Refactoring UI Audit

tl;dr: this is what the skills shelf is for. Not Anthropic’s sample pdf skill — a procedure packed from a book so the agent runs the same audit every time. Input must be something visual (URL, screenshot, frame, code). Output is a ranked report with principle names and concrete CSS-level fixes.

What it actually does

The SKILL.md we unpacked from refactoring-ui-audit.skill (25,359 bytes, 2026-09-09) tells the agent to walk eight Refactoring UI ideas in order, then cut 15–40 raw findings down to Critical / High-impact / Polish.

1. Hierarchy
2. Layout & spacing
3. Typography
4. Color
5. Depth
6. Images
7. Finishing touches
8. Personality & systems

Detail checklists live in references/01–08. First audit of a page: read all eight. Follow-up (“just typography”): read one file. The skill is explicit: don’t invent rules that aren’t in those references, and don’t audit from a verbal description — ask for a screenshot or URL.

Install

.skill is a zip. Unzip to a skills directory the host already watches:

# Claude Code — personal
unzip refactoring-ui-audit.skill -d ~/.claude/skills/refactoring-ui-audit

# project-scoped
unzip refactoring-ui-audit.skill -d .claude/skills/refactoring-ui-audit

claude.ai: zip upload under Settings → Features (paid + code execution). Custom skills there are per-user.

There is no npm package and no Anthropic marketplace id. If the host cannot load a local SKILL.md, this skill does not run.

Report contract (from the skill, not paraphrased marketing)

Each finding: short title, Where (element/selector), Principle (named Refactoring UI idea), Fix (values, not “improve hierarchy”). Suggested length 800–1500 words; past 2000 is padding. Critical bucket stays small on purpose.

Gotchas

• The book is not in the zip as a PDF. The skill encodes the method. Users who want the full book still buy Refactoring UI. Do not paste the reference files onto this site.
• Not a generator. It critiques an existing surface. Pair with something like frontend-design if you want net-new UI, not an audit.
• Page type matters. The skill says a dense admin and a marketing landing have different right answers. If the agent still SaaS-landings everything, the skill is not loaded.
• sandbox_test untested — we read the package, we did not run an audit in the sandbox.

Related: Claude skills (what the shelf is), UI Skills catalog, you don’t need animations.