agents-md（特工 MD）

AGENTS.md Audit

AGENTS.md is the source of truth for agent instructions. Always write to AGENTS.md, never directly to CLAUDE.md. Symlink CLAUDE.md so Claude Code loads it:

ln -s AGENTS.md CLAUDE.md

AGENTS.md files are execution contracts, not knowledge bases.

Litmus test for every line: "Would removing this cause the agent to make a mistake?" If no, cut it. Bloated instruction files cause agents to ignore actual rules.

Reference Files

Quick Example

Input: AGENTS.md with stale commands and generic advice Quick audit result: 5/10 (Fail) Key issues: Missing test command, generic "follow best practices" advice, dead link to deleted folder Fix: Add npm test, replace generic advice with specific gotcha, remove dead link After: 9/10 (Pass)

How to Use

Default path:

• Start with quick audit using references/quick-checklist.md (10 checks)
• Escalate to full audit (references/quality-criteria.md) only when quick audit fails, file is high-risk, or user requests it
• Apply edits only after reporting findings and getting confirmation

Progressive loading:

• Always load the checklist for the selected audit mode
• Load references/refactor-workflow.md only for low-signal files (below target score, stale commands, or root file over ~150 lines)
• Load references/templates.md only when drafting a new file or rebuilding from scratch
• Load references/root-content-guidance.md only when deciding what stays in root vs moved out

Audit Workflow

Copy this checklist to track progress:

Audit Progress:
- [ ] Step 1: Discover files
- [ ] Step 2: Select audit mode (quick or full)
- [ ] Step 3: Run audit against checklist
- [ ] Step 4: Report findings with score table
- [ ] Step 5: Propose minimal diffs
- [ ] Step 6: Validate changes
- [ ] Step 7: Apply and verify

Step 1: Discover files

Run:

find . \( -name "AGENTS.md" -o -name "CLAUDE.md" -o -name "CLAUDE.local.md" \) 2>/dev/null | sort

Also check for a home-level file: ~/.claude/CLAUDE.md (applies to all sessions).

AGENTS.md is the source of truth. If a project has a CLAUDE.md that is not a symlink to AGENTS.md, recommend renaming it to AGENTS.md and creating the symlink:

mv CLAUDE.md AGENTS.md
ln -s AGENTS.md CLAUDE.md

Instruction files can exist at multiple levels — project root, parent directories, and child directories are all loaded automatically. CLAUDE.local.md is the gitignored personal variant. Audit each level independently.

For monorepos, include workspace-level AGENTS.md files.

Step 2: Select audit mode

• Quick audit: Default for most files (10 checks, target >= 8/10)
• Full audit: When quick audit fails, file is high-risk, or user requests full scoring

Step 3: Run audit

• Quick audit target: >= 8/10 checks from references/quick-checklist.md
• Full audit file-quality target: >= 91% of applicable points from references/quality-criteria.md
• Full audit execution target: 2/2 when producing an edit proposal
• Score each root file independently

Step 4: Report findings

Output a concise report before edits:

## AGENTS.md Audit Report

| File | Mode | Score | Grade | Key Issues |
|------|------|-------|-------|------------|
| ./AGENTS.md | Quick | 6/10 | Fail | Missing test command, stale path, doc-heavy section |

Step 5: Propose minimal diffs

• Fix broken/stale commands first
• Remove generic, duplicate, or obsolete guidance
• Move deep detail into linked files using @path/to/file.md import syntax
• Use emphasis ("IMPORTANT:", "YOU MUST") on critical rules that agents tend to skip
• Keep rewrites incremental and preserve useful wording when possible

Show each proposed change with rationale and a diff snippet.

Step 6: Validate changes

Validation loop:

1. Run smoke checks for core commands (dev, test, build, lint/typecheck) when applicable
2. If commands cannot be run, verify script existence and note the limitation
3. Check that linked paths resolve
4. Confirm no contradictory rules remain
5. If issues found → revise → validate again
6. Only proceed when validation passes

Step 7: Apply and verify

• Apply approved edits
• After each PR, add at most one new gotcha only if it prevented or fixed a real mistake
• Verify changes by re-running relevant commands

Gotchas

• Don't rewrite the entire file when targeted edits would pass audit — incremental fixes preserve useful wording and reduce review burden.
• Don't add gotchas that aren't grounded in a real failure — hypothetical warnings become noise the agent learns to ignore.
• Don't audit CLAUDE.local.md the same way as AGENTS.md — local files are personal and not committed, so enforce less strictly.
• Don't remove emphasis markers (IMPORTANT, YOU MUST) from rules agents consistently skip — these exist because the default phrasing was not enough.
• Don't treat a high quick-audit score as permission to skip validation — stale commