A design-system gatekeeper for AI coding agents
StyleSeed compiles your design decisions into rules your agent reads before every screen — then scores the rendered result and makes it revise.

What it does StyleSeed is an agent skill (one registration, 22 internal workflows) for Claude Code, Codex, and Cursor. You make the design decisions; it compiles them into rules the agent consults before building each screen, then runs a render → score → revise loop: screens scoring below 80 on its 0–100 gate get sent back before you see them. It ships 74 craft rules targeting the classic AI-UI tells — default indigo, emoji icons, the icon-in-a-chip above every card, same-weight cards hiding real priority.
The interesting bit The benchmark honesty is refreshing: on 120 rendered cells, raw rules alone were inconsistent (Codex +1.6, Claude Code −3.7), while the enforced gate improved both by +5.3. The authors explicitly frame it as “repeatable application of expert judgment, not expert-level output,” and even published a scorecard of their own README page scoring 58/100. That’s a project eating its own dog food and showing the receipt.
Key highlights
- 74 craft rules plus a scored gate with an enforced revise-below-80 loop — not just a prompt dump
- 8 output grammars (consumer, editorial, technical, commerce, etc.) so checks fit the result type instead of one universal taste
- Generative palette engine: any key color becomes OKLCH ramps, semantic roles, and contrast evidence
/styleseed referencecompiles screenshots, URLs, Figma, or an existing UI into an evidence-backed project grammar- Adapters beyond web UI: social carousels, decks, documents, graphics
Caveats
- The workflow must actually be invoked — a markdown file alone can’t block delivery, and 80 points is explicitly not expert approval
- The default install command follows mutable
main; the README points evaluators at a pinned v5.0.0 release archive instead - Heavy feature surface (recipes, skins, studio pipeline, adapters) — a lot to absorb for a tool whose core value is a rules-plus-gate loop
Verdict Worth a look if AI agents keep shipping the same indigo-gradient slop across your screens and sessions. If you already have a strict design system and disciplined humans, it’s mostly guardrails you don’t need.
Frequently asked
- What is bitjaru/styleseed?
- StyleSeed compiles your design decisions into rules your agent reads before every screen — then scores the rendered result and makes it revise.
- Is styleseed open source?
- Yes — bitjaru/styleseed is open source, released under the MIT license.
- What language is styleseed written in?
- bitjaru/styleseed is primarily written in JavaScript.
- How popular is styleseed?
- bitjaru/styleseed has 1k stars on GitHub.
- Where can I find styleseed?
- bitjaru/styleseed is on GitHub at https://github.com/bitjaru/styleseed.