claude-skills

A free audit toolkit for Claude Code. Fourteen slash commands covering accessibility, performance, SEO, privacy, design, deploy safety, and codebase audits.

What's a skill?

A skill is a reusable slash command for Claude Code. It's a markdown file that lives in ~/.claude/skills/<name>/SKILL.md. When you type /<name> in Claude Code, the skill runs the playbook in that file.

No code, no plugins, no SDK. If you can write a checklist, you can write a skill.

Start here

First time using skills? Try this order:

1. Install /onboard (instructions below)
2. Run cd <your-project> && claude to start Claude Code in your project
3. Type /onboard — Claude will read your codebase and brief you on it
4. Then install /a11y and run it on the same project — your first audit

If you like what you see, install the rest with the "all of them" command below.

Install

One skill

mkdir -p ~/.claude/skills/a11y && \
  curl -L https://raw.githubusercontent.com/Surfrrosa/claude-skills/main/a11y/SKILL.md \
  -o ~/.claude/skills/a11y/SKILL.md

Swap a11y for any skill name in this repo.

All of them (without overwriting existing skills)

git clone https://github.com/Surfrrosa/claude-skills.git ~/claude-skills-tmp && \
  cp -rn ~/claude-skills-tmp/*/ ~/.claude/skills/ && \
  rm -rf ~/claude-skills-tmp

The -n flag tells cp to skip any skill folders you already have. If you'd like to replace an existing skill with the version from this repo, delete that folder first.

After install, restart Claude Code (or run /help) so it picks up the new skills.

What's included

The toolkit (audit-class skills)

Skill	What it does
/full-sweep	Orchestrates the audit-class skills in a single Plan Mode run. Auto-commits the mechanical wins, files the rest as issues.
/drift	Finds every place code drifts from a canonical source of truth: hand-typed reference data, model IDs bypassing config, hardcoded URLs / emails / prices, magic numbers, year-bound constants, copy that fell out of sync with brand docs.
/cohesion	Finds every place a page drifts from the project's design system. Different button styles, hard-coded colors that should be tokens, parallel classes recreating canonical components, inline styles redefining global variables.
/coupling	Orthogonality audit. Asks the Pragmatic Programmer's question — "if I change feature X, how many modules light up?" — with mechanical evidence (fan-in / fan-out, circular imports, cross-layer violations, git co-change patterns).
/walkthrough	UX coverage audit. State-machine gaps, missing features (no logout, no password reset), entry-state combinations nobody walked through, pre-hydration render flashes.
/thatsweird	Browser and OS edge-case sweep. Chrome auto-dark inversion on dark sites, iOS rubber-band flash, iOS input zoom, prefers-reduced-motion violations.
/design	Design psychology audit. Asks "does the system serve the user?" not "does the page match the system?" Evaluates typography appropriateness, color register, brand-voice ↔ visual match.

Foundations

Skill	What it does
/a11y	Accessibility (WCAG) audit on source code and optionally live URLs. Can auto-fix safe categories.
/perf	Performance audit. Bundle size, image weight, render-blocking resources, and Core Web Vitals via PageSpeed.
/seo	SEO health audit across projects. Score them, identify gaps, optionally fix.
/privacy	Audit data collection, consent flows, exposed secrets, and privacy policy accuracy.

Workflow

Skill	What it does
/onboard	Digest a repository's architecture, docs, conventions, and current state before starting work.
/ship	Pre-deploy safety checklist. Catches build failures, leaked secrets, debug artifacts, missing env vars.
/session	End-of-session log generator. What changed, decisions made, next steps.

Which skills apply to your stack?

Skill	Web (Next.js / Astro / Flask)	Python backend / Node API	React Native / Expo
/a11y, /walkthrough	yes	skip	yes
/perf, /seo	yes	partial	skip
/privacy, /drift, /coupling, /onboard, /ship, /session	yes	yes	yes
/cohesion, /thatsweird, /design	yes	skip	skip

If a skill doesn't apply to your stack, it will self-detect and skip with a clear message — you don't have to memorize this.

Customization

Each skill resolves projects from a small registry table at the top of the file. The placeholder is myapp; add your own projects as you go.

For the audit-class skills (/drift, /cohesion, /coupling), there's a small one-time setup: you tell the skill what your project's canonical sources are (config module, design system file, etc.). The skill uses that as ground truth.

See CUSTOMIZATION.md for a walkthrough.

Sub-skill dependencies for /full-sweep

/full-sweep is an orchestrator. It expects sub-audits to be installed in ~/.claude/skills/. This repo includes most of them; a few (/zombie, /security-review, /vuln, /sentry) you'll need to source separately or substitute equivalents.

Used it on a real project?

We'd love to hear what you found. Open a field report — anonymous or named, your call. Even a one-line "ran /a11y on X, found 7 issues, 1 was a false positive" helps tune the skills and helps other users calibrate expectations.

What's useful in a field report:

• Which skill(s) you ran
• Stack / project size (rough)
• How many findings, and how many were false positives
• One surprise (something it caught that you didn't expect, OR something it missed)

You don't have to name the project. Anonymized reports are just as valuable.

License

MIT. Use them, fork them, share them. See LICENSE.

Issues + contributing

Something off? Open an issue. Improvements welcome — see CONTRIBUTING.md.