audit

name	audit
description	Read-only analysis of global Claude Code configuration — flags issues and suggests improvements
disable-model-invocation	true
effort	high
allowed-tools	Read Glob Grep Bash(ls *) Bash(find *) Bash(which *) Bash(cat *) Bash(wc *) Bash(claude *) Bash(uname *) Bash(echo *) Bash(head *) Bash(command *)

Audit the user's global Claude Code configuration. ultrathink about what you find.

This is read-only. Do NOT create, edit, or delete any files.

---

Pre-loaded snapshot

These run inside the project sandbox — safe on any install.

Installed CLI tools

!which git gh node pnpm npm yarn bun python python3 uv pip cargo go ruby docker kubectl 2>/dev/null || true

OS & home

!uname -s !echo ~

Claude Code version

!claude --version 2>/dev/null || echo "(claude CLI not in PATH)"

---

Phase 1 — Read global config

Use your tools to read these paths (they may not all exist — that's expected):

1. ~/.claude/CLAUDE.md — personal instructions (also count lines with wc -l)
2. ~/.claude/settings.json — hooks and permissions
3. ~/.claude/rules/ — list rule files
4. ~/.claude/skills/ — find all SKILL.md files
5. ~/.claude.json — MCP server config
6. ~/.claude/projects/ — find MEMORY.md files (up to 10)

---

Analysis Checklist

Work through each check. Report findings as ✓ (good), ⚠ (suggestion), or ✗ (problem).

1. CLAUDE.md Health

• Does ~/.claude/CLAUDE.md exist? If not → ✗
• Is it under 200 lines? Over 200 → ⚠ split into .claude/rules/
• Does it contain project-specific commands that belong in a project CLAUDE.md? → ⚠
• Any stale references (tools not installed, paths that don't exist)? → ✗
• Does it use @import for modularity? Not required, but ⚠ if over 100 lines without it

2. Settings & Hooks

• Does ~/.claude/settings.json exist?
• Check for safety hooks. Flag if missing:
  • rm -rf blocking → ⚠
  • git push --force blocking → ⚠
  • git add . / git add -A blocking → ⚠

3. CLI Tool Safety

For each installed CLI tool that can modify external state, check for protective hooks:

• gh — can create PRs, close issues, delete repos
• docker — can remove containers/images
• kubectl — can modify cluster state Report unprotected high-risk tools as ⚠

4. MCP Servers

• List configured servers from ~/.claude.json
• If Node installed but no Playwright MCP → ⚠ (common miss)
• Flag any servers with overly broad permissions

5. Skills

• List global skills with their descriptions
• Check each for valid frontmatter (name, description)
• Flag overly permissive allowed-tools scopes

6. Rules

• List global rules
• Flag overlapping or contradictory rules

7. Memory Health

• Count project memory directories
• Flag any MEMORY.md over 200 lines or 25KB (truncation risk — whichever comes first)

---

Report Format

# CC Config Audit

## Summary
[1-2 sentences: overall health]

## ✗ Problems
- [issue + how to fix]

## ⚠ Suggestions
- [improvement + why it matters]

## ✓ Good
- [things well-configured]

## Next steps
1. [most impactful fix]
2. [second]
3. [third]

Be specific. Only report what you actually found — no generic advice.