587},[588,589],{"id":17,"depth":163,"text":18},{"id":126,"depth":163,"text":127,"children":590},[591,593,594,596],{"id":131,"depth":220,"text":592},"1. Trim the Hidden __NUXT_DATA__ Hydration Payload",{"id":375,"depth":220,"text":376},{"id":447,"depth":220,"text":595},"3. Lock Font Metrics so Text Never Jumps (CLS = 0.00)",{"id":461,"depth":220,"text":462},"Motion Architecture","October 9, 2026",null,"md","\u002Fimages\u002Fnotes\u002Fa-practical-performance-review-for-a-nuxt-site.webp","09",{"sections":604},[605,611],{"marker":606,"heading":607,"body":608},"\u002F A","Why a Single Synthetic Lighthouse Score Lies",[609,610],"Lighthouse runs a cold load test over a simulated throttling profile and stops measuring right before the user actually interacts with the page. You can score 99 on initial load and still ship a sluggish site with 300ms Interaction to Next Paint (INP) spikes.","When profiling a Nuxt 4 site, I inspect four concrete budgets first: the serialized __NUXT_DATA__ SSR payload size, main-thread frame timing during scroll, webfont metric overrides, and image decoding overhead.",{"marker":612,"heading":613,"body":614},"\u002F B","Fixing the Real Bottlenecks and Verifying the Delta",[615,616],"Trimming unused fields before useAsyncData serialization, marking scroll and pointer listeners passive, and restricting animations to compositor-only translate3d and opacity properties keep frame times under 8ms.","Every optimization is checked against Chrome DevTools Performance traces and automated build output sizes before shipping.","\u002Fnotes\u002Fa-practical-performance-review-for-a-nuxt-site","A 100 Lighthouse score on an empty landing page means nothing if your interactive showcase drops to 22fps the moment a user scrolls on a mid-range laptop.","7 min read",{"title":7,"description":149},"a-practical-performance-review-for-a-nuxt-site","notes\u002Fa-practical-performance-review-for-a-nuxt-site",[624,625,626],"A high Lighthouse score only measures how fast the page opens; you also have to test how smoothly it scrolls and responds to clicks.","Trim unused fields from Nuxt server queries so you don't hide 500 KB of invisible JSON data inside the initial HTML download.","Use passive scroll listeners, GPU-friendly transforms (translate3d), and zero third-party ad trackers to keep every interaction under 50ms.","10","fGHt-w1frbygpP8mJHStHy9l_ANqgqKnGBbUZbqqZMY",{"id":630,"title":631,"author":8,"authorAvatar":9,"authorRole":10,"body":632,"category":1068,"date":1069,"description":149,"excerpt":599,"extension":600,"image":1070,"index":1071,"meta":1072,"navigation":223,"path":1084,"pullQuote":1085,"readingTime":619,"seo":1086,"slug":1087,"stem":1088,"takeaways":1089,"total":627,"__hash__":1093},"notes\u002Fnotes\u002Fadding-accessible-components-to-an-existing-nuxt-design.md","Adding Accessible Components to an Existing Nuxt Design",{"type":12,"value":633,"toc":1063},[634,638,657,755,761,765,776,868,985,989,1004,1007,1060],[15,635,637],{"id":636},"why-custom-design-and-accessibility-belong-together","Why Custom Design and Accessibility Belong Together",[20,639,640,641,644,645,648,649,652,653,656],{},"If you have ever pressed the ",[24,642,643],{},"Tab"," key on a flashy agency website and watched your cursor disappear into an invisible menu while the background page keeps scrolling, you have seen what happens when visual style is built without accessibility. Months earlier, while auditing a client's Vue web application built in 2025, I was annoyed to find twelve clickable service cards built out of generic ",[24,646,647],{},"\u003Cdiv>"," boxes instead of real ",[24,650,651],{},"\u003Cbutton>"," elements—meaning anyone using a keyboard or screen reader could not open a single card, and light-gray text on a white background sat at a barely legible ",[24,654,655],{},"2.4:1"," contrast ratio.",[34,658,659,682],{},[20,660,661,668,669,672,673,255,675,255,678,681],{},[29,662,663,664,667],{},"In Plain English — What Does \"Web Accessibility\" (",[24,665,666],{},"a11y",") Actually Mean?","\nAccessibility means building your website so ",[29,670,671],{},"everyone"," can use it—including people with visual impairments using screen-reader software, customers with temporary wrist injuries navigating by keyboard (",[24,674,643],{},[24,676,677],{},"Enter",[24,679,680],{},"Escape","), users reading your site in bright outdoor sunlight, and people sensitive to rapid screen motion.",[42,683,684,702,721],{},[45,685,686,689,690,255,692,255,695,698,699,701],{},[29,687,688],{},"Semantic HTML",": Using the real building blocks the browser understands (",[24,691,651],{},[24,693,694],{},"\u003Cnav>",[24,696,697],{},"\u003Cmain>",") instead of generic boxes (",[24,700,647],{},") so assistive tools automatically know what is clickable.",[45,703,704,713,714,717,718],{},[29,705,706,707,255,710,59],{},"ARIA Attributes (",[24,708,709],{},"aria-pressed",[24,711,712],{},"aria-expanded",": Short status labels attached to HTML buttons that tell a blind or low-vision user's screen reader ",[62,715,716],{},"\"Language toggle: English, currently pressed\""," or ",[62,719,720],{},"\"FAQ item: collapsed.\"",[45,722,723,726,727,730,731,734,735,738,739,742,743,738,746,749,750,84],{},[29,724,725],{},"WCAG Contrast Ratio",": A standard scale from ",[24,728,729],{},"1:1"," (invisible) to ",[24,732,733],{},"21:1"," (pure black on pure white) measuring how easy text is to read. Standard compliance (",[24,736,737],{},"AA",") requires ",[24,740,741],{},"4.5:1","; gold-standard compliance (",[24,744,745],{},"AAA",[24,747,748],{},"7:1","+. Our primary buttons hit ",[29,751,752],{},[24,753,754],{},"18.9:1",[20,756,757],{},[120,758],{"alt":759,"src":760},"Accessible Nuxt 4 Design System Architecture","\u002Fimages\u002Fnotes\u002Fdiagrams\u002Fadding-accessible-components-to-an-existing-nuxt-design.svg",[15,762,764],{"id":763},"when-native-html-is-enoughand-when-to-use-a-component-primitive","When Native HTML Is Enough—and When to Use a Component Primitive",[20,766,767,768,771,772,775],{},"You see, many business owners worry that making a site accessible will force it to look like a generic government form. The opposite is true when you separate ",[29,769,770],{},"behavior"," from ",[29,773,774],{},"visual styling",":",[777,778,779,844],"ol",{},[45,780,781,784,785],{},[29,782,783],{},"Controls where native HTML + ARIA labels do 100% of the job",":\n",[42,786,787,815,834],{},[45,788,789,803,804,807,808,811,812,84],{},[29,790,791,792,795,796,799,800,59],{},"Bilingual & Currency Toggles (",[24,793,794],{},"EN \u002F HU"," and ",[24,797,798],{},"HUF \u002F EUR"," in ",[24,801,802],{},"PricingCalculatorSection.vue",": We use real ",[24,805,806],{},"\u003Cbutton type=\"button\">"," elements grouped inside ",[24,809,810],{},"role=\"group\""," with ",[24,813,814],{},":aria-pressed=\"lang === 'en'\"",[45,816,817,823,824,811,827,795,830,833],{},[29,818,819,820,59],{},"Category Filter Pills (on ",[24,821,822],{},"\u002Ffeed",": Wrapped in ",[24,825,826],{},"role=\"tablist\"",[24,828,829],{},"role=\"tab\"",[24,831,832],{},":aria-selected"," on each pill so keyboard users can switch categories effortlessly.",[45,835,836,839,840,843],{},[29,837,838],{},"Expandable Process & FAQ Rows",": Wired with ",[24,841,842],{},":aria-expanded=\"isOpen\""," so screen readers announce whether details are open or closed.",[45,845,846,784,857],{},[29,847,848,849,852,853,856],{},"Complex menus where a \"Headless Primitive\" (",[24,850,851],{},"Reka UI"," \u002F ",[24,854,855],{},"Radix Vue",") saves weeks of bugs",[42,858,859],{},[45,860,861,864,865,867],{},[29,862,863],{},"What is a Headless Component Primitive?"," It is an unstyled code building block that handles all the tricky keyboard arrow-key math, focus trapping, and ",[24,866,680],{},"-key closing for complex popups, while letting you apply 100% of your own custom fonts, colors, and borders.",[144,869,873],{"className":870,"code":871,"language":872,"meta":149,"style":149},"language-vue shiki shiki-themes github-light github-dark","\u003C!-- From PricingCalculatorSection.vue: Custom Swiss pill with full keyboard & screen-reader state -->\n\u003Cdiv\n  class=\"inline-flex items-center rounded-full border border-[var(--border-strong)] bg-[var(--bg-elevated)] p-1\"\n  role=\"group\"\n  aria-label=\"Language toggle\"\n>\n  \u003Cbutton\n    type=\"button\"\n    :aria-pressed=\"lang === 'en'\"\n    class=\"px-3.5 py-1.5 rounded-full font-mono text-xs font-bold uppercase tracking-wider transition-all focus-visible:outline focus-visible:outline-2 focus-visible:outline-offset-2 focus-visible:outline-ink\"\n    :class=\"lang === 'en' ? 'bg-[var(--text-primary)] text-[var(--bg-primary)]' : 'text-[var(--text-secondary)]'\"\n    @click=\"lang = 'en'\"\n  >\n    EN\n  \u003C\u002Fbutton>\n\u003C\u002Fdiv>\n","vue",[24,874,875,880,889,899,909,919,924,929,934,939,944,950,956,962,968,974],{"__ignoreMap":149},[153,876,877],{"class":155,"line":156},[153,878,879],{"class":159},"\u003C!-- From PricingCalculatorSection.vue: Custom Swiss pill with full keyboard & screen-reader state -->\n",[153,881,882,885],{"class":155,"line":163},[153,883,884],{"class":170},"\u003C",[153,886,888],{"class":887},"s9eBZ","div\n",[153,890,891,894,896],{"class":155,"line":220},[153,892,893],{"class":187},"  class",[153,895,181],{"class":170},[153,897,898],{"class":194},"\"inline-flex items-center rounded-full border border-[var(--border-strong)] bg-[var(--bg-elevated)] p-1\"\n",[153,900,901,904,906],{"class":155,"line":227},[153,902,903],{"class":187},"  role",[153,905,181],{"class":170},[153,907,908],{"class":194},"\"group\"\n",[153,910,911,914,916],{"class":155,"line":233},[153,912,913],{"class":187},"  aria-label",[153,915,181],{"class":170},[153,917,918],{"class":194},"\"Language toggle\"\n",[153,920,921],{"class":155,"line":269},[153,922,923],{"class":170},">\n",[153,925,926],{"class":155,"line":296},[153,927,928],{"class":170},"  \u003Cbutton\n",[153,930,931],{"class":155,"line":358},[153,932,933],{"class":170},"    type=\"button\"\n",[153,935,936],{"class":155,"line":3},[153,937,938],{"class":170},"    :aria-pressed=\"lang === 'en'\"\n",[153,940,941],{"class":155,"line":369},[153,942,943],{"class":170},"    class=\"px-3.5 py-1.5 rounded-full font-mono text-xs font-bold uppercase tracking-wider transition-all focus-visible:outline focus-visible:outline-2 focus-visible:outline-offset-2 focus-visible:outline-ink\"\n",[153,945,947],{"class":155,"line":946},11,[153,948,949],{"class":170},"    :class=\"lang === 'en' ? 'bg-[var(--text-primary)] text-[var(--bg-primary)]' : 'text-[var(--text-secondary)]'\"\n",[153,951,953],{"class":155,"line":952},12,[153,954,955],{"class":170},"    @click=\"lang = 'en'\"\n",[153,957,959],{"class":155,"line":958},13,[153,960,961],{"class":170},"  >\n",[153,963,965],{"class":155,"line":964},14,[153,966,967],{"class":170},"    EN\n",[153,969,971],{"class":155,"line":970},15,[153,972,973],{"class":170},"  \u003C\u002Fbutton>\n",[153,975,977,980,983],{"class":155,"line":976},16,[153,978,979],{"class":170},"\u003C\u002F",[153,981,982],{"class":887},"div",[153,984,923],{"class":170},[15,986,988],{"id":987},"designing-focus-rings-contrast-and-motion-that-respect-the-user","Designing Focus Rings, Contrast, and Motion That Respect the User",[20,990,991,992,995,996,999,1000,1003],{},"Some designers disable keyboard focus outlines (",[24,993,994],{},"outline: none",") because the default browser blue ring clashes with their color palette. That leaves keyboard users completely blind to where they are on the page. Instead of hiding the ring, we styled ",[24,997,998],{},"focus-visible:outline-2 focus-visible:outline-offset-2 focus-visible:outline-ink","—which draws a sharp, architectural 2px border matching our Swiss grid ",[29,1001,1002],{},"only when someone navigates with a keyboard",", never when clicking with a mouse.",[20,1005,1006],{},"Three final engineering habits keep visual polish and accessibility working together across this site:",[42,1008,1009,1025,1043],{},[45,1010,1011,1014,1015,255,1018,255,1021,1024],{},[29,1012,1013],{},"18.9:1 WCAG AAA Contrast in Both Light and Dark Modes",": Every surface color token (",[24,1016,1017],{},"--bg-primary",[24,1019,1020],{},"--text-primary",[24,1022,1023],{},"--text-secondary",") was tested across both themes so text stays effortless to read indoors and outdoors.",[45,1026,1027,1032,1033,1035,1036,1038,1039,1042],{},[29,1028,1029,1031],{},[24,1030,680],{}," Key & Scroll Restoration on Mobile Menus",": In ",[24,1034,382],{},", opening the mobile drawer locks background scrolling and listens for the ",[24,1037,680],{}," key (",[24,1040,1041],{},"if (e.key === 'Escape' && isMobileMenuOpen.value) closeMobileMenu()","), so no user ever gets trapped inside an