PixelJury
gchahal1982/pixeljury
Elimina o visual genérico de sites feitos por IA com feedback de design visual
Ver no GitHubAgent que avalia e melhora a qualidade visual de sites gerados por IA, identificando elementos genéricos (AI slop) e fornecendo correções específicas. Integra-se com Cursor, Claude Code, v0 e Lovable, oferecendo scores antes/depois para medir progresso de design.
Instalação
npx convez-ai gchahal1982/pixeljury1 comando no terminal: detecta o formato do repositório (marketplace ou plugin avulso) e instala via Claude Code. Requer Node 18+ e o Claude Code instalado.
Métricas
Casos de uso
- Melhorar aparência visual de sites gerados por IA
- Iterar design com feedback automático e scores mensuráveis
- Detectar e corrigir padrões genéricos de AI slop em interfaces
Veredito

Ferramenta prática que fecha a lacuna real entre código funcional e design polido. Traz perspectiva de designer visual ao fluxo de vibe-coding, com métricas que permitem otimização iterativa — sem exigir expertise em design.
README
A tradução para o português deste README ainda está na fila.
PixelJury
Make your AI-built site stop looking AI-built. 🎨
Your code works — so why does the page still look like a robot made it? PixelJury shows you exactly what's giving it away, hands your AI agent the fix, and proves the glow-up with a score.
You vibe-coded a site with Cursor, Claude Code, v0, Lovable or Bolt. It works. But it has that look — the purple gradient hero, the Inter font, the emoji feature cards, the fake "10k+ / 99.9% / 24-7" stats. Generic. AI-made. Slop.
PixelJury is the taste your agent is missing. It opens your running page, looks at the actual pixels like a picky designer would, and tells you — specifically — what's making it look cheap. Then it writes a ready-to-paste fix for your agent and gives you a design score so you can watch it get better. One example below went from 16 → 84 in a single pass. ✨
No design degree, no Figma, no API key required. One command:
npx pixeljury review http://localhost:3000
⭐ If "my AI site looks like AI made it" is a pain you've felt, star the repo — it helps others find it.
Contents
See it work (before/after) · The iteration loop · Quickstart · How it works · Providers · The rubric · Who it's for · FAQ · Contributing · Roadmap
Why not just use a "design skill"? Those help your agent write nicer code up front — but they never see what actually shows up on screen. PixelJury is the second pair of eyes after the page renders: it looks at the real pixels and catches the slop that slipped through. Use both — generate with whatever you like, then let PixelJury tell you if it's actually good.
See it work — before / after
Same product ("FlowSync"), judged by PixelJury. Both screenshots below are produced by the
tool's own render stage; both pages live in examples/ so you can reproduce them.
Before — examples/sloppy-saas | After — examples/polished-saas |
|---|---|
| 16 / 100 — Broken or pure slop | 84 / 100 — Good, fixable gaps |
| 3 hard fails · 6 slop-trope deductions | 0 hard fails · 0 deductions |
| ✗ mobile overflow · ✗ footer contrast · ✗ 42px touch target | ✓ responsive · ✓ AA contrast · ✓ 44px+ targets |
✗ Inter · gradient hero · emoji icons · left-border cards · 5× repeated cards · ✨ AI-powered badge | ✓ committed editorial direction, none of the tropes |
Full per-dimension breakdowns: before.score.json ·
after.score.json.
The screenshots, hard fails, and slop-trope deductions are fully deterministic — reproducible with no API key. The 0–100 headline number also factors in a vision model's read of the six dimensions (the scores above are a Claude vision pass against rubric v0.1). The key-less
mockprovider still runs the whole deterministic half.
The point isn't the score — it's that you can move it
A linter tells you what's wrong once. PixelJury gives you a number you iterate against: run
it, hand the fix-prompt.md to your agent, run it again, watch it climb. The same page above
went 16 → 84 in one pass — and it doesn't stop there.
| Iteration | Score | What moved it |
|---|---|---|
| v0 — raw AI output | 16 | The generation you started with. 3 hard fails, every slop trope. |
v1 — after one fix-prompt pass | 84 | Killed the hard fails (responsive, AA contrast, 44px targets); dropped the gradient/emoji/fake-stats; committed to a real type + color system. |
| v2 — the next pass (to break 90) | 90+ (target) | The critique PixelJury still surfaces: one distinctive layout move (Originality 80→88), let the accent carry the primary CTA (Hierarchy 84→90), swap in real customer logos (Polish). |
That loop — review → fix → re-review — is the product. The score is just how you know the last change actually helped, and where the next 6 points are hiding.
