UX audit skill family

What's new 2026-09-03 — sixth skill joins; performance skill learns browser MCPs

• ux-audit-disclosure-depth 1.3.0 (new): scores every rendered value on six evidence-backed factors, returns INTERACT / CANDIDATE / SERVED / FLAT / OVER-DISCLOSED with the hidden question and the lowest depth rung, and proposes matching patterns from a bundled 40-pattern library with prerequisite-based fit.
• ux-audit-performance-interaction 1.6.1: at the start of a live audit it scans the session's own tools for a browser-control MCP and maps what it finds onto the protocol; every measurement now names its topology (headless/server = relative-only, the user's real browser = absolute).
• Pack at this release: psychology 2.1.0 · composition-layout 1.1.0 · compliance-design-system 1.1.0 · performance-interaction 1.6.1 · disclosure-depth 1.3.0 · sourcing-component 1.1.0.

About

UX audit skill family

Six Agent Skills that audit a product surface the way a careful reviewer would — from evidence in the code and the rendered UI, never from taste — and hand back one report in chat, mergeable across the family.

What they read, between them

• ux-audit-psychology — words and flows: clear to a new user, honest in its motivation design?
• ux-audit-composition-layout — geometry: does the arrangement show the structure before a word is read?
• ux-audit-compliance-design-system — tokens: does every rendered value come from the project's own system?
• ux-audit-performance-interaction — time: does the main thread stay responsive mid-gesture, against frame and input budgets?
• ux-audit-disclosure-depth — depth: which values are questions in disguise, and the cheapest way to answer them?
• ux-sourcing-component — when an audit needs a component: reuse, compose, build, or (last resort) install.

What a report looks like. One markdown document: scope and what was read, a census of everything audited, findings ranked by severity with Impact · Effort · Build tags, items routed to sibling skills, and a NEEDS-CONTEXT list naming exactly which evidence was missing. Scores become ranges when evidence is thin; nothing is invented.

When to reach for which. A screen is confusing → psychology. It "looks mixed" → composition. Hardcoded colours or spacing → compliance. It freezes or lags → performance. A detail page "feels flat" → disclosure depth. Any of them says "build this" → sourcing decides how.

The skills carry no project nouns, so they run on any product; your own budgets, tokens and rules live in your CLAUDE.md, not in the skills.