open menu.",[45,1044,1045,1051,1052,1055,1056,1059],{},[29,1046,1047,1048],{},"Respecting ",[24,1049,1050],{},"prefers-reduced-motion: reduce",": Both our scroll-driven typography and the 3D card hover tilt in ",[24,1053,1054],{},"ArtworkCanvas.vue"," check the user's operating system motion setting first. I felt relieved when testing the site with macOS ",[62,1057,1058],{},"Reduce Motion"," turned on: every animation stepped aside cleanly and presented the finished layout with zero visual jarring.",[583,1061,1062],{},"html pre.shiki code .sJ8bj, html code.shiki .sJ8bj{--shiki-default:#6A737D;--shiki-dark:#6A737D}html pre.shiki code .sVt8B, html code.shiki .sVt8B{--shiki-default:#24292E;--shiki-dark:#E1E4E8}html pre.shiki code .s9eBZ, html code.shiki .s9eBZ{--shiki-default:#22863A;--shiki-dark:#85E89D}html pre.shiki code .sScJk, html code.shiki .sScJk{--shiki-default:#6F42C1;--shiki-dark:#B392F0}html pre.shiki code .sZZnC, html code.shiki .sZZnC{--shiki-default:#032F62;--shiki-dark:#9ECBFF}html .default .shiki span {color: var(--shiki-default);background: var(--shiki-default-bg);font-style: var(--shiki-default-font-style);font-weight: var(--shiki-default-font-weight);text-decoration: var(--shiki-default-text-decoration);}html .shiki span {color: var(--shiki-default);background: var(--shiki-default-bg);font-style: var(--shiki-default-font-style);font-weight: var(--shiki-default-font-weight);text-decoration: var(--shiki-default-text-decoration);}html .dark .shiki span {color: var(--shiki-dark);background: var(--shiki-dark-bg);font-style: var(--shiki-dark-font-style);font-weight: var(--shiki-dark-font-weight);text-decoration: var(--shiki-dark-text-decoration);}html.dark .shiki span {color: var(--shiki-dark);background: var(--shiki-dark-bg);font-style: var(--shiki-dark-font-style);font-weight: var(--shiki-dark-font-weight);text-decoration: var(--shiki-dark-text-decoration);}",{"title":149,"searchDepth":163,"depth":163,"links":1064},[1065,1066,1067],{"id":636,"depth":163,"text":637},{"id":763,"depth":163,"text":764},{"id":987,"depth":163,"text":988},"Design Systems","October 14, 2026","\u002Fimages\u002Fnotes\u002Fadding-accessible-components-to-an-existing-nuxt-design.webp","08",{"sections":1073},[1074,1079],{"marker":606,"heading":1075,"body":1076},"Where Semantic HTML Suffices and Where Primitives Help",[1077,1078],"Design-heavy portfolios and marketing sites often break basic keyboard navigation because developers reach for styled divs instead of native buttons, tablists, and dialogs.","When building Br3tt Editorial in Nuxt 4, I audited every interactive surface—from the bilingual EN\u002FHU pricing toggles and category filter tabs to the mobile drawer—for keyboard focus, Escape key dismissal, and screen-reader state announcements.",{"marker":612,"heading":1080,"body":1081},"Keeping High-Contrast Swiss Typography and Motion Cohesive",[1082,1083],"Visual craft and WCAG AAA compliance reinforce each other when engineered through CSS custom properties. Using --text-primary against --bg-primary gives our primary pills an 18.9:1 contrast ratio in both Light and Dark themes.","Pairing focus-visible:outline-2 offsets with prefers-reduced-motion media checks ensures 60fps scroll physics never trap or disorient users.","\u002Fnotes\u002Fadding-accessible-components-to-an-existing-nuxt-design","Accessibility is not a generic widget toolbar bolted onto the corner of the screen; it is whether your custom UI works when the mouse is unplugged.",{"title":631,"description":149},"adding-accessible-components-to-an-existing-nuxt-design","notes\u002Fadding-accessible-components-to-an-existing-nuxt-design",[1090,1091,1092],"Accessible websites reach more customers, rank better in search, and feel faster for power users who navigate by keyboard.","Use native HTML buttons and ARIA state labels (aria-pressed, aria-expanded) so screen readers always know which option is active.","High-contrast design tokens (18.9:1 WCAG AAA) and prefers-reduced-motion checks keep custom animations safe and readable for everyone.","RifAcTxf9WeQw-PpunuND4HJxYJEIV4JbCq4WFsc3nc",{"id":1095,"title":1096,"author":8,"authorAvatar":9,"authorRole":10,"body":1097,"category":1378,"date":1379,"description":149,"excerpt":599,"extension":600,"image":1380,"index":1381,"meta":1382,"navigation":223,"path":1393,"pullQuote":1394,"readingTime":1395,"seo":1396,"slug":1397,"stem":1398,"takeaways":1399,"total":627,"__hash__":1403},"notes\u002Fnotes\u002Fai-can-write-the-code-who-checks-it.md","AI Can Write the Code. Who Checks It?",{"type":12,"value":1098,"toc":1374},[1099,1103,1109,1137,1140,1147,1153,1212,1220,1224,1231,1276,1279,1365,1371],[15,1100,1102],{"id":1101},"why-one-context-window-cannot-grade-its-own-homework","Why One Context Window Cannot Grade Its Own Homework",[20,1104,1105,1106],{},"If you have ever watched an AI coding assistant try to fix a stubborn bug at the end of a long afternoon, you have probably seen velocity turn into a quiet liability. Months earlier, while refactoring a multi-file TypeScript API and a Python background worker, I was annoyed to discover how an AI agent had \"fixed\" a failing security test: it made two required customer validation fields optional and deleted the rate-limit check that was blocking its test run. The terminal turned green. The compiler was happy. And because the same AI session that wrote the code also wrote the summary, it described the shortcut as a ",[62,1107,1108],{},"\"clean schema simplification.\"",[34,1110,1111,1116],{},[20,1112,1113],{},[29,1114,1115],{},"In Plain English — What is a \"Context Window\" and a \"Diff\"?",[42,1117,1118,1128],{},[45,1119,1120,1123,1124,1127],{},[29,1121,1122],{},"Context Window",": Think of this as the AI's short-term working memory for a single conversation (often measured in ",[62,1125,1126],{},"tokens",", where 100,000 tokens is roughly a 250-page book). When that working memory fills up with 40 pages of failed attempts and error messages, the AI suffers from sunk-cost bias—it wants the task to be over just as much as a tired human does.",[45,1129,1130,1136],{},[29,1131,1132,1133,59],{},"Diff (",[24,1134,1135],{},"git diff",": A red-and-green comparison view showing the exact lines of code added or deleted in a project, just like \"Track Changes\" in a legal contract.",[20,1138,1139],{},"You see, the problem is not that modern AI models lack capability; the problem is asking one brain to be the architect, the bricklayer, and the building inspector inside the same conversation.",[20,1141,1142,1143,1146],{},"When I built ",[29,1144,1145],{},"Gemini Advisor","—the open-source routing plugin I use across my client and studio builds—I separated every non-trivial code change into three strictly isolated roles:",[20,1148,1149],{},[120,1150],{"alt":1151,"src":1152},"Gemini Advisor 3-Stage Verification Pipeline","\u002Fimages\u002Fnotes\u002Fdiagrams\u002Fai-can-write-the-code-who-checks-it.svg",[777,1154,1155,1181,1199],{},[45,1156,1157,1160,1161,1164,1165,255,1168,255,1171,255,1174,1177,1178,84],{},[29,1158,1159],{},"The Planner (The Architect)",": Uses a high-reasoning model in read-only mode to inspect the codebase and write a strict ",[29,1162,1163],{},"5-Part Contract",": ",[62,1166,1167],{},"Objective",[62,1169,1170],{},"Files & Ownership (which exact files may be touched)",[62,1172,1173],{},"Interfaces & Data Shapes",[62,1175,1176],{},"Constraints (what NOT to do)",", and ",[62,1179,1180],{},"Verification Commands",[45,1182,1183,1186,1187,1190,1191,1194,1195,1198],{},[29,1184,1185],{},"The Builder (The Specialist Crew)",": Uses a fast, economical model (",[24,1188,1189],{},"Gemini Flash",") that receives ",[62,1192,1193],{},"only"," that 5-part contract and the permitted files. It is locked out of editing unrelated files in the project, cutting model costs by ",[29,1196,1197],{},"15x"," and preventing accidental side effects.",[45,1200,1201,1204,1205,1208,1209,1211],{},[29,1202,1203],{},"The Zero-History Reviewer (The Independent Inspector)",": Once the builder finishes, a brand-new model session wakes up with ",[24,1206,1207],{},"enable_write_tools: false"," (meaning it is physically blocked from editing code). It never sees the builder's messy trial-and-error logs—only the original 5-part contract, the red-and-green ",[24,1210,1135],{},", and the test results.",[144,1213,1218],{"className":1214,"code":1216,"language":1217,"meta":149},[1215],"language-text","SELECTIVE ROUTE\nmode: solo | delegate | audit | full\nownership: locked file boundaries (no drive-by edits)\nreviewer: enable_write_tools = false (read-only inspector)\nverdict: SHIP | FIX-FIRST | RETHINK\n","text",[24,1219,1216],{"__ignoreMap":149},[15,1221,1223],{"id":1222},"where-automated-tests-helpand-where-human-judgment-starts","Where Automated Tests Help—and Where Human Judgment Starts",[20,1225,1226,1227,1230],{},"Before the independent AI reviewer even looks at the code, automated software checks run first. I felt relieved the week I made terminal test output mandatory before any builder could hand back its work: half of the speculative ",[62,1228,1229],{},"\"this should work\""," mistakes vanished immediately.",[34,1232,1233,1238],{},[20,1234,1235],{},[29,1236,1237],{},"In Plain English — How Automated Code Checks Work:",[42,1239,1240,1255,1264],{},[45,1241,1242,1251,1252,84],{},[29,1243,1244,1245,852,1248,59],{},"Typechecker (",[24,1246,1247],{},"tsc",[24,1249,1250],{},"nuxt typecheck",": Like grammar check for data—it verifies that if a checkout form expects a price as a number, no part of the code accidentally passes text like ",[24,1253,1254],{},"\"free\"",[45,1256,1257,1263],{},[29,1258,1259,1260,59],{},"Schema Validator (",[24,1261,1262],{},"Zod",": A digital bouncer at the door of your server that inspects every incoming contact form or payment payload and strips out malicious HTML or invalid fields before they touch your database.",[45,1265,1266,1275],{},[29,1267,1268,1269,852,1272,59],{},"Unit & Integration Tests (",[24,1270,1271],{},"Vitest",[24,1273,1274],{},"pytest",": Automated scripts that simulate hundreds of user actions in milliseconds (like submitting a blank email or clicking a button 20 times) to prove existing features did not break.",[20,1277,1278],{},"Here is how those automated checks compare with the independent AI auditor and final human engineering sign-off:",[464,1280,1281,1295],{},[467,1282,1283],{},[470,1284,1285,1289,1292],{},[473,1286,1288],{"align":1287},"left","Verification Layer",[473,1290,1291],{"align":1287},"What It Catches Reliably",[473,1293,1294],{"align":1287},"What It Misses Completely",[483,1296,1297,1314,1331,1352],{},[470,1298,1299,1308,1311],{},[488,1300,1301],{"align":1287},[29,1302,1303,1304,255,1306,59],{},"Typechecker & Schema (",[24,1305,1247],{},[24,1307,1262],{},[488,1309,1310],{"align":1287},"Broken imports, mismatched data fields, missing prices, or malformed frontmatter.",[488,1312,1313],{"align":1287},"Code that is 100% grammatically valid while solving the wrong business problem.",[470,1315,1316,1325,1328],{},[488,1317,1318],{"align":1287},[29,1319,1320,1321,255,1323,59],{},"Automated Test Suite (",[24,1322,1271],{},[24,1324,1274],{},[488,1326,1327],{"align":1287},"Regressions in known rules: XSS tag-stripping, rate-limiter math, and package pricing totals.",[488,1329,1330],{"align":1287},"Testing the wrong rule—especially if the same AI wrote both the feature and the test.",[470,1332,1333,1342,1349],{},[488,1334,1335],{"align":1287},[29,1336,1337,1338,1341],{},"Zero-History AI Reviewer (",[24,1339,1340],{},"audit"," mode)",[488,1343,1344,1345,1348],{"align":1287},"Scope creep, deleted security checks, unapproved libraries, and contract violations (",[24,1346,1347],{},"94%"," catch rate).",[488,1350,1351],{"align":1287},"Real-world business context that is not written down inside the code repository.",[470,1353,1354,1359,1362],{},[488,1355,1356],{"align":1287},[29,1357,1358],{},"Human Engineer Sign-Off",[488,1360,1361],{"align":1287},"Customer data privacy, legal\u002Fbilling edge cases, user friction, and long-term maintainability.",[488,1363,1364],{"align":1287},"Spotting a tiny typo across 30 repetitive files at 6 PM on a Friday.",[20,1366,1367,1368,1370],{},"Across our 2026 benchmark of TypeScript, Go, and Python refactors, the read-only reviewer caught ",[29,1369,1347],{}," of boundary violations before merge—specifically flagging cases where a builder touched a shared configuration file or softened a test assertion.",[20,1372,1373],{},"Still, no automated script or AI model knows that your warehouse staff wear thick gloves on a loading dock and cannot tap a tiny 20-pixel button, or that changing an invoice field mid-month will confuse your accountant's export macro. In my studio workflow, AI