Quickstart
# point it at your running dev server
npx pixeljury review http://localhost:3000 --provider anthropic
PixelJury · rubric v0.1
Design score: 61/100 — Generic AI output
Hard fails
✗ Page is 70px wider than the 390px viewport (horizontal scroll on mobile) caps at 70
✗ Text contrast 3.1:1 (needs 4.5:1) — e.g. "© 2026 FlowSync…" caps at 65
Problems
✗ "inter" is the primary typeface — an AI/template default −5
✗ Large gradient background panel (1440×420px) −5
✗ The same card pattern is reused 5× — reads as templated −5
✗ Pill/badge with AI-marketing copy — "✨ AI-powered…" −3
✗ Layout & hierarchy weak hierarchy 50
→ pixeljury/critique.md pixeljury/fix-prompt.md pixeljury/screenshot.png
Run your agent on fix-prompt.md, then re-run to see the score move.
Then hand pixeljury/fix-prompt.md to your agent, let it rework the page, and run it again
to watch the score climb. The before/after is the point.
What it produces
Every run writes to ./pixeljury/:
| File | What it is |
|---|---|
screenshot.png | full-page desktop render |
screenshot-390.png | mobile (390px) render |
critique.md | human-readable verdict with per-dimension scores |
fix-prompt.md | agent-ready instructions, hard-fails-first |
score.json | machine-readable result (for CI gating in 0.3) |
How it works
[1 RENDER] → [2 STATIC SIGNALS] → [3 VISION SCORE] → [4 COMPOSE]
Playwright deterministic BYO-key vision rubric formula
desktop+390 hard fails + tropes 6 weighted dims → score.json + critique + fix-prompt
- Render — headless Chromium loads your page; captures desktop + 390px screenshots and a DOM/CSS snapshot.
- Static signals — deterministic, no LLM: contrast, mobile overflow, touch targets, tiny body text, overused fonts, gradient panels, repeated cards, AI-copy badges. These produce the reproducible hard fails and half the deductions, for free.
- Vision score — your screenshots + the static summary + the rubric go to a vision model, which scores the six weighted dimensions. Bring your own key — no PixelJury backend, no auth wall, no cost to us.
- Compose —
composite − deductions, capped by any hard fail, per the rubric formula.
The score is not vibes: every point traces to a rule in rubric.md.
Providers
npx pixeljury review <url> --provider anthropic # ANTHROPIC_API_KEY
npx pixeljury review <url> --provider openai # OPENAI_API_KEY
npx pixeljury review <url> --provider gemini # GEMINI_API_KEY
npx pixeljury review <url> --provider ollama # local, no key (e.g. llama3.2-vision)
npx pixeljury review <url> --provider claude-code # your Claude Code login (no API key)
npx pixeljury review <url> --provider codex # your Codex / ChatGPT login (no API key)
npx pixeljury review <url> --provider mock # deterministic, no key, no network
For the API providers, the key is read from --key, then PIXELJURY_KEY, then the standard
env var. Override the model with --model. With no key found, PixelJury falls back to mock
so it always runs end-to-end (the static signals are still real).
Use your Claude Code or Codex subscription — no API key
If you already pay for Claude Code or Codex, PixelJury can score through the CLI you already
have logged in, instead of a separate metered key. It shells out to your local claude /
codex binary and uses its auth — PixelJury has no backend and never sees or stores your
credentials.
npx pixeljury review http://localhost:3000 --provider claude-code
npx pixeljury review http://localhost:3000 --provider codex
- Claude Code — works with a Pro/Max subscription (run
claudeonce to sign in, orclaude setup-tokenfor a long-lived token). Point at a different binary withPIXELJURY_CLAUDE_BIN. - Codex — works with a ChatGPT Plus/Pro/Business login (
codexonce to sign in). ⚠️ OpenAI recommends API-key auth (--provider openai) for automation and advises against subscription auth in public/CI contexts;--provider codexis best for individual use on your own machine and draws from your shared Codex quota.PIXELJURY_CODEX_BINoverrides the binary path.
