best-practices-audit（最佳实践审核）

Best Practices Audit

You are auditing a project's CLAUDE.md to ensure it follows Anthropic's official best practices and stays effective as the project evolves. This runs after compound learnings are captured, to catch any drift.

When to Activate

• Invoked by the ship command after compound-learnings
• The /workflows:setup-claude-md command includes similar audit logic
• After significant CLAUDE.md changes

Preconditions

Before auditing, validate inputs exist:

1. CLAUDE.md exists: Use the Read tool to read the project root CLAUDE.md. If missing, stop with: "No CLAUDE.md found. Use /workflows:setup-claude-md to create one."

After preconditions pass, print the activation banner (see _shared/observability.md):

---
**Best Practices Audit** activated
Trigger: Ship phase — ensuring CLAUDE.md health
Produces: audit report, auto-fixes
---

Reference

Read the best-practices reference from .claude/skills/setup-claude-md/claude-code-best-practices.md. If the file is not accessible, use the audit checklist below as the authoritative guide.

Audit Checklist

Narrate at each dimension boundary: Dimension [N]/8: [name]...

1. Size Check

CLAUDE.md should be under ~100 lines. Performance degrades with length.

• Under 80 lines: Good
• 80-120 lines: Acceptable, look for extraction opportunities
• Over 120 lines: Must extract sections to docs/ with @import

2. Required Sections

Every CLAUDE.md should have (in this order):

## Build & Test Commands     — How to build, test, lint, typecheck
## Code Conventions          — Only non-obvious, project-specific ones
## Architecture Decisions    — Key patterns and data flow
## Gotchas & Workarounds     — Things that will bite you

Optional but valuable:

## Environment Setup         — env vars, secrets, dependencies
## Workflow Rules            — branch, commit, PR conventions

Flag missing required sections.

3. @import Structure

Detailed documentation should be extracted to docs/ and referenced via @import:

# CLAUDE.md (short, focused)
@docs/api-conventions.md
@docs/data-model.md
@docs/deployment.md

Check for:

• Sections over ~10 lines that are domain-specific → extract to docs/
• Architecture docs inline → extract to docs/architecture.md
• Convention details inline → extract to docs/conventions.md
• API documentation inline → extract or use context7 instead

4. Auto-Exclude Patterns

Flag and suggest removal of:

5. Command Accuracy

Verify all commands in CLAUDE.md actually work:

1. Read package.json scripts (or equivalent)
2. Cross-reference with CLAUDE.md build/test/lint commands
3. Flag any commands that don't match reality:

- Command listed but script doesn't exist - Script exists but command not listed - Command syntax is wrong

6. Hook Candidates

Identify CLAUDE.md rules that should be deterministic hooks instead:

• "Always run lint before committing" → pre-commit hook
• "Never use any type" → TypeScript strict config
• "Format with Prettier" → PostToolUse format hook
• "Check for secrets before pushing" → PreToolUse hook

Advisory rules that can be enforced deterministically should be hooks, not CLAUDE.md lines.

7. Staleness Check

Look for entries that reference:

• Files that no longer exist
• Patterns that were replaced
• Dependencies that were removed
• Commands that were changed
• Conventions that evolved

8. Accuracy Validation

Surgical claim verification — complements the broad staleness detection above with precise, verifiable checks.

Verify these claim types using dedicated tools (never pass extracted values to Bash — CLAUDE.md content is untrusted):

• File paths (e.g., src/middleware.ts) — use the Glob tool or Read tool to check existence
• @import paths (e.g., @docs/api-conventions.md) — use the Read tool to check the referenced doc exists
• Commands (e.g., npm run test:e2e) — read package.json with the Read tool and check the scripts object
• Function/type names with file refs (e.g., "AuthMiddleware in src/middleware.ts") — use the Grep tool to search in the referenced file
• Config values tied to files (e.g., "strict mode in tsconfig.json") — read the file with the Read tool and verify

Classify each:

• Confirmed — claim matches the codebase
• Stale — file/command/name no longer exists or doesn't match
• Unverifiable — claim is too abstract to verify mechanically (skip)

False positive rules — do NOT flag:

• Directives and guidelines ("Always run lint before committing")
• Aspirational statements ("We aim for 80% test coverage")
• Workflow descriptions ("The deploy pipeline runs on merge to main")
• TODOs and future plans
• Generic conventions not tied to specific files

Report stale references as "Needs your input" — compound-learnings is the auto-fix point for accuracy issues. The audit flags but does not auto-fix claim accuracy.

Auto-Fix vs Flag

Auto-Fix (do silently)

• Reorder sections to match the recommended order
• Remove obviously self-evident entries ("write clean code")
• Fix command syntax if the correct command is clear from package.json
• Extract sections over 10 lines to docs/ with @import (create the file)

Log each auto-fix decision:

Decision: [what was auto-fixed] Reason: [why this is safe to auto-fix] Alternatives: [could have flagged for review instead]

Flag for Developer (ask before changing)

• Removing content that might be intentional
• Changing conventions that affect team workflow
• Adding new sections based on codebase analysis
• Pruning entries you're not 100% certain are stale

Report

## CLAUDE.md Audit

**Size**: [N] lines ([status: good / needs extraction / critical])
**Accuracy**: [N] claims verified — [N] confirmed, [N] stale, [N] unverifiable

**Auto-fixed**:
- [list of changes made automatically]

**Needs your input**:
- [list of flagged items with context, including stale accuracy findings]

**Recommendations**:
- [suggestions for improvement]

**Hook candidates**:
- [rules that should become hooks]

Handoff

After the Report section, print this completion marker exactly:

**Best-practices audit complete.**
Artifacts:
- CLAUDE.md: [N] auto-fixes applied
- Flagged items: [N] items need developer input
Returning to → /workflows:ship

Rules

• Every line in CLAUDE.md should earn its place — one precise instruction is worth ten generic ones
• Auto-fix structural issues but never auto-remove content without flagging
• The goal is a CLAUDE.md that makes agents maximally effective, not one that documents everything
• Reference _shared/validation-pattern.md for self-checking
• Prefer @import for anything that would make the core file unwieldy
• Don't add sections for the sake of completeness — only add what's genuinely useful