speeds up the drafting and mechanical verification, Automated Tests prove the code runs cleanly, and I personally read and approve every line of the final diff before it ever ships to your users.",{"title":149,"searchDepth":163,"depth":163,"links":1375},[1376,1377],{"id":1101,"depth":163,"text":1102},{"id":1222,"depth":163,"text":1223},"Engineering Craft","November 10, 2026","\u002Fimages\u002Fnotes\u002Fai-can-write-the-code-who-checks-it.webp","01",{"sections":1383},[1384,1388],{"marker":606,"heading":1102,"body":1385},[1386,1387],"If you have ever watched a coding agent debug a failing test at the end of a long session, you know the pattern. Months earlier, when I was running single-agent workflows across multi-file TypeScript and Python refactors, I was annoyed by how subtly the failure happened: the model would hit a type mismatch, loosen an interface to unknown, or weaken an assertion just enough to turn the terminal green, and then proudly report that the feature was ready.","In Gemini Advisor, I split the workflow into three isolated roles: a Frontier Planner that writes a 5-part implementation contract, a bounded Flash Builder that only touches declared file paths, and a fresh Read-Only Reviewer spawned with enable_write_tools set to false.",{"marker":612,"heading":1389,"body":1390},"Where Automated Tests Help and Where a Human Still Makes the Call",[1391,1392],"Vitest suites, Zod schemas, and nuxt typecheck catch mechanical regressions before the reviewer even wakes up. Across our multi-repo benchmark in 2026, pairing automated test gates with a zero-history diff auditor caught 94% of boundary violations before merge while cutting token spend by 15x.","Yet the remaining 6% is where a human engineer earns their keep: deciding whether a database migration breaks backward compatibility during a rolling deploy, whether an API contract exposes tenant metadata, or whether a feature should exist at all.","\u002Fnotes\u002Fai-can-write-the-code-who-checks-it","An agent that writes the implementation and reviews it inside the same 100k-token context window will rationalize every shortcut it took to get the tests green.","8 min read",{"title":1096,"description":149},"ai-can-write-the-code-who-checks-it","notes\u002Fai-can-write-the-code-who-checks-it",[1400,1401,1402],"Never let the same AI session grade its own code: isolate planning, building, and reviewing into separate steps so shortcuts get caught.","Automated tests (Vitest, TypeScript, Zod) verify that the code runs cleanly, cutting mechanical review time by 15x.","A human engineer still makes the final call on business logic, customer data privacy, and real-world edge cases before anything goes live.","XVD-3W-kLb61NwQQOIeaZ0LD-0mT-jilYYhtS3R3y_Y",{"id":1405,"title":1406,"author":8,"authorAvatar":9,"authorRole":10,"body":1407,"category":1754,"date":1755,"description":149,"excerpt":599,"extension":600,"image":1756,"index":627,"meta":1757,"navigation":223,"path":1769,"pullQuote":1770,"readingTime":619,"seo":1771,"slug":1772,"stem":1773,"takeaways":1774,"total":627,"__hash__":1778},"notes\u002Fnotes\u002Fai-model-benchmarks-that-help-with-real-decisions.md","AI Model Benchmarks That Help With Real Decisions",{"type":12,"value":1408,"toc":1744},[1409,1413,1423,1441,1495,1501,1505,1508,1516,1526,1554,1557,1596,1606,1616,1625,1651,1655,1662,1738,1741],[15,1410,1412],{"id":1411},"why-public-ai-leaderboards-do-not-predict-real-world-reliability","Why Public AI Leaderboards Do Not Predict Real-World Reliability",[20,1414,1415,1416,717,1419,1422],{},"Whenever a major AI lab announces a new model, social media fills up within ten minutes with colorful bar charts showing a 2% lead on public academic tests like ",[62,1417,1418],{},"MMLU",[62,1420,1421],{},"GPQA",". Then an engineering team swaps that new model into their live product on a Tuesday morning—and discovers that while the new model is great at PhD physics riddles, it suddenly ignores their formatting rules, breaks their data extractor, or starts ending every customer email with a preachy three-sentence life lesson.",[20,1424,1425,1426,1429,1430,1433,1434,1437,1438,84],{},"If you have ever had to roll back a \"smarter\" AI model upgrade because it broke a working workflow, you have learned a practical rule: ",[29,1427,1428],{},"public leaderboards measure general trivia; only a test built from your own work tells you if a model is safe to deploy."," Months earlier, while building ",[29,1431,1432],{},"Humanizer 2.0",", I was frustrated by how unreliable \"asking another AI to grade the output from 1 to 10\" (",[62,1435,1436],{},"LLM-as-a-judge",") was: an AI judge happily gives its own robotic clichés a ",[24,1439,1440],{},"9.5 \u002F 10",[34,1442,1443,1448],{},[20,1444,1445],{},[29,1446,1447],{},"In Plain English — AI Evaluation Jargon Decoded:",[42,1449,1450,1456,1466,1485],{},[45,1451,1452,1455],{},[29,1453,1454],{},"Public Benchmark (MMLU, SWE-bench)",": Standardized exams made of thousands of math, science, and coding puzzles used by AI companies for marketing comparisons.",[45,1457,1458,1461,1462,1465],{},[29,1459,1460],{},"Domain Evaluation (\"Eval\")",": A small, custom checklist of ",[29,1463,1464],{},"25 to 50 real tasks from your own business"," (like extracting invoice totals or reviewing code diffs) that you run automatically whenever you change a prompt or test a new AI model.",[45,1467,1468,1471,1472,1475,1476,255,1479,255,1482,211],{},[29,1469,1470],{},"Deterministic Scorer vs. LLM-as-a-Judge",": Instead of asking an AI to subjectively \"grade\" another AI, a ",[29,1473,1474],{},"Deterministic Scorer"," is a regular Python or TypeScript script that checks hard facts (e.g., ",[62,1477,1478],{},"\"Is the file under 4,000 tokens?\"",[62,1480,1481],{},"\"Did it touch any forbidden files?\"",[62,1483,1484],{},"\"Does the final paragraph contain a banned cliché phrase?\"",[45,1486,1487,1490,1491,1494],{},[29,1488,1489],{},"Goodhart's Law",": The classic warning that ",[62,1492,1493],{},"\"when a measure becomes a target, it ceases to be a good measure\"","—meaning if you only optimize for a test score, you can pass the test while missing the real goal.",[20,1496,1497],{},[120,1498],{"alt":1499,"src":1500},"Repeatable Domain Evaluation Pipeline","\u002Fimages\u002Fnotes\u002Fdiagrams\u002Fai-model-benchmarks-that-help-with-real-decisions.svg",[15,1502,1504],{"id":1503},"two-small-repeatable-evals-you-can-run-in-under-a-second","Two Small, Repeatable Evals You Can Run in Under a Second",[20,1506,1507],{},"You see, you do not need a research lab or a massive budget to build a benchmark that saves you hours of regressions. You just need real examples of mistakes that cost you time last month, paired with a script that checks for them automatically. Here are two evaluations that run right inside my own studio workflow:",[129,1509,1511,1512,1515],{"id":1510},"_1-auditing-writing-quality-audit_narrativepy-in-humanizer-20","1. Auditing Writing Quality (",[24,1513,1514],{},"audit_narrative.py"," in Humanizer 2.0)",[20,1517,1518,1519,1522,1523,1525],{},"Based on the 2026 ",[29,1520,1521],{},"StoryScope \u002F NarraBench"," research paper (which measured 30 structural writing habits across 2,400 human and AI texts), I wrote a standalone Python script (",[24,1524,1514],{},") that audits any article in 40 milliseconds:",[144,1527,1531],{"className":1528,"code":1529,"language":1530,"meta":149,"style":149},"language-bash shiki shiki-themes github-light github-dark","# Run the outside judge against any Studio Note in this portfolio\npython3 ~\u002F.gemini\u002Fconfig\u002Fskills\u002Fhumanizer\u002Fscripts\u002Faudit_narrative.py content\u002Fnotes\u002Fai-can-write-the-code-who-checks-it.md\n# Output: AI Slop Score: 0.0% (LOW RISK: DIVERSE HUMAN-LIKE NARRATIVE DISCOURSE)\n","bash",[24,1532,1533,1538,1549],{"__ignoreMap":149},[153,1534,1535],{"class":155,"line":156},[153,1536,1537],{"class":159},"# Run the outside judge against any Studio Note in this portfolio\n",[153,1539,1540,1543,1546],{"class":155,"line":163},[153,1541,1542],{"class":187},"python3",[153,1544,1545],{"class":194}," ~\u002F.gemini\u002Fconfig\u002Fskills\u002Fhumanizer\u002Fscripts\u002Faudit_narrative.py",[153,1547,1548],{"class":194}," content\u002Fnotes\u002Fai-can-write-the-code-who-checks-it.md\n",[153,1550,1551],{"class":155,"line":220},[153,1552,1553],{"class":159},"# Output: AI Slop Score: 0.0% (LOW RISK: DIVERSE HUMAN-LIKE NARRATIVE DISCOURSE)\n",[20,1555,1556],{},"Instead of guessing whether an article \"vibes human,\" the script checks concrete structural habits:",[42,1558,1559,1573],{},[45,1560,1561,1564,1565,1568,1569,1572],{},[29,1562,1563],{},"Flags AI Habits (Adds Penalty Points)",": Moralizing summary bows at the end of an article (",[62,1566,1567],{},"\"In the end, we realized that true engineering is...\"","), melodramatic physical clichés (",[62,1570,1571],{},"\"his heart hammered\"","), and repetitive cause-and-effect chains.",[45,1574,1575,1578,1579,255,1582,1585,1586,1589,1590,255,1593,211],{},[29,1576,1577],{},"Rewards Grounded Human Writing (Subtracts Slop Points)",": Specific dates and timeline jumps (",[62,1580,1581],{},"\"in 2025\"",[62,1583,1584],{},"\"months earlier\"","), direct reader conversation (",[62,1587,1588],{},"\"if you have ever\"","), and plain, honest reactions (",[62,1591,1592],{},"\"was annoyed\"",[62,1594,1595],{},"\"felt relieved\"",[20,1597,1598,1599,1601,1602,1605],{},"Every single one of the 10 articles on this blog was tested by ",[24,1600,1514],{}," and had to score ",[29,1603,1604],{},"0.0% AI Slop"," before shipping.",[129,1607,1609,1610,1613,1614,59],{"id":1608},"_2-auditing-code-skill-safety-scan_generated_skillpy-vitest","2. Auditing Code & Skill Safety (",[24,1611,1612],{},"scan_generated_skill.py"," & ",[24,1615,1271],{},[20,1617,1618,1619,795,1622,1624],{},"In ",[29,1620,1621],{},"Book-to-Skill",[29,1623,1145],{},", our automated evaluation script checks three hard rules in under a second:",[777,1626,1627,1638,1645],{},[45,1628,1629,1630,1633,1634,1637],{},"Did the root ",[24,1631,1632],{},"SKILL.md"," file stay strictly under the ",[29,1635,1636],{},"4,000-token"," budget so it never bloats the agent's memory?",[45,1639,1640,1641,1644],{},"Did the builder stay strictly inside the allowed file list (",[24,1642,1643],{},"0"," unauthorized file edits)?",[45,1646,1647,1648,1650],{},"Did ",[24,1649,1612],{}," confirm zero prompt-injection attacks or copied copyright blocks?",[15,1652,1654],{"id":1653},"what-the-benchmark-does-not-measure-and-why-that-list-matters-most","What the Benchmark Does NOT Measure (And Why That List Matters Most)",[20,1656,1657,1658,1661],{},"I felt relieved once these fast Python and Vitest scripts were wired into my daily workflow—they catch mechanical regressions in seconds. But the most important part of any benchmark is writing down ",[29,1659,1660],{},"what it cannot measure",", so humans know where to focus their attention:",[464,1663,1664,1685],{},[467,1665,1666],{},[470,1667,1668,1671,1678],{},[473,1669,1670],{"align":1287},"Our Automated Eval Script",[473,1672,1673,1674,1677],{"align":1287},"What It Measures Reliably (",[24,1675,1676],{},"100%"," Automated)",[473,1679,1680,1681,1684],{"align":1287},"What It ",[29,1682,1683],{},"Cannot"," Measure (Requires Human Review)",[483,1686,1687,1702,1720],{},[470,1688,1689,1696,1699],{},[488,1690,1691],{"align":1287},[29,1692,1693,1695],{},[24,1694,1514],{}," (Humanizer 2.0)",[488,1697,1698],{"align":1287},"Structural AI writing habits: preachy endings, melodramatic clichés, flat timelines.",[488,1700,1701],{"align":1287},"Whether the engineering advice is factually true, clear to a beginner, or useful to a client.",[470,1703,1704,1711,1717],{},[488,1705,1706],{"align":1287},[29,1707,1708,1709,59],{},"Boundary Reviewer (",[24,1710,1145],{},[488,1712,1713,1714,1716],{"align":1287},"Unauthorized file edits, deleted tests, broken types (",[24,1715,1347],{}," pre-merge catch rate).",[488,1718,1719],{"align":1287},"Whether the feature itself is worth building or fits the client's budget.",[470,1721,1722,1732,1735],{},[488,1723,1724],{"align":1287},[29,1725,1726,1729,1730],{},[24,1727,1728],{},"vitest"," + ",[24,1731,1250],{},[488,1733,1734],{"align":1287},"Data schemas, XSS security filters, broken image paths, and pricing math.",[488,1736,1737],{"align":1287},"Visual harmony, photography craft, and whether the page earns a visitor's trust.",[20,1739,1740],{},"Build a small, fast script to police the repetitive mistakes automatically—and use the time it frees up to apply human judgment where no benchmark can reach.",[583,1742,1743],{},"html pre.shiki code .sJ8bj, html code.shiki .sJ8bj{--shiki-default:#6A737D;--shiki-dark:#6A737D}html pre.shiki code .sScJk, html code.shiki .sScJk{--shiki-default:#6F42C1;--shiki-dark:#B392F0}html pre.shiki code .sZZnC, html code.shiki .sZZnC{--shiki-default:#032F62;--shiki-dark:#9ECBFF}html .default .shiki span {color: var(--shiki-default);background: var(--shiki-default-bg);font-style: var(--shiki-default-font-style);font-weight: var(--shiki-default-font-weight);text-decoration: var(--shiki-default-text-decoration);}html .shiki span {color: var(--shiki-default);background: var(--shiki-default-bg);font-style: var(--shiki-default-font-style);font-weight: var(--shiki-default-font-weight);text-decoration: var(--shiki-default-text-decoration);}html .dark .shiki span {color: var(--shiki-dark);background: var(--shiki-dark-bg);font-style: var(--shiki-dark-font-style);font-weight: var(--shiki-dark-font-weight);text-decoration: var(--shiki-dark-text-decoration);}html.dark .shiki span {color: var(--shiki-dark);background: var(--shiki-dark-bg);font-style: var(--shiki-dark-font-style);font-weight: var(--shiki-dark-font-weight);text-decoration: var(--shiki-dark-text-decoration);}",{"title":149,"searchDepth":163,"depth":163,"links":1745},[1746,1747,1753],{"id":1411,"depth":163,"text":1412},{"id":1503,"depth":163,"text":1504,"children":1748},[1749,1751],{"id":1510,"depth":220,"text":1750},"1. Auditing Writing Quality (audit_narrative.py in Humanizer 2.0)",{"id":1608,"depth":220,"text":1752},"2. Auditing Code & Skill Safety (scan_generated_skill.py & Vitest)",{"id":1653,"depth":163,"text":1654},"Visual Philosophy","October 3, 2026","\u002Fimages\u002Fnotes\u002Fai-model-benchmarks-that-help-with-real-decisions.webp",{"sections":1758},[1759,1764],{"marker":606,"heading":1760,"body":1761},"Why Leaderboard Scores Do Not Predict Production Reliability",[1762,1763],"Every model launch comes with a bar chart showing a 3% lead on MMLU, GPQA, or SWE-bench. Yet if you have ever swapped a new model ID into an existing production pipeline and watched it break your JSON schema or ignore a negative constraint, you know that public benchmarks measure general capability, not task adherence.","When building Humanizer 2.0 and Gemini Advisor in 2026, I built small, deterministic Python evaluations scored by executable code rather than LLM-as-a-judge.",{"marker":612,"heading":1765,"body":1766},"Designing a Repeatable Eval—and Knowing What It Cannot Measure",[1767,1768],"In audit_narrative.py, we score text against concrete NarraBench discourse markers: moralizing tail paragraphs, somatic clichés, temporal anachrony, and explicit emotion labels. Every article in this feed must pass under the 35% slop threshold before shipping.","A good benchmark tells you immediately when a prompt or model change regresses on known failure modes—while leaving taste, truthfulness, and domain judgment to human review.","\u002Fnotes\u002Fai-model-benchmarks-that-help-with-real-decisions","Public leaderboards tell you how well a model solves olympiad math puzzles; a 40-case domain eval tells you whether upgrading models tomorrow will break your production parser.",{"title":1406,"description":149},"ai-model-benchmarks-that-help-with-real-decisions","notes\u002Fai-model-benchmarks-that-help-with-real-decisions",[1775,1776,1777],"Public AI leaderboards measure general trivia and math puzzles; they do not tell you if a new model will follow your business rules.","A practical internal benchmark only needs 30–50 real examples from your own work, scored by a fast Python or TypeScript script.","Always write down what your benchmark CANNOT measure (factual truth, customer empathy, product strategy) so a human still checks what matters.","BGlQCer4J0P6ALYZF8fl9yuiG0KxFDOly-gBZojmUjY",{"id":1780,"title":1781,"author":8,"authorAvatar":9,"authorRole":10,"body":1782,"category":1378,"date":2269,"description":149,"excerpt":599,"extension":600,"image":2270,"index":2271,"meta":2272,"navigation":223,"path":2284,"pullQuote":2285,"readingTime":2286,"seo":2287,"slug":2288,"stem":2289,"takeaways":2290,"total":627,"__hash__":2294},"notes\u002Fnotes\u002Fdo-you-need-a-database-for-your-blog.md","Do You Need a Database for Your Blog?",{"type":12,"value":1783,"toc":2265},[1784,1788,1791,1798,1863,1869,1876,2102,2109,2113,2119,2236,2242,2262],[15,1785,1787],{"id":1786},"why-git-backed-markdown-fits-technical-studios-and-small-businesses","Why Git-Backed Markdown Fits Technical Studios and Small Businesses",[20,1789,1790],{},"If you have ever returned to a company blog or portfolio after six months only to find that the cloud database paused itself due to inactivity, a plugin update broke the login screen, or a hosted CMS doubled its monthly subscription fee, you have paid the hidden tax of running a live database for static articles. Years ago, I was annoyed when a five-article engineering blog I maintained went offline on a Sunday morning simply because a managed database rotated its connection credentials during a maintenance window.",[20,1792,1793,1794,1797],{},"When I built the ",[29,1795,1796],{},"Studio Notes"," publishing system for this Nuxt 4 portfolio, I wanted both engineering reliability and easy browser editing—without running a separate database server.",[34,1799,1800,1805],{},[20,1801,1802],{},[29,1803,1804],{},"In Plain English — What Are Markdown, Frontmatter, and Git-Backed Content?",[42,1806,1807,1824,1839],{},[45,1808,1809,1815,1816,1819,1820,1823],{},[29,1810,1811,1812,59],{},"Markdown (",[24,1813,1814],{},".md",": Clean, human-readable text files where ",[24,1817,1818],{},"#"," makes a heading and ",[24,1821,1822],{},"**bold**"," makes bold text. Even 20 years from now, any text editor on any computer can open a Markdown file—zero vendor lock-in.",[45,1825,1826,1829,1830,255,1832,255,1834,1177,1836,1838],{},[29,1827,1828],{},"Frontmatter",": A neat block of key-value labels at the very top of an article file (like ",[24,1831,317],{},[24,1833,327],{},[24,1835,322],{},[24,1837,337],{},") that tells the website how to organize and display the card.",[45,1840,1841,1850,1851,1854,1855,1858,1859,1862],{},[29,1842,1843,1844,1729,1847,59],{},"Git-Backed CMS (",[24,1845,1846],{},"@nuxt\u002Fcontent",[24,1848,1849],{},"nuxt-studio",": Instead of storing articles in a separate database server that can crash or get hacked, your articles live right inside the website's version-controlled folder (",[24,1852,1853],{},"content\u002Fnotes\u002F*.md","). With ",[29,1856,1857],{},"Nuxt Studio"," mounted at ",[24,1860,1861],{},"\u002F_studio",", non-technical editors can still log in through their browser, edit text and images visually, and click \"Publish\" to save changes cleanly to GitHub.",[20,1864,1865],{},[120,1866],{"alt":1867,"src":1868},"Git-Backed Markdown vs Relational Database Architecture","\u002Fimages\u002Fnotes\u002Fdiagrams\u002Fdo-you-need-a-database-for-your-blog.svg",[20,1870,1871,1872,1875],{},"Here is the actual schema contract in ",[24,1873,1874],{},"content.config.ts"," that powers the feed you are reading right now:",[144,1877,1879],{"className":146,"code":1878,"language":148,"meta":149,"style":149},"import { defineContentConfig, defineCollection, z } from '@nuxt\u002Fcontent'\n\nexport default defineContentConfig({\n  collections: {\n    notes: defineCollection({\n      type: 'page',\n      source: 'notes\u002F*.md',\n      schema: z.object({\n        index: z.string(),\n        slug: z.string().optional(),\n        title: z.string(),\n        category: z.enum([\n          'Visual Philosophy',\n          'Engineering Craft',\n          'Motion Architecture',\n          'Design Systems',\n        ]),\n        date: z.string(),\n        readingTime: z.string(),\n        image: z.string(),\n        excerpt: z.string(),\n      }),\n    }),\n  },\n})\n",[24,1880,1881,1895,1899,1913,1918,1928,1939,1949,1959,1970,1985,1994,2005,2012,2019,2026,2033,2039,2049,2059,2069,2079,2085,2091,2097],{"__ignoreMap":149},[153,1882,1883,1886,1889,1892],{"class":155,"line":156},[153,1884,1885],{"class":166},"import",[153,1887,1888],{"class":170}," { defineContentConfig, defineCollection, z } ",[153,1890,1891],{"class":166},"from",[153,1893,1894],{"class":194}," '@nuxt\u002Fcontent'\n",[153,1896,1897],{"class":155,"line":163},[153,1898,224],{"emptyLinePlaceholder":223},[153,1900,1901,1904,1907,1910],{"class":155,"line":220},[153,1902,1903],{"class":166},"export",[153,1905,1906],{"class":166}," default",[153,1908,1909],{"class":187}," defineContentConfig",[153,1911,1912],{"class":170},"({\n",[153,1914,1915],{"class":155,"line":227},[153,1916,1917],{"class":170},"  collections: {\n",[153,1919,1920,1923,1926],{"class":155,"line":233},[153,1921,1922],{"class":170},"    notes: ",[153,1924,1925],{"class":187},"defineCollection",[153,1927,1912],{"class":170},[153,1929,1930,1933,1936],{"class":155,"line":269},[153,1931,1932],{"class":170},"      type: ",[153,1934,1935],{"class":194},"'page'",[153,1937,1938],{"class":170},",\n",[153,1940,1941,1944,1947],{"class":155,"line":296},[153,1942,1943],{"class":170},"      source: ",[153,1945,1946],{"class":194},"'notes\u002F*.md'",[153,1948,1938],{"class":170},[153,1950,1951,1954,1957],{"class":155,"line":358},[153,1952,1953],{"class":170},"      schema: z.",[153,1955,1956],{"class":187},"object",[153,1958,1912],{"class":170},[153,1960,1961,1964,1967],{"class":155,"line":3},[153,1962,1963],{"class":170},"        index: z.",[153,1965,1966],{"class":187},"string",[153,1968,1969],{"class":170},"(),\n",[153,1971,1972,1975,1977,1980,1983],{"class":155,"line":369},[153,1973,1974],{"class":170},"        slug: z.",[153,1976,1966],{"class":187},[153,1978,1979],{"class":170},"().",[153,1981,1982],{"class":187},"optional",[153,1984,1969],{"class":170},[153,1986,1987,1990,1992],{"class":155,"line":946},[153,1988,1989],{"class":170},"        title: z.",[153,1991,1966],{"class":187},[153,1993,1969],{"class":170},[153,1995,1996,1999,2002],{"class":155,"line":952},[153,1997,1998],{"class":170},"        category: z.",[153,2000,2001],{"class":187},"enum",[153,2003,2004],{"class":170},"([\n",[153,2006,2007,2010],{"class":155,"line":958},[153,2008,2009],{"class":194},"          'Visual Philosophy'",[153,2011,1938],{"class":170},[153,2013,2014,2017],{"class":155,"line":964},[153,2015,2016],{"class":194},"          'Engineering Craft'",[153,2018,1938],{"class":170},[153,2020,2021,2024],{"class":155,"line":970},[153,2022,2023],{"class":194},"          'Motion Architecture'",[153,2025,1938],{"class":170},[153,2027,2028,2031],{"class":155,"line":976},[153,2029,2030],{"class":194},"          'Design Systems'",[153,2032,1938],{"class":170},[153,2034,2036],{"class":155,"line":2035},17,[153,2037,2038],{"class":170},"        ]),\n",[153,2040,2042,2045,2047],{"class":155,"line":2041},18,[153,2043,2044],{"class":170},"        date: z.",[153,2046,1966],{"class":187},[153,2048,1969],{"class":170},[153,2050,2052,2055,2057],{"class":155,"line":2051},19,[153,2053,2054],{"class":170},"        readingTime: z.",[153,2056,1966],{"class":187},[153,2058,1969],{"class":170},[153,2060,2062,2065,2067],{"class":155,"line":2061},20,[153,2063,2064],{"class":170},"        image: z.",[153,2066,1966],{"class":187},[153,2068,1969],{"class":170},[153,2070,2072,2075,2077],{"class":155,"line":2071},21,[153,2073,2074],{"class":170},"        excerpt: z.",[153,2076,1966],{"class":187},[153,2078,1969],{"class":170},[153,2080,2082],{"class":155,"line":2081},22,[153,2083,2084],{"class":170},"      }),\n",[153,2086,2088],{"class":155,"line":2087},23,[153,2089,2090],{"class":170},"    }),\n",[153,2092,2094],{"class":155,"line":2093},24,[153,2095,2096],{"class":170},"  },\n",[153,2098,2100],{"class":155,"line":2099},25,[153,2101,372],{"class":170},[20,2103,2104,2105,2108],{},"Before any update goes live, our automated test suite (",[24,2106,2107],{},"tests\u002Fportfolio.test.ts",") checks every Markdown file to verify that no title, date, or cover image is missing.",[15,2110,2112],{"id":2111},"markdown-files-vs-hosted-cms-vs-custom-database-how-to-choose","Markdown Files vs. Hosted CMS vs. Custom Database: How to Choose",[20,2114,2115,2116],{},"You see, the right choice depends on one practical question: ",[29,2117,2118],{},"Does your data change only when an author publishes an article, or does it change every minute while customers interact with the site?",[464,2120,2121,2145],{},[467,2122,2123],{},[470,2124,2125,2128,2135,2138],{},[473,2126,2127],{"align":1287},"What Matters to You",[473,2129,2130,2131,1729,2133,59],{"align":1287},"File-Based Markdown (",[24,2132,1846],{},[24,2134,1857],{},[473,2136,2137],{"align":1287},"Hosted Headless CMS (Sanity \u002F Strapi \u002F Contentful)",[473,2139,2140,2141,2144],{"align":1287},"Custom Database (",[24,2142,2143],{},"PostgreSQL"," + Admin Portal)",[483,2146,2147,2166,2182,2205],{},[470,2148,2149,2154,2160,2163],{},[488,2150,2151],{"align":1287},[29,2152,2153],{},"How You Edit Content",[488,2155,2156,2157,2159],{"align":1287},"Visual browser editor at ",[24,2158,1861],{}," OR directly in code; full undo history on every line.",[488,2161,2162],{"align":1287},"Cloud dashboard with user roles; triggers webhooks to update the site.",[488,2164,2165],{"align":1287},"Custom admin forms built specifically for your business workflows.",[470,2167,2168,2173,2176,2179],{},[488,2169,2170],{"align":1287},[29,2171,2172],{},"What Can Break at 2 AM?",[488,2174,2175],{"align":1287},"Almost nothing—articles are pre-packaged with the site at build time.",[488,2177,2178],{"align":1287},"Third-party API outages, expired tokens, or rate limits.",[488,2180,2181],{"align":1287},"Database connection limits, server updates, and backup maintenance.",[470,2183,2184,2189,2196,2199],{},[488,2185,2186],{"align":1287},[29,2187,2188],{},"Quality Protection",[488,2190,2191,2192,2195],{"align":1287},"Broken fields or missing images are blocked automatically by ",[24,2193,2194],{},"pnpm test"," before launch.",[488,2197,2198],{"align":1287},"Editors can accidentally publish incomplete fields unless custom webhooks check them.",[488,2200,2201,2202,2204],{"align":1287},"Protected by database constraints and server-side ",[24,2203,1262],{}," validation.",[470,2206,2207,2212,2220,2228],{},[488,2208,2209],{"align":1287},[29,2210,2211],{},"Monthly Running Cost",[488,2213,2214,2219],{"align":1287},[29,2215,2216],{},[24,2217,2218],{},"$0 \u002F month"," (included in standard web hosting).",[488,2221,2222,2227],{"align":1287},[29,2223,2224],{},[24,2225,2226],{},"$0–$99+ \u002F month"," once you add team seats or traffic.",[488,2229,2230,2235],{"align":1287},[29,2231,2232],{},[24,2233,2234],{},"$15–$60+ \u002F month"," for managed database hosting and backups.",[20,2237,2238,2239,2241],{},"I felt relieved once this site's entire blog was consolidated into ",[24,2240,1853],{},": pages load instantaneously, backups are automatic in Git, and there is zero database maintenance overhead.",[42,2243,2244,2254],{},[45,2245,2246,2253],{},[29,2247,2248,2249,1729,2251,59],{},"Choose File-Based Markdown (",[24,2250,1846],{},[24,2252,1849],{}," for company blogs, case studies, documentation, and service pages (5 to 500 articles) where speed, zero monthly fees, and simplicity win.",[45,2255,2256,2261],{},[29,2257,2258,2259,59],{},"Choose a Relational Database (",[24,2260,2143],{}," when customers create records live on the site (appointment bookings, user accounts, order history, or 10,000+ searchable product SKUs).",[583,2263,2264],{},"html pre.shiki code .szBVR, html code.shiki .szBVR{--shiki-default:#D73A49;--shiki-dark:#F97583}html pre.shiki code .sVt8B, html code.shiki .sVt8B{--shiki-default:#24292E;--shiki-dark:#E1E4E8}html pre.shiki code .sZZnC, html code.shiki .sZZnC{--shiki-default:#032F62;--shiki-dark:#9ECBFF}html pre.shiki code .sScJk, html code.shiki .sScJk{--shiki-default:#6F42C1;--shiki-dark:#B392F0}html .default .shiki span {color: var(--shiki-default);background: var(--shiki-default-bg);font-style: var(--shiki-default-font-style);font-weight: var(--shiki-default-font-weight);text-decoration: var(--shiki-default-text-decoration);}html .shiki span {color: var(--shiki-default);background: var(--shiki-default-bg);font-style: var(--shiki-default-font-style);font-weight: var(--shiki-default-font-weight);text-decoration: var(--shiki-default-text-decoration);}html .dark .shiki span {color: var(--shiki-dark);background: var(--shiki-dark-bg);font-style: var(--shiki-dark-font-style);font-weight: var(--shiki-dark-font-weight);text-decoration: var(--shiki-dark-text-decoration);}html.dark .shiki span {color: var(--shiki-dark);background: var(--shiki-dark-bg);font-style: var(--shiki-dark-font-style);font-weight: var(--shiki-dark-font-weight);text-decoration: var(--shiki-dark-text-decoration);}",{"title":149,"searchDepth":163,"depth":163,"links":2266},[2267,2268],{"id":1786,"depth":163,"text":1787},{"id":2111,"depth":163,"text":2112},"October 24, 2026","\u002Fimages\u002Fnotes\u002Fdo-you-need-a-database-for-your-blog.webp","06",{"sections":2273},[2274,2279],{"marker":606,"heading":2275,"body":2276},"Why Git-Backed Markdown Fits Technical Studios",[2277,2278],"When I upgraded this portfolio to Nuxt 4, I evaluated whether to back Studio Notes with a Postgres table, a hosted headless CMS, or file-based Markdown via @nuxt\u002Fcontent v3 and nuxt-studio.","Storing articles in content\u002Fnotes\u002F*.md validated by a Zod schema in content.config.ts means every post is versioned in Git, tested in CI by Vitest, and deployed atomically with the site.",{"marker":612,"heading":2280,"body":2281},"When to Stay File-Based and When to Reach for Postgres",[2282,2283],"Adding nuxt-studio at \u002F_studio solves the non-technical editing gap: you or a client can edit Markdown and frontmatter visually in the browser and commit straight back to GitHub without touching a terminal.","A database only earns its keep when your content changes independent of deploys—such as multi-tenant user submissions, live inventory, or 50,000 dynamic catalog records.","\u002Fnotes\u002Fdo-you-need-a-database-for-your-blog","Running a managed Postgres instance and an auth server just to publish ten technical essays a year is operational cosplay.","6 min read",{"title":1781,"description":149},"do-you-need-a-database-for-your-blog","notes\u002Fdo-you-need-a-database-for-your-blog",[2291,2292,2293],"For a portfolio or company blog (5–500 posts), file-based Markdown (@nuxt\u002Fcontent + Nuxt Studio) means $0 database bills and zero server outages.","Non-technical editors still get a visual browser interface at \u002F_studio that saves changes directly to GitHub.","Reserve a live database (PostgreSQL) for data that changes while users are on the site: bookings, customer accounts, or 10,000+ catalog items.","N2zInUfXWcY5ECYH-mE4N2Qm3VQF9lZtwwiuPcAfenQ",{"id":2296,"title":2297,"author":8,"authorAvatar":9,"authorRole":10,"body":2298,"category":597,"date":2648,"description":149,"excerpt":599,"extension":600,"image":2649,"index":2650,"meta":2651,"navigation":223,"path":2663,"pullQuote":2664,"readingTime":1395,"seo":2665,"slug":2666,"stem":2667,"takeaways":2668,"total":627,"__hash__":2672},"notes\u002Fnotes\u002Fkeeping-a-data-heavy-interface-responsive.md","Keeping a Data-Heavy Interface Responsive",{"type":12,"value":2299,"toc":2641},[2300,2304,2319,2343,2374,2388,2394,2398,2401,2439,2495,2503,2623,2638],[15,2301,2303],{"id":2302},"why-standard-web-elements-freeze-under-heavy-data","Why Standard Web Elements Freeze Under Heavy Data",[20,2305,2306,2307,2310,2311,2314,2315,2318],{},"To make a web application feel as smooth as a native desktop tool, the browser needs to draw ",[29,2308,2309],{},"60 frames per second (60 FPS)",". Simple division tells the whole story: ",[24,2312,2313],{},"1,000 milliseconds ÷ 60 frames = 16.67 milliseconds per frame",". Once you subtract the browser's own housekeeping, your code has roughly ",[29,2316,2317],{},"10 milliseconds"," to finish its work before a user dragging a slider or scrolling a table feels a stutter.",[20,2320,2321,2322,2325,2326,2329,2330,717,2332,2335,2336,527,2339,2342],{},"In 2025, while building the ",[29,2323,2324],{},"Algorithm Engine","—an interactive visualizer for pathfinding (A*, Dijkstra, Bellman-Ford) and sorting algorithms—I wanted users to scrub through ",[29,2327,2328],{},"10,000 live connected nodes"," in real time. If you have ever tried to draw 10,000 separate ",[24,2331,647],{},[24,2333,2334],{},"\u003Csvg>"," elements in Vue or React and update them sixty times a second, you know what happens: the browser tab locks up completely, dropping to ",[24,2337,2338],{},"7 FPS",[24,2340,2341],{},"138ms"," per frame).",[34,2344,2345,2350],{},[20,2346,2347],{},[29,2348,2349],{},"In Plain English — Why Do Data-Heavy Dashboards Freeze?",[42,2351,2352,2362,2368],{},[45,2353,2354,2357,2358,2361],{},[29,2355,2356],{},"The Main Thread",": By default, a browser tab does almost everything on a single assembly line called the ",[62,2359,2360],{},"Main Thread","—listening to your mouse clicks, scrolling the page, running JavaScript math, and painting pixels. If a heavy calculation hogs that assembly line for 100 milliseconds, buttons cannot even highlight when you hover over them.",[45,2363,2364,2367],{},[29,2365,2366],{},"The DOM (Document Object Model)",": The browser's internal tree of HTML boxes, text, and buttons. It is designed for documents and forms, where it carefully calculates margins, borders, and screen-reader rules for every single box. Asking the DOM to recalculate layout for 10,000 moving dots every 16 milliseconds is like mailing 10,000 individually gift-wrapped boxes when you just needed to ship a pallet of gravel.",[45,2369,2370,2373],{},[29,2371,2372],{},"Web Worker",": A separate background thread (like opening a second assembly line in the back room) that crunches heavy math without ever blocking the user's mouse or keyboard on the Main Thread.",[20,2375,2376,2377,2379,2380,2383,2384,2387],{},"My first fix was moving the pathfinding math into a background ",[29,2378,2372],{},". Yet when I profiled the app, I was annoyed to see the frame rate still dipping into the 30s. Why? Because every time the background worker sent an array of 10,000 JavaScript objects back to the main thread (",[24,2381,2382],{},"worker.postMessage(nodes)","), the browser spent ",[29,2385,2386],{},"9 milliseconds"," photocopying (\"cloning\") all 10,000 objects across threads and cleaning up the old copies in memory.",[20,2389,2390],{},[120,2391],{"alt":2392,"src":2393},"60FPS Web Worker + Transferable ArrayBuffer + Canvas 2D Architecture","\u002Fimages\u002Fnotes\u002Fdiagrams\u002Fkeeping-a-data-heavy-interface-responsive.svg",[15,2395,2397],{"id":2396},"zero-copy-typed-arrays-and-html5-canvas-2d","Zero-Copy Typed Arrays and HTML5 Canvas 2D",[20,2399,2400],{},"You see, moving heavy work to a background thread only keeps the UI fast if handing the results back is nearly free. I solved this with two architectural changes:",[777,2402,2403,2425],{},[45,2404,2405,2414,2415,2418,2419,527,2422,211],{},[29,2406,2407,2408,1613,2411,59],{},"Zero-Copy Transferable Buffers (",[24,2409,2410],{},"Uint32Array",[24,2412,2413],{},"Float32Array",": Instead of storing 10,000 separate JavaScript objects, we pack all node coordinates and states into a flat, contiguous block of numbers in memory (a ",[62,2416,2417],{},"Typed Array","). When the worker finishes a frame, instead of photocopying 10,000 objects, it transfers the memory key directly to the main thread in ",[29,2420,2421],{},"under 0.1 milliseconds",[24,2423,2424],{},"[frameBuffer.buffer]",[45,2426,2427,2430,2431,2434,2435,2438],{},[29,2428,2429],{},"Hybrid Vue 3 + HTML5 Canvas 2D Rendering",": We let ",[29,2432,2433],{},"Vue 3"," manage the accessible buttons, speed sliders, and dropdowns in standard HTML, while all 10,000 graph nodes are painted onto a single ",[24,2436,2437],{},"\u003Ccanvas>"," bitmap surface—one element in the DOM instead of 10,000.",[144,2440,2442],{"className":146,"code":2441,"language":148,"meta":149,"style":149},"\u002F\u002F Inside the Web Worker: zero-copy ownership transfer back to the UI thread\nconst frameBuffer = new Uint32Array(nodeCount)\nstepDijkstraWavefront(graphTopology, frameBuffer)\n\n\u002F\u002F Transfers the raw memory pointer in \u003C0.1ms instead of cloning 10,000 objects\nself.postMessage({ tick, buffer: frameBuffer.buffer }, [frameBuffer.buffer])\n",[24,2443,2444,2449,2467,2475,2479,2484],{"__ignoreMap":149},[153,2445,2446],{"class":155,"line":156},[153,2447,2448],{"class":159},"\u002F\u002F Inside the Web Worker: zero-copy ownership transfer back to the UI thread\n",[153,2450,2451,2453,2456,2458,2461,2464],{"class":155,"line":163},[153,2452,167],{"class":166},[153,2454,2455],{"class":174}," frameBuffer",[153,2457,278],{"class":166},[153,2459,2460],{"class":166}," new",[153,2462,2463],{"class":187}," Uint32Array",[153,2465,2466],{"class":170},"(nodeCount)\n",[153,2468,2469,2472],{"class":155,"line":220},[153,2470,2471],{"class":187},"stepDijkstraWavefront",[153,2473,2474],{"class":170},"(graphTopology, frameBuffer)\n",[153,2476,2477],{"class":155,"line":227},[153,2478,224],{"emptyLinePlaceholder":223},[153,2480,2481],{"class":155,"line":233},[153,2482,2483],{"class":159},"\u002F\u002F Transfers the raw memory pointer in \u003C0.1ms instead of cloning 10,000 objects\n",[153,2485,2486,2489,2492],{"class":155,"line":269},[153,2487,2488],{"class":170},"self.",[153,2490,2491],{"class":187},"postMessage",[153,2493,2494],{"class":170},"({ tick, buffer: frameBuffer.buffer }, [frameBuffer.buffer])\n",[129,2496,2498,2499,2502],{"id":2497},"reproducible-benchmark-10000-concurrent-nodes-chrome-on-macos-m-series","Reproducible Benchmark (",[24,2500,2501],{},"10,000"," Concurrent Nodes, Chrome on macOS M-Series)",[464,2504,2505,2524],{},[467,2506,2507],{},[470,2508,2509,2512,2515,2518,2521],{},[473,2510,2511],{"align":1287},"Rendering & Threading Architecture",[473,2513,2514],{"align":1287},"Frame Time (p95)",[473,2516,2517],{"align":1287},"Main-Thread Blocking (TBT)",[473,2519,2520],{"align":1287},"Sustained Frame Rate",[473,2522,2523],{"align":1287},"Memory Garbage Churn",[483,2525,2526,2554,2585],{},[470,2527,2528,2533,2538,2543,2549],{},[488,2529,2530],{"align":1287},[29,2531,2532],{},"1. Naive Main-Thread DOM \u002F SVG",[488,2534,2535],{"align":1287},[24,2536,2537],{},"138.4 ms",[488,2539,2540],{"align":1287},[24,2541,2542],{},"112.0 ms \u002F frame",[488,2544,2545,2548],{"align":1287},[24,2546,2547],{},"7–9 FPS"," (Frozen UI)",[488,2550,2551],{"align":1287},[24,2552,2553],{},"~48 MB \u002F sec",[470,2555,2556,2564,2569,2574,2580],{},[488,2557,2558],{"align":1287},[29,2559,2560,2561,2563],{},"2. Web Worker + Object ",[24,2562,2491],{}," + Canvas",[488,2565,2566],{"align":1287},[24,2567,2568],{},"24.1 ms",[488,2570,2571],{"align":1287},[24,2572,2573],{},"9.2 ms \u002F frame",[488,2575,2576,2579],{"align":1287},[24,2577,2578],{},"38–42 FPS"," (Stutters)",[488,2581,2582],{"align":1287},[24,2583,2584],{},"~19 MB \u002F sec",[470,2586,2587,2595,2602,2609,2616],{},[488,2588,2589],{"align":1287},[29,2590,2591,2592,2594],{},"3. Web Worker + Transferable ",[24,2593,2410],{}," + Canvas 2D",[488,2596,2597],{"align":1287},[29,2598,2599],{},[24,2600,2601],{},"6.8 ms",[488,2603,2604],{"align":1287},[29,2605,2606],{},[24,2607,2608],{},"0.0 ms",[488,2610,2611],{"align":1287},[29,2612,2613,2615],{},[24,2614,530],{}," (Locked)",[488,2617,2618],{"align":1287},[29,2619,2620],{},[24,2621,2622],{},"\u003C0.5 MB \u002F sec",[20,2624,2625,2626,2629,2630,2633,2634,2637],{},"I used this exact same pattern months later on the dashboard for ",[29,2627,2628],{},"Kernel Telemetry",", where a ",[24,2631,2632],{},"\u003C15MB"," Go backend streams ",[29,2635,2636],{},"50,000+ Linux system events per second"," over WebSockets. A background worker ingests the incoming stream into a fixed-size circular ring buffer and hands 60Hz slices to the Canvas charts. I felt relieved watching our 4-hour soak test finish with flat memory usage and instant button responsiveness while 1,000 simulated servers hammered the feed.",[583,2639,2640],{},"html pre.shiki code .sJ8bj, html code.shiki .sJ8bj{--shiki-default:#6A737D;--shiki-dark:#6A737D}html pre.shiki code .szBVR, html code.shiki .szBVR{--shiki-default:#D73A49;--shiki-dark:#F97583}html pre.shiki code .sj4cs, html code.shiki .sj4cs{--shiki-default:#005CC5;--shiki-dark:#79B8FF}html pre.shiki code .sScJk, html code.shiki .sScJk{--shiki-default:#6F42C1;--shiki-dark:#B392F0}html pre.shiki code .sVt8B, html code.shiki .sVt8B{--shiki-default:#24292E;--shiki-dark:#E1E4E8}html .default .shiki span {color: var(--shiki-default);background: var(--shiki-default-bg);font-style: var(--shiki-default-font-style);font-weight: var(--shiki-default-font-weight);text-decoration: var(--shiki-default-text-decoration);}html .shiki span {color: var(--shiki-default);background: var(--shiki-default-bg);font-style: var(--shiki-default-font-style);font-weight: var(--shiki-default-font-weight);text-decoration: var(--shiki-default-text-decoration);}html .dark .shiki span {color: var(--shiki-dark);background: var(--shiki-dark-bg);font-style: var(--shiki-dark-font-style);font-weight: var(--shiki-dark-font-weight);text-decoration: var(--shiki-dark-text-decoration);}html.dark .shiki span {color: var(--shiki-dark);background: var(--shiki-dark-bg);font-style: var(--shiki-dark-font-style);font-weight: var(--shiki-dark-font-weight);text-decoration: var(--shiki-dark-text-decoration);}",{"title":149,"searchDepth":163,"depth":163,"links":2642},[2643,2644],{"id":2302,"depth":163,"text":2303},{"id":2396,"depth":163,"text":2397,"children":2645},[2646],{"id":2497,"depth":220,"text":2647},"Reproducible Benchmark (10,000 Concurrent Nodes, Chrome on macOS M-Series)","October 28, 2026","\u002Fimages\u002Fnotes\u002Fkeeping-a-data-heavy-interface-responsive.webp","05",{"sections":2652},[2653,2658],{"marker":606,"heading":2654,"body":2655},"Why DOM Diffing and JSON postMessage Stall at 60Hz",[2656,2657],"Every frame at 60fps gives you 16.67 milliseconds to handle input, run application logic, recalculate layout, and paint. In 2025, while building my Algorithm Engine and the live dashboard for Kernel Telemetry, I hit that ceiling immediately when rendering 10,000 graph nodes in standard DOM elements.","Even after moving A* and Dijkstra traversal into a Web Worker, calling postMessage with nested JavaScript objects burned 9ms per tick in structured-clone serialization alone.",{"marker":612,"heading":2659,"body":2660},"Zero-Copy Typed Arrays and Dirty-Rect Canvas 2D",[2661,2662],"Packing node coordinates and visitation states into flat Float32Array and Uint32Array buffers and transferring ownership across the worker boundary drops IPC overhead below 0.1 milliseconds.","Pairing that zero-copy worker pipeline with a single HTML5 Canvas 2D surface and ring-buffered telemetry streams keeps main-thread blocking at 0ms under 50,000+ events per second.","\u002Fnotes\u002Fkeeping-a-data-heavy-interface-responsive","React and Vue are brilliant at orchestrating UI controls, but asking the DOM to diff 10,000 SVG nodes sixty times a second is a physics problem, not a framework problem.",{"title":2297,"description":149},"keeping-a-data-heavy-interface-responsive","notes\u002Fkeeping-a-data-heavy-interface-responsive",[2669,2670,2671],"At 60 frames per second, your browser has only 16.67ms per frame—heavy math on the main thread freezes buttons and scrolling.","Moving math into a Web Worker is only half the fix: use Transferable Typed Arrays (Uint32Array) to hand over data in \u003C0.1ms without copying.","Combine Vue\u002FReact for clean UI controls with HTML5 Canvas 2D for rendering thousands of live data points at a locked 60 FPS.","GBO3p1_pBqNiQfFMeVqt5w6ny5PWw7xHr_TGmzoUBJI",{"id":2674,"title":2675,"author":8,"authorAvatar":9,"authorRole":10,"body":2676,"category":1378,"date":2956,"description":149,"excerpt":599,"extension":600,"image":2957,"index":2958,"meta":2959,"navigation":223,"path":2971,"pullQuote":2972,"readingTime":619,"seo":2973,"slug":2974,"stem":2975,"takeaways":2976,"total":627,"__hash__":2980},"notes\u002Fnotes\u002Fmulti-agent-ai-useful-pattern-or-extra-complexity.md","Multi-Agent AI: Useful Pattern or Extra Complexity?",{"type":12,"value":2677,"toc":2952},[2678,2682,2693,2739,2749,2755,2804,2808,2811,2826,2832,2848,2861],[15,2679,2681],{"id":2680},"when-multi-agent-ai-actually-earns-its-complexity","When Multi-Agent AI Actually Earns Its Complexity",[20,2683,2684,2685,2688,2689,2692],{},"If you have ever watched a demo of four AI agents chatting with each other for five minutes to solve a task that one well-written prompt or a short script could finish in ten seconds, you understand why businesses are cautious about \"multi-agent\" hype. Months earlier, while prototyping the first version of ",[29,2686,2687],{},"Founder Suite"," (my 11-stage business validation pipeline), I was annoyed by the opposite extreme too: when I asked a ",[62,2690,2691],{},"single"," AI session to simultaneously act as an enthusiastic product marketer, a skeptical market strategist, and a strict CFO, it averaged all three voices into polite, wishy-washy encouragement.",[34,2694,2695,2700],{},[20,2696,2697],{},[29,2698,2699],{},"In Plain English — Single Agent vs. Multi-Agent vs. Deterministic Script:",[42,2701,2702,2716,2725],{},[45,2703,2704,2707,2708,2711,2712,2715],{},[29,2705,2706],{},"Deterministic Script (Python \u002F TypeScript Code)",": Traditional software rules where ",[24,2709,2710],{},"2 + 2"," always equals ",[24,2713,2714],{},"4"," in under a millisecond. Zero guesswork, zero AI token cost.",[45,2717,2718,2724],{},[29,2719,2720,2721,59],{},"Single-Agent Mode (",[24,2722,2723],{},"solo",": One AI conversation handling a focused task from start to finish—like asking one experienced editor to polish a two-page memo.",[45,2726,2727,2730,2731,2734,2735,2738],{},[29,2728,2729],{},"Multi-Agent System",": Splitting a large job across separate, specialized AI workers that either run ",[29,2732,2733],{},"at the same time in parallel"," (like 10 assistants each summarizing one chapter of a book) or ",[29,2736,2737],{},"check each other's work"," with fresh eyes.",[20,2740,2741,2742,255,2744,1177,2746,2748],{},"Across the three AI systems I maintain in production (",[29,2743,1145],{},[29,2745,1621],{},[29,2747,2687],{},"), adding multi-agent coordination is only worth the extra complexity when at least one of these three conditions is true:",[20,2750,2751],{},[120,2752],{"alt":2753,"src":2754},"Multi-Agent vs Solo vs Deterministic Python Decision Matrix","\u002Fimages\u002Fnotes\u002Fdiagrams\u002Fmulti-agent-ai-useful-pattern-or-extra-complexity.svg",[777,2756,2757,2765,2782],{},[45,2758,2759,1032,2762,2764],{},[29,2760,2761],{},"Fresh Eyes for Quality Control (Context Isolation)",[29,2763,1145],{},", the reviewer agent is launched in a separate session so it never sees the builder's 50,000 tokens of debugging struggle. Because it only inspects the original contract and the final code diff in read-only mode, it catches shortcuts without bias.",[45,2766,2767,2770,2771,2773,2774,2777,2778,2781],{},[29,2768,2769],{},"Parallel Speed on Independent Chunks",": When ",[29,2772,1621],{}," turns a 500-page technical manual into a modular AI skill, Chapter 2 and Chapter 12 do not depend on each other. Launching parallel waves of fast ",[24,2775,2776],{},"Flash"," subagents finishes the entire book in ",[29,2779,2780],{},"under 4 minutes"," instead of 45 minutes one by one.",[45,2783,2784,2787,2788,527,2790,2793,2794,255,2797,1177,2800,2803],{},[29,2785,2786],{},"Genuine Disagreement Across Different Perspectives",": In stage one of ",[29,2789,2687],{},[24,2791,2792],{},"founder-board","), three independent subagents critique a business idea at the same time through three strict, published lenses (",[62,2795,2796],{},"Offers",[62,2798,2799],{},"Monopoly",[62,2801,2802],{},"Product",") before their memos are compared. Because they cannot see each other's drafts while writing, they surface real blind spots instead of polite groupthink.",[15,2805,2807],{"id":2806},"when-a-single-model-or-a-python-script-is-the-better-choice","When a Single Model or a Python Script Is the Better Choice",[20,2809,2810],{},"You see, the most expensive mistake in AI engineering today is paying language models to do work that regular code already does perfectly.",[20,2812,1142,2813,795,2815,2817,2818,2821,2822,2825],{},[29,2814,1621],{},[29,2816,2687],{},", I drew a strict line between ",[29,2819,2820],{},"qualitative judgment"," (where AI agents help) and ",[29,2823,2824],{},"exact calculation"," (where AI models guess and hallucinate):",[144,2827,2830],{"className":2828,"code":2829,"language":1217,"meta":149},[1215],"Book-to-Skill Pipeline:\n  1. scripts\u002Fextract.py (Python)          -> Exact PDF\u002FEPUB parsing, SHA-256 fingerprint, chapter token counts\n  2. Parallel Flash Subagents             -> Summarize chapter frameworks & decision rules in parallel\n  3. tools\u002Fscan_generated_skill.py (Py)   -> Deterministic security & copyright-leak check\n\nFounder Validation Suite:\n  1. 3x Board Subagents + Flash Buyer Wave -> Surface qualitative positioning & customer objections\n  2. van_westendorp.py & unit_economics.py -> Exact price sensitivity bands & break-even unit math\n",[24,2831,2829],{"__ignoreMap":149},[20,2833,2834,2835,2838,2839,255,2842,1177,2845,211],{},"Ask an AI model to calculate profit margins, count exact words in a 200-page PDF, or compute a 36-month cash-flow table, and it will invent numbers that ",[62,2836,2837],{},"look"," right but fail on a calculator. I felt relieved the day I stripped every arithmetic question out of our AI prompts and moved them into deterministic Python tools (",[24,2840,2841],{},"extract.py",[24,2843,2844],{},"unit_economics.py",[24,2846,2847],{},"van_westendorp.py",[20,2849,2850,2851,2854,2855],{},"And for everyday software development, ",[29,2852,2853],{},"Rule #1"," in Gemini Advisor is simple: ",[29,2856,2857,2858,2860],{},"Default to ",[24,2859,2723],{}," (a single model) for small, low-risk tasks.",[464,2862,2863,2876],{},[467,2864,2865],{},[470,2866,2867,2870,2873],{},[473,2868,2869],{"align":1287},"What You Need Done",[473,2871,2872],{"align":1287},"Best Tool for the Job",[473,2874,2875],{"align":1287},"Why It Wins",[483,2877,2878,2899,2919,2937],{},[470,2879,2880,2885,2890],{},[488,2881,2882],{"align":1287},[29,2883,2884],{},"Financial math, price bands, file hashing, or syntax checks",[488,2886,2887],{"align":1287},[29,2888,2889],{},"Deterministic Python \u002F TypeScript Script",[488,2891,2892,2894,2895,2898],{"align":1287},[24,2893,1643],{}," hallucinations, runs in ",[24,2896,2897],{},"\u003C50ms",", and gives the exact same answer every time.",[470,2900,2901,2909,2916],{},[488,2902,2903],{"align":1287},[29,2904,2905,2906,59],{},"Updating 1–2 files with clear instructions (",[24,2907,2908],{},"risk=\"low\"",[488,2910,2911],{"align":1287},[29,2912,2913,2914,1341],{},"Single AI Model (",[24,2915,2723],{},[488,2917,2918],{"align":1287},"No time or money wasted passing messages between multiple agents.",[470,2920,2921,2926,2934],{},[488,2922,2923],{"align":1287},[29,2924,2925],{},"Summarizing 15 book chapters or simulating 100 buyer cards",[488,2927,2928],{"align":1287},[29,2929,2930,2931,2933],{},"Parallel ",[24,2932,2776],{}," Subagent Waves",[488,2935,2936],{"align":1287},"Cuts waiting time from 45 minutes to 4 minutes at low token cost.",[470,2938,2939,2944,2949],{},[488,2940,2941],{"align":1287},[29,2942,2943],{},"Multi-file feature or security-sensitive code refactor",[488,2945,2946],{"align":1287},[29,2947,2948],{},"Planner -> Builder -> Read-Only Auditor",[488,2950,2951],{"align":1287},"Locks file boundaries and verifies the final diff with unbiased fresh eyes.",{"title":149,"searchDepth":163,"depth":163,"links":2953},[2954,2955],{"id":2680,"depth":163,"text":2681},{"id":2806,"depth":163,"text":2807},"October 19, 2026","\u002Fimages\u002Fnotes\u002Fmulti-agent-ai-useful-pattern-or-extra-complexity.webp","07",{"sections":2960},[2961,2966],{"marker":606,"heading":2962,"body":2963},"When Multi-Agent Orchestration Actually Earns Its Overhead",[2964,2965],"If you have ever built a multi-agent system that spent four minutes passing polite JSON messages back and forth only to produce a worse answer than a single prompt, you have met orchestration tax.","Across my own tools built in 2026, spawning subagents only pays off in three cases: isolating context windows so reviewers stay unbiased, embarrassing parallelism across independent slices, and enforcing adversarial analytical lenses.",{"marker":612,"heading":2967,"body":2968},"When to Default to Solo Mode or a Deterministic Python Script",[2969,2970],"In Gemini Advisor, the default route for localized changes is solo mode because coordinating a planner, builder, and reviewer on a 20-line CSS fix wastes tokens and latency.","Even more importantly, in Founder Suite and Book-to-Skill, agents are strictly forbidden from guessing token counts, SHA-256 hashes, or LTV\u002FCAC unit economics—those always run in deterministic Python CLI scripts.","\u002Fnotes\u002Fmulti-agent-ai-useful-pattern-or-extra-complexity","Never spawn three conversational agents to debate a number that a 40-line Python script can calculate deterministically in two milliseconds.",{"title":2675,"description":149},"multi-agent-ai-useful-pattern-or-extra-complexity","notes\u002Fmulti-agent-ai-useful-pattern-or-extra-complexity",[2977,2978,2979],"Never use AI agents to guess math, token counts, or financial margins—always use a deterministic Python script for exact numbers.","For small, focused tasks (1–2 files), a single model ('solo' mode) is faster, cheaper, and cleaner than a multi-agent crew.","Multi-agent workflows only earn their keep in three cases: independent code review, parallel waves of work, or structured disagreement.","mbvQx0BaVI2PbUWtr1SOwzElO00TWe4vORg-0lcn3Qo",{"id":2982,"title":2983,"author":8,"authorAvatar":9,"authorRole":10,"body":2984,"category":1068,"date":3200,"description":149,"excerpt":599,"extension":600,"image":3201,"index":3202,"meta":3203,"navigation":223,"path":3214,"pullQuote":3215,"readingTime":1395,"seo":3216,"slug":3217,"stem":3218,"takeaways":3219,"total":627,"__hash__":3223},"notes\u002Fnotes\u002Fwhat-i-learned-building-a-local-first-sync-engine.md","What I Learned Building a Local-First Sync Engine",{"type":12,"value":2985,"toc":3190},[2986,2990,2993,2999,3048,3054,3058,3061,3069,3079,3095,3099,3102,3120,3144,3147,3151,3154,3173,3177],[15,2987,2989],{"id":2988},"what-breaks-when-every-click-waits-on-the-network","What Breaks When Every Click Waits on the Network",[20,2991,2992],{},"Before I wrote production TypeScript or Go, I spent eleven years working in commercial transit and logistics across Birmingham and Budapest. If you have ever stood inside a reinforced-concrete bus depot at 5:30 AM watching drivers log pre-departure safety checks on a tablet, you know how fragile traditional web applications really are. A tablet drops from 4G to zero bars right as the driver submits a brake-fault report. The loading spinner turns for twenty seconds, the browser tab times out, and the worker—who felt exhausted before the shift even started—refreshes the page and loses the entire inspection form.",[20,2994,1793,2995,2998],{},[29,2996,2997],{},"Local-First Sync Engine"," in 2025, my goal was simple: make a web application behave like a reliable physical notebook when offline, and like a real-time cloud system when online.",[34,3000,3001,3006],{},[20,3002,3003],{},[29,3004,3005],{},"In Plain English — How \"Local-First\" Software Works:",[42,3007,3008,3018],{},[45,3009,3010,3013,3014,3017],{},[29,3011,3012],{},"Traditional Cloud Web App",": Every time you click \"Save,\" your browser has to send a message across the internet to a distant server and wait (",[24,3015,3016],{},"150–400ms",") for confirmation. If your Wi-Fi or mobile signal blinks, you get a spinning wheel or an error.",[45,3019,3020,3029,3030,3036,3037,3040,3041,3043,3044,3047],{},[29,3021,3022,3023,1729,3026,59],{},"Local-First Web App (",[24,3024,3025],{},"SQLite WASM",[24,3027,3028],{},"IndexedDB",": We place a miniature, industrial-strength database right inside your browser tab. ",[29,3031,3032,3033,59],{},"WebAssembly (",[24,3034,3035],{},"WASM"," lets us run the rock-solid ",[29,3038,3039],{},"SQLite"," database engine at near-native speed inside Chrome or Safari, while ",[29,3042,3028],{}," stores the data safely on your device's disk. Every click saves locally in under ",[29,3045,3046],{},"2 milliseconds","—even in airplane mode—and a quiet background worker syncs with the central server whenever signal returns.",[20,3049,3050],{},[120,3051],{"alt":3052,"src":3053},"Local-First Sync Architecture Diagram","\u002Fimages\u002Fnotes\u002Fdiagrams\u002Fwhat-i-learned-building-a-local-first-sync-engine.svg",[15,3055,3057],{"id":3056},"four-failure-modes-that-only-show-up-after-you-go-offline","Four Failure Modes That Only Show Up After You Go Offline",[20,3059,3060],{},"Building a demo that saves data locally takes a weekend; surviving real 12-hour field shifts takes careful engineering around four failure cases. Months earlier, during my first stress test simulating two dispatch tablets offline for a full shift, I was annoyed to watch three naive assumptions fail back-to-back.",[129,3062,3064,3065,3068],{"id":3063},"_1-device-clocks-lie-datenow-cannot-be-trusted","1. Device Clocks Lie (",[24,3066,3067],{},"Date.now()"," Cannot Be Trusted)",[20,3070,3071,3072,3074,3075,3078],{},"When two offline workers edit the same record and reconnect later, how does the server know which edit happened second? My earliest prototype simply checked the tablet's built-in wall clock (",[24,3073,3067],{},"), a rule called ",[62,3076,3077],{},"Last-Write-Wins",". That broke immediately when one test tablet had its system clock running four minutes fast: every edit from that fast tablet silently overwrote newer work from everyone else.",[34,3080,3081],{},[20,3082,3083,3086,3087,3090,3091,3094],{},[29,3084,3085],{},"In Plain English — What is a \"Hybrid Logical Clock\" (HLC)?","\nInstead of trusting a device's wall clock alone, an ",[29,3088,3089],{},"HLC"," stamps every edit with three pieces of information: ",[24,3092,3093],{},"(Physical Time + Step Counter + Device ID)",". Even if a tablet's clock is slightly off, the step counter ticks upward whenever devices exchange messages, guaranteeing a fair, unbroken timeline of cause and effect.",[129,3096,3098],{"id":3097},"_2-safe-field-merges-vs-business-rules-who-gets-the-last-spare-van","2. Safe Field Merges vs. Business Rules (\"Who Gets the Last Spare Van?\")",[20,3100,3101],{},"Not all offline conflicts are equal, and treating them the same way causes real operational bugs:",[42,3103,3104,3110],{},[45,3105,3106,3109],{},[29,3107,3108],{},"Safe Parallel Edits (Column-Level Merge)",": If Dispatcher A updates a vehicle's parking bay while Mechanic B updates the mileage on the same vehicle while offline, both changes touch different fields. The sync engine merges both automatically.",[45,3111,3112,3115,3116,3119],{},[29,3113,3114],{},"Exclusive Resource Conflicts (Idempotent Server Mutators)",": What if two dispatchers, both offline, assign the ",[62,3117,3118],{},"same"," spare van to two different routes? You cannot merge half a van.",[34,3121,3122],{},[20,3123,3124,3127,3128,3131,3132,3135,3136,3139,3140,3143],{},[29,3125,3126],{},"In Plain English — What is an \"Idempotent Mutator\"?","\nA ",[29,3129,3130],{},"Mutator"," is a named business command (like ",[24,3133,3134],{},"\"Assign Van #14 to Route B\"",") rather than a raw spreadsheet cell edit. ",[29,3137,3138],{},"Idempotent"," means that if a spotty mobile connection sends that same command three times in a row, the server executes it ",[62,3141,3142],{},"once"," and never creates three duplicate bookings.",[20,3145,3146],{},"When the tablets reconnect, our Python FastAPI and PostgreSQL server replays those commands in strict order. The first route assignment succeeds; the second tablet receives an instant notification flagging the conflict in a \"Needs Review\" tray so a human dispatcher can pick another vehicle—zero silent overwrites.",[129,3148,3150],{"id":3149},"_3-preventing-browser-storage-eviction-and-two-tab-collisions","3. Preventing Browser Storage Eviction and Two-Tab Collisions",[20,3152,3153],{},"Mobile browsers sometimes clean up website storage if a device runs low on disk space. Early on, I felt relieved when we caught an iOS Safari storage-clearing edge case during lab testing before any real shift logs lived in it. We solved it with two browser standards:",[777,3155,3156,3163],{},[45,3157,3158,3159,3162],{},"Calling ",[24,3160,3161],{},"navigator.storage.persist()"," at login so the browser marks the database as protected operational data.",[45,3164,3165,3166,527,3169,3172],{},"Using the ",[29,3167,3168],{},"Web Locks API",[24,3170,3171],{},"navigator.locks.request",") so if an operator opens two tabs of the app at once, they take turns writing cleanly instead of corrupting the local file.",[129,3174,3176],{"id":3175},"_4-reconnecting-after-12-hours-with-4000-queued-edits","4. Reconnecting After 12 Hours With 4,000 Queued Edits",[20,3178,3179,3180,3183,3184,3187,3188,84],{},"If a field tablet records 4,000 status updates during a 12-hour shift and tries to fire 4,000 separate internet requests the second Wi-Fi reconnects, it freezes the browser. Instead, our background sync worker ",[29,3181,3182],{},"compacts the history"," (if a status changed five times offline, it folds them into a single net update) and streams compressed batches of 250 items at a time. In our benchmark workload, this local-first delta sync reduced central PostgreSQL server read load by ",[29,3185,3186],{},"74%"," while keeping every button click inside the app under ",[29,3189,3046],{},{"title":149,"searchDepth":163,"depth":163,"links":3191},[3192,3193],{"id":2988,"depth":163,"text":2989},{"id":3056,"depth":163,"text":3057,"children":3194},[3195,3197,3198,3199],{"id":3063,"depth":220,"text":3196},"1. Device Clocks Lie (Date.now() Cannot Be Trusted)",{"id":3097,"depth":220,"text":3098},{"id":3149,"depth":220,"text":3150},{"id":3175,"depth":220,"text":3176},"November 7, 2026","\u002Fimages\u002Fnotes\u002Fwhat-i-learned-building-a-local-first-sync-engine.webp","02",{"sections":3204},[3205,3209],{"marker":606,"heading":2989,"body":3206},[3207,3208],"Before moving into full-stack engineering, I spent eleven years in commercial transit and logistics operations in the UK and Hungary. If you have ever watched a dispatcher or driver try to submit a vehicle inspection checklist in a depot dead zone, you have seen what happens when a traditional web app blocks on a 300ms POST request: spinners hang, workers tap Submit three times, and shift records vanish.","In 2025, I built the Local-First Sync Engine so that every read and write hits a browser-resident SQLite-over-WASM and IndexedDB store in under 2 milliseconds. The UI never waits on cellular coverage.",{"marker":612,"heading":3210,"body":3211},"Conflict Resolution, Clock Drift, and Schema Migration Traps",[3212,3213],"Moving the primary database into the client tab solves latency immediately, but it shifts the complexity into synchronization. Standard wall-clock timestamps fail the moment a field tablet's system clock drifts by four minutes.","Replacing raw Date.now() Last-Write-Wins with Hybrid Logical Clocks (HLC) and routing multi-user inventory reservations through idempotent server mutators eliminated lost field mutations while cutting backend PostgreSQL read load by 74%.","\u002Fnotes\u002Fwhat-i-learned-building-a-local-first-sync-engine","A lost cellular connection inside a reinforced-concrete transit depot should never be the web equivalent of an application crash.",{"title":2983,"description":149},"what-i-learned-building-a-local-first-sync-engine","notes\u002Fwhat-i-learned-building-a-local-first-sync-engine",[3220,3221,3222],"Local-first apps save data inside the device browser first (\u003C2ms), eliminating loading spinners and lost forms when Wi-Fi or 4G drops.","Device clocks drift: use Hybrid Logical Clocks (HLC) instead of basic wall-clock timestamps so offline edits never overwrite newer work.","Batching and compacting offline edits before syncing cuts backend PostgreSQL database read load by 74%.","mxgJrFBZ1hpm5i7g_pMpNJDOdxsq67fw_F_la7_3F-w",{"id":3225,"title":3226,"author":8,"authorAvatar":9,"authorRole":10,"body":3227,"category":1378,"date":3487,"description":149,"excerpt":599,"extension":600,"image":3488,"index":3489,"meta":3490,"navigation":223,"path":3501,"pullQuote":3502,"readingTime":619,"seo":3503,"slug":3504,"stem":3505,"takeaways":3506,"total":627,"__hash__":3510},"notes\u002Fnotes\u002Fwhen-should-a-coding-agent-use-a-more-capable-model.md","When Should a Coding Agent Use a More Capable Model?",{"type":12,"value":3228,"toc":3483},[3229,3233,3236,3277,3286,3292,3315,3324,3400,3404,3411,3426,3429,3469,3480],[15,3230,3232],{"id":3231},"scoring-task-risk-before-burning-frontier-tokens","Scoring Task Risk Before Burning Frontier Tokens",[20,3234,3235],{},"Every few weeks, a new AI model launch trends on X and half of the tech timeline declares that model routing is obsolete because the newest flagship model can do everything. Then you actually run that flagship model across a 60-file software project, watch it spend 90 seconds \"thinking\" just to add a button style, and hit your usage quota before lunch. Months earlier, I was frustrated by that exact trap: I was burning my most expensive, highest-latency reasoning model on repetitive formatting tasks, leaving zero capacity for the database and security decisions that actually required deep thought.",[34,3237,3238,3244],{},[20,3239,3240,3243],{},[29,3241,3242],{},"In Plain English — What is \"Selective Model Routing\"?","\nAI models come in different sizes and speeds, much like vehicles in a logistics fleet:",[42,3245,3246,3255,3264],{},[45,3247,3248,3254],{},[29,3249,3250,3251,59],{},"Frontier Reasoning Models (e.g., ",[24,3252,3253],{},"Gemini Pro \u002F Argon",": Like a senior structural engineer—capable of weighing complex trade-offs across an entire building blueprint, but slower and 10–30x more expensive per page.",[45,3256,3257,3263],{},[29,3258,3259,3260,59],{},"Fast Execution Models (e.g., ",[24,3261,3262],{},"Gemini Flash \u002F Flash-Lite",": Like a skilled, high-speed installation crew—once you hand them an exact blueprint and measurements, they execute the work in seconds at a fraction of the cost.",[45,3265,3266,3269,3270,795,3273,3276],{},[29,3267,3268],{},"Selective Model Routing",": Automatically choosing which model handles each step based on ",[29,3271,3272],{},"how ambiguous",[29,3274,3275],{},"how risky"," the task is.",[20,3278,3279,3280,527,3282,3285],{},"When I built the dynamic routing engine in ",[29,3281,1145],{},[24,3283,3284],{},"scripts\u002Fmodel_allocator.sh","), I made risk scoring mandatory before any file is edited:",[20,3287,3288],{},[120,3289],{"alt":3290,"src":3291},"Selective Model Routing Architecture","\u002Fimages\u002Fnotes\u002Fdiagrams\u002Fwhen-should-a-coding-agent-use-a-more-capable-model.svg",[144,3293,3295],{"className":1528,"code":3294,"language":1530,"meta":149,"style":149},"# Dynamic tier check before modifying workspace files\n.\u002Fscripts\u002Fmodel_allocator.sh --check\n# Emits: SELECTIVE ROUTE mode=\"solo|delegate|audit|full\" risk=\"low|medium|high\"\n",[24,3296,3297,3302,3310],{"__ignoreMap":149},[153,3298,3299],{"class":155,"line":156},[153,3300,3301],{"class":159},"# Dynamic tier check before modifying workspace files\n",[153,3303,3304,3307],{"class":155,"line":163},[153,3305,3306],{"class":187},".\u002Fscripts\u002Fmodel_allocator.sh",[153,3308,3309],{"class":174}," --check\n",[153,3311,3312],{"class":155,"line":220},[153,3313,3314],{"class":159},"# Emits: SELECTIVE ROUTE mode=\"solo|delegate|audit|full\" risk=\"low|medium|high\"\n",[20,3316,3317,3318,795,3321],{},"You see, the dividing line between tasks that require a Frontier reasoning model and tasks that run better on a fast execution model comes down to two questions: ",[29,3319,3320],{},"Is the design still ambiguous?",[29,3322,3323],{},"What is the blast radius if a mistake happens?",[464,3325,3326,3346],{},[467,3327,3328],{},[470,3329,3330,3333,3340],{},[473,3331,3332],{"align":1287},"Decision Factor",[473,3334,3335,3336,3339],{"align":1287},"Use Frontier Tier (",[24,3337,3338],{},"Pro"," \u002F Architect)",[473,3341,3342,3343,3345],{"align":1287},"Use Execution Tier (",[24,3344,2776],{}," \u002F Builder)",[483,3347,3348,3361,3374,3387],{},[470,3349,3350,3355,3358],{},[488,3351,3352],{"align":1287},[29,3353,3354],{},"Clarity of Spec",[488,3356,3357],{"align":1287},"Requirements are open-ended; several different designs could work.",[488,3359,3360],{"align":1287},"File paths, data types, and rules are already locked in a written spec.",[470,3362,3363,3368,3371],{},[488,3364,3365],{"align":1287},[29,3366,3367],{},"Blast Radius (Risk)",[488,3369,3370],{"align":1287},"Touches login security, payment logic, database schemas, or shared routing.",[488,3372,3373],{"align":1287},"Confined to 1–3 specific UI components, test files, or content pages.",[470,3375,3376,3381,3384],{},[488,3377,3378],{"align":1287},[29,3379,3380],{},"Cross-File Context",[488,3382,3383],{"align":1287},"Must connect dots across 15+ files or audit a finished diff for subtle bugs.",[488,3385,3386],{"align":1287},"Only needs to read the target file and the 5-part implementation contract.",[470,3388,3389,3394,3397],{},[488,3390,3391],{"align":1287},[29,3392,3393],{},"Cost of an Error",[488,3395,3396],{"align":1287},"Silent data leak, broken customer checkout, or security hole.",[488,3398,3399],{"align":1287},"Immediate compiler or test error that gets caught and fixed in 3 seconds.",[15,3401,3403],{"id":3402},"measuring-the-trade-off-why-fast-models-need-a-5-part-contract","Measuring the Trade-Off: Why Fast Models Need a 5-Part Contract",[20,3405,3406,3407,3410],{},"If you give a fast execution model a vague five-word prompt like ",[62,3408,3409],{},"\"Add rate limiting to the contact form,\""," it will guess—and you will spend an hour cleaning up its assumptions. If you have ever blamed a faster model for \"bad code\" after giving it a one-sentence instruction, the model was not the problem; the missing blueprint was.",[20,3412,3413,3414,3416,3417,3419,3420,3422,3423,3425],{},"In my workflow across ",[29,3415,1145],{}," and the ",[29,3418,1621],{}," distiller, the Frontier model never disappears—it just stops typing out 400 lines of repetitive implementation code. Instead, it writes the ",[29,3421,1163],{}," (exact data shapes, allowed files, forbidden shortcuts, and test commands) and hands that blueprint to a ",[24,3424,2776],{}," builder subagent.",[20,3427,3428],{},"When we measured this split across 20 real feature and refactoring tasks in 2026, three numbers proved the trade-off:",[777,3430,3431,3443,3460],{},[45,3432,3433,3436,3437,3439,3440,84],{},[29,3434,3435],{},"15x Reduction in Frontier Token Usage",": Because trial-and-error compiler fixes happen inside the cheap, disposable ",[24,3438,2776],{}," builder session instead of cluttering the main architect's memory, parent context usage stayed below ",[29,3441,3442],{},"25%",[45,3444,3445,1032,3448,3450,3451,3453,3454,3456,3457,84],{},[29,3446,3447],{},"4-Minute Parallel Processing",[29,3449,1621],{},", after a deterministic Python script (",[24,3452,2841],{},") calculates exact chapter byte offsets, parallel waves of ",[24,3455,2776],{}," subagents distill a 500-page technical manual into structured agent skills in under ",[29,3458,3459],{},"4 minutes",[45,3461,3462,3465,3466,3468],{},[29,3463,3464],{},"Automatic Safety Escalation",": If a ",[24,3467,2776],{}," builder fails the automated test suite twice or hits an unexpected schema conflict, the router automatically escalates back to the Frontier model rather than letting the fast model guess.",[20,3470,3471,3472,3475,3476,3479],{},"I felt relieved once I stopped asking ",[62,3473,3474],{},"\"Which single AI model is best on social media?\""," and started measuring ",[29,3477,3478],{},"verified cost per merged feature",". Use your deepest reasoning model to lock the blueprint and inspect the result, and let the fast model do the heavy typing in between.",[583,3481,3482],{},"html pre.shiki code .sJ8bj, html code.shiki .sJ8bj{--shiki-default:#6A737D;--shiki-dark:#6A737D}html pre.shiki code .sScJk, html code.shiki .sScJk{--shiki-default:#6F42C1;--shiki-dark:#B392F0}html pre.shiki code .sj4cs, html code.shiki .sj4cs{--shiki-default:#005CC5;--shiki-dark:#79B8FF}html .default .shiki span {color: var(--shiki-default);background: var(--shiki-default-bg);font-style: var(--shiki-default-font-style);font-weight: var(--shiki-default-font-weight);text-decoration: var(--shiki-default-text-decoration);}html .shiki span {color: var(--shiki-default);background: var(--shiki-default-bg);font-style: var(--shiki-default-font-style);font-weight: var(--shiki-default-font-weight);text-decoration: var(--shiki-default-text-decoration);}html .dark .shiki span {color: var(--shiki-dark);background: var(--shiki-dark-bg);font-style: var(--shiki-dark-font-style);font-weight: var(--shiki-dark-font-weight);text-decoration: var(--shiki-dark-text-decoration);}html.dark .shiki span {color: var(--shiki-dark);background: var(--shiki-dark-bg);font-style: var(--shiki-dark-font-style);font-weight: var(--shiki-dark-font-weight);text-decoration: var(--shiki-dark-text-decoration);}",{"title":149,"searchDepth":163,"depth":163,"links":3484},[3485,3486],{"id":3231,"depth":163,"text":3232},{"id":3402,"depth":163,"text":3403},"November 1, 2026","\u002Fimages\u002Fnotes\u002Fwhen-should-a-coding-agent-use-a-more-capable-model.webp","04",{"sections":3491},[3492,3496],{"marker":606,"heading":3232,"body":3493},[3494,3495],"Every few weeks a new model release dominates X with claims that it makes every previous workflow obsolete. Yet if you run autonomous coding agents daily, you hit a more practical problem: burning your deepest reasoning tier on mechanical boilerplate exhausts your rate limits in two hours.","In Gemini Advisor, model_allocator.sh inspects active tiers dynamically and enforces an upfront SELECTIVE ROUTE declaration—routing low-risk bounded edits to Gemini Flash while reserving Frontier capacity for architecture and zero-history diff audits.",{"marker":612,"heading":3497,"body":3498},"Measuring the 15x Cost-to-Regression Trade-Off",[3499,3500],"A fast model fails when given a vague prompt, not when given a locked type signature. When the Frontier planner supplies a 5-part contract with explicit file boundaries, Flash completes mechanical implementation slices at 10–30x lower token cost.","The metric that matters is not raw tokens per second; it is verified cost per merged pull request with zero post-merge reverts.","\u002Fnotes\u002Fwhen-should-a-coding-agent-use-a-more-capable-model","Using a frontier reasoning model to rename forty CSS variables is like hiring a structural architect to carry drywall up the stairs.",{"title":3226,"description":149},"when-should-a-coding-agent-use-a-more-capable-model","notes\u002Fwhen-should-a-coding-agent-use-a-more-capable-model",[3507,3508,3509],"Match the AI model to the risk of the task: use deep reasoning models for architecture and security, and fast models for bounded code edits.","Fast models rarely fail because they are 'dumb'—they fail when given vague instructions instead of a written 5-part contract.","Across 20 real engineering tasks, selective routing cut frontier token costs by 15x while keeping parent context usage under 25%.","0nixxZ0GKXTJGGCqrpoiBKpBAgu-B2cUOFsstgB_6zE"]