Even better — let your agent run the loop. Drop
AGENTS.md(Codex/Cursor) or the/pixeljuryslash command (Claude Code) into your project, and your agent will review → fix → re-review until the score clears a threshold, on your existing subscription.
The rubric is open — argue with it
rubric.md is the scoring brain: six weighted dimensions (typography,
hierarchy, color, spacing, originality, polish), deterministic hard fails, and named
slop-trope deductions. Originality is the soul of it — if someone was told AI built this,
would they believe it instantly? — and it's the one thing a code linter structurally cannot
score.
Disagree with a weight? Open a PR. The argument is the marketing.
Who it's for
- Developers vibe-coding UIs with Cursor, Claude Code, v0, Lovable, Bolt, Replit, or Windsurf who want a second opinion before they ship.
- Indie hackers and solo founders shipping landing pages fast and tired of the "generic AI SaaS" look.
- Teams that want a repeatable design-quality bar in code review and CI — a number on every frontend PR instead of "looks fine to me."
- Anyone who has looked at an AI-generated page and thought this works, but it looks like AI made it.
Use cases
- Catch AI slop (gradient heroes, emoji icons, fake stats, templated cards) before launch.
- Accessibility smoke test — WCAG contrast, 44px touch targets, mobile overflow at 390px.
- Design QA in CI (0.3) — gate a PR under a score threshold.
- Before/after proof when your agent reworks a page — the score moves, and you can show it.
Install / requirements
- Node ≥ 18. PixelJury uses Playwright; on first run install the browser once:
npx playwright install chromium - An API key for one of the providers above (or use
--provider ollama/mock).
Roadmap
- 0.1 ✅
review— render → score → critique → fix-prompt. (you are here) - 0.2 —
fix,compare(before/after image), the viral demo unit. - 0.3 — GitHub Action +
--ciscore gating +reportHTML + more vision adapters. - 0.4 — custom rubric weights + a brand profile (score against your system, not a generic rubric).
FAQ
Is this a linter? No. Static linters read your code. PixelJury renders the page and judges the pixels — it can see that a hero looks generic, which no code rule can.
Does it need an API key? For the visual score, yes (OpenAI / Anthropic / Gemini), or run a
local model with Ollama. With no key it falls back to a mock provider — the deterministic
half (hard fails + slop tropes) still runs. There's no PixelJury backend and no telemetry.
Does it work with Cursor / Claude Code / v0 / Lovable / Bolt? Yes — PixelJury is agent- and
framework-agnostic. It judges whatever URL you give it and writes a fix-prompt.md any coding
agent can act on.
Will it work on my framework? It renders a URL in headless Chromium, so React, Vue, Svelte, Next.js, Astro, plain HTML — anything that serves a page — all work.
Is it free / open source? MIT licensed, free, no account required.
How is the score calculated? Six weighted dimensions scored by a vision model, minus
deterministic slop deductions, capped by any hard fail. Every point traces to a rule in
rubric.md.
Contributing
PRs and issues welcome — see CONTRIBUTING.md. The most valuable
contributions right now: arguing with the rubric (open a PR against rubric.md),
new vision adapters, and before/after examples for the gallery.
⭐ If PixelJury caught some slop for you, star the repo — it's how other people find it.
Development
npm install
npm test # deterministic engine + parser tests (no browser needed)
npx playwright install chromium
node packages/cli/bin/pixeljury.js review http://localhost:3000 --provider mock
Monorepo layout:
packages/cli # npx entry + the `review` command + terminal report
packages/core # Playwright render + static signals + score composer
packages/vision # BYO-key vision adapters + structured-output parsing
rubric.md # the scoring brain (ships in the repo)
examples/ # before/after gallery + a runnable sloppy-saas demo page
License
MIT © PixelJury contributors