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, emoji feature cards, 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, tells you specifically what's making it look cheap, then writes a ready-to-paste fix for your agent and gives you a design score so you can watch it get better. The example below went from 16 → 84 in a single pass. ✨

npx pixeljury review http://localhost:3000

No design degree, no Figma, no API key required.

See it work — before / after

Run it, hand the generated fix-prompt.md to your agent, run it again, watch the score climb. That loop is the product.

What you get

Every run writes to ./pixeljury/:

| File | What it is | |---|---| | screenshot.png / screenshot-390.png | desktop + mobile renders | | critique.md | human-readable verdict with per-dimension scores | | fix-prompt.md | agent-ready instructions, hard-fails first | | score.json | machine-readable result |

Providers — bring a key, or use a subscription you already have

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
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

With no key, PixelJury falls back to mock so it always runs end-to-end — the deterministic checks (contrast, mobile overflow, touch targets, slop tropes) are still real.

How it works

RENDER (Playwright) → STATIC SIGNALS (deterministic) → VISION SCORE (BYO key) → COMPOSE

Six weighted design dimensions (typography, hierarchy, color, spacing, originality, polish) scored by a vision model, minus deterministic slop deductions, capped by any hard fail. Every point traces to an open, versioned rubric.

Who it's for

Developers vibe-coding UIs with Cursor / Claude Code / v0 / Lovable / Bolt who want a second opinion before they ship — and teams that want a repeatable design-quality bar in code review and CI instead of "looks fine to me."

Full docs, the rubric, and how to contribute: https://github.com/gchahal1982/pixeljury · MIT