plugin87/ux-ui-agent-skills

The problem

Ask a coding agent for a dashboard and you get the statistical default: an indigo-to-purple gradient, four identical cards with no focal point, emoji as icons, grey-on-white body text, and a blue Delete Account button. Nothing in the typical prompt or repo tells the agent otherwise, and nothing catches it afterwards, because taste is not a build failure. This project ships a doctrine layer plus objective gates that measure the rendered page and fail CI on the specifics above.

What's different

Most approaches to this problem are either a component library (you get the library's components and its opinions, but the agent still writes the glue) or a prompt/persona file (cheap to adopt, unenforceable). This kit occupies a third position: it is a pure instruction and knowledge layer with no runtime, and its output is checked by gates that open a real browser. The claims that actually distinguish it: gates run against a real render — contrast in every state, target size, keyboard, reduced motion, overflow at 280px — rather than against source text; a token-by-intent check that rejects a destructive label filled with a saturated non-danger hue, and deliberately does not skip untokenised pages where intent gets chosen by convenience; and a slop/taste audit that flags single-radius, one-flat-shadow, hardcoded-value output. The 138 design systems and 52 component specs are content volume, and worth treating as such — breadth of reference material is easy to claim and hard to verify.

Who it's for

Front-end and design-systems engineers who already use a coding agent (Claude Code, Cursor, Codex, Copilot, Aider) and want its UI output to pass accessibility review without a human re-auditing every diff. Teams with a WCAG obligation or a public design system get the most value. Skip it if you are prototyping throwaway screens or already have a mature design system with its own enforced lint rules — you would be adding a second doctrine next to a working one. It is not a UI library; it does not give you components you import.

Cost of adoption

The README claims no build tools, dependencies or runtime, MIT licensed, installable via npm (ux-ui-agent-skills). The real cost is doctrinal: you adopt CLAUDE.md and .claude/rules/ into your project, which means the kit's opinions on spacing, radius, shadow and colour intent become review criteria. Expect some false positives against an existing palette before you tune. The gate suite runs a browser, so CI gets slower and heavier. Documentation is unusually detailed and self-critical about a gate that previously skipped untokenised pages, which is a good maturity signal — but also a reminder that a young rules engine still has edge cases. Version 2.10.0 indicates rapid churn; pin a version.

Verdict

This is a brand-new project — created 2026-03-16, under 300 stars a week ago and +896 stars on the weekly board at #3 — so the current attention is a spike, not yet an established user base. The design is coherent and the self-auditing framing (gates pinned by tests, a blind subagent eval) is more honest than most agent-skill repos. The risk is the usual one for instruction-layer projects: the rules are only as good as the next model's willingness to follow them, and the gate suite is the part that actually holds. If you care about accessibility on agent-written UI, it is worth a trial on one repo; do not roll it across a team before your own gate run comes back clean.