The meta-guide · every site also has its own /guide
Twenty-five websites in one autonomous run: the honest recipe, so you can reproduce it with your own AI agent. No secret sauce withheld. The process is the product.
The single biggest quality lever wasn't code. It was art direction that forces divergence. Left to defaults, AI-built sites converge on the same dark purple-gradient hero. So every site got a one-paragraph brief with four locked ingredients:
Bans matter as much as briefs: no lorem ipsum, no emoji-bullet feature grids, no three-cards-with-icons, no purple gradients, no Inter-by-default.
Each brief went to an independent builder agent with full autonomy over a single directory. Pure static sites (index.html, CSS, JS, assets) with libraries from CDNs (three.js, GSAP) and fonts from a Google-Fonts mirror. Every site also had to ship a /guide route explaining its own construction, styled in its own design language.
Three of the twenty-five (the fashion house, the restaurant, the photo monograph) were image-led: those agents wrote photographer-style shot lists, generated their imagery with an AI image model, and were instructed to pivot to code-driven visuals if generation failed. Never ship a broken image.
The heart of the pipeline is a ~60-line Playwright script every agent ran after every change:
node shot.js http://localhost:7101 shots/site-pass1
→ writes -top / -mid / -end / -full / -mobile screenshots
→ prints every console error, page error, failed requestThe agent then reads its own screenshots and judges them. This closes the loop that most AI-generated sites never close: the model doesn't hope the code works. It looks.
After the first complete build, each agent was required to run at least three full critique-and-improve cycles: screenshot everything, fix every console error, then find and implement at least five concrete design improvements: composition, type hierarchy, spacing rhythm, color tension, motion polish, detail density. The standing order:
The direction of travel is always richer, more finished, more intentional, never simpler.
Each builder deployed its own site to Vercel (vercel deploy --prod --yes) and verified the live URL with fresh screenshots. Then came the twist that raised the floor of the whole collection: a second agent that had never seen the code audited the live site like an award juror (typography, spacing, contrast, mobile coherence, dead zones, AI-slop tells), fixed everything it found in the source, made at least two craft-elevating improvements of its own, redeployed, and filed an honest 1–10 score with remaining flaws on record. Those are the scores on the index cards.
All fifty agents (25 builders + 25 critics) ran inside one workflow script, a pipeline where each site flowed straight from build to critique without waiting for the others. One hard-won lesson: running all twenty-five at once meant a usage-limit interruption left everything half-finished. Restructured into waves of five that fully complete before the next wave starts, every finished site banked permanently, and interrupted runs resumed for the cost of only the unfinished work.
The prompt skeleton that built every site here:
You are an elite design agent building ONE showcase website.
Portfolio-grade: Awwwards Site-of-the-Day quality. Mediocrity is failure.
BRAND: [fictional brand with a point of view]
ART DIRECTION: [light/dark, exact hex palette, named type direction]
SIGNATURE TECHNIQUE: [one advanced technique to commit to hard]
RULES: real copy throughout, no lorem ipsum; a /guide route explaining
how you built it; never low-contrast text; mobile must be coherent;
kill the generic-AI look: no purple gradients, no default fonts.
LOOP (at least 3 passes): serve locally → screenshot desktop + mobile
+ full page → read your own screenshots → fix every console error →
find 5 concrete design improvements → implement them → repeat.
DEPLOY: push to hosting, screenshot the live URL, fix what's broken.
THEN: hand it to a second agent instructed to be merciless.That's the entire trick. Divergent briefs. Agents that look at their own work. A minimum iteration count. A fresh-eyes critic with permission to be harsh. The rest is taste.
Built over one long day: 25 sites, 50 agents, 3+ passes each, ~2,900 tool invocations, 8M+ tokens, thousands of self-screenshots. Every site live on Vercel with its own /guide.