excatt Audit Skill
Project-specific verification rules that generic linters cannot enforce.
| Skill | Checks |
|---|---|
/verify |
Build, types, lint, test, security |
/audit |
Business logic, architecture patterns, naming conventions |
/gap-analysis |
Design doc vs implementation matchRate |
When to Use
Triggers: audit, check rules, run audit, audit manage Do NOT use for: build/lint/test (/verify), design vs implementation (/gap-analysis)
Mode Selection
/audit → rules exist? → RUN mode (execute rules)
→ no rules? → MANAGE mode (bootstrap)
/audit manage → MANAGE mode (create/delete/sync rules)
Rules location: .claude/audit-rules/*.md
RUN Mode
Step 1: Collect & Match
Read all .claude/audit-rules/*.md. Match each rule's scope against changed files:
BASE=$(git remote show origin 2>/dev/null | grep 'HEAD branch' | awk '{print $NF}')
BASE=${BASE:-main}
git diff "$BASE"...HEAD --name-only
Skip rules whose scope has no matching changes.
Auto-exempt files (always skip, no rule needed):
- Lock files:
pnpm-lock.yaml,uv.lock,package-lock.json,Cargo.lock - Generated/build output:
dist/,build/,.next/,__pycache__/ - Docs:
README.md,CHANGELOG.md,LICENSE - Test fixtures:
fixtures/,__fixtures__/,test-data/ - Vendor/third-party:
vendor/,node_modules/ - CI/CD config:
.github/,.gitlab-ci.yml,Dockerfile - CLAUDE.md and
.claude/internals
Step 2: Execute
For each matched rule:
- Run bash commands from
## Checksection (if any) - AI-analyze results against
## Pass Criteria - Classify: PASS / FAIL / WARN / SKIP
Step 3: Report
AUDIT REPORT
=============
api-response-format PASS
error-handling FAIL 2 violations
src/api/users.ts:45 bare catch without error type
src/api/posts.ts:78 empty catch block
naming-convention PASS
---------------------------------------------
Rules: 3 | Passed: 2 | Failed: 1
Status: NEEDS FIX
Status: all pass → CLEAR / any error-severity fail → NEEDS FIX / warning-only → REVIEW
Step 4: Fix (if issues found)
Ask user via AskUserQuestion:
- Apply All Fixes — Automatically apply all recommended fixes
- Review Each Fix — Approve each fix individually
- Skip Fixes — Report only, make no changes
After fixes, re-run only failed rules and show before/after comparison.
MANAGE Mode
Bootstrap (No Rules)
Detect project type from config files and propose seed rules:
| Signal | Type | Suggested Rules |
|---|---|---|
next in package.json |
Next.js | Server/Client separation, API response format |
fastapi in pyproject.toml |
FastAPI | HTTPException usage, router naming |
express/nest in package.json |
Node API | Error middleware, response wrapper |
| Any project | Universal | Naming convention |
Change Analysis
git diff "$BASE"...HEAD --name-only --stat
Compare changed domains against existing rule scopes → propose CREATE/UPDATE/DELETE.
User Confirmation
Always confirm before creating or deleting rules:
Proposed:
+ api-response-format (new, covers src/api/**)
- legacy-auth-check (stale, files removed)
Apply? [Y/n]
Rule Format
Each rule is a markdown file with YAML frontmatter (name, scope, severity) and structured sections (What, Check, Pass Criteria, Examples).
Full guide: references/rule-format.md
Example rules: references/examples/
Integration
Pre-commit flow: /verify → /audit → commit
Review loop: Implement → Spec Review → Quality Review → /verify → /audit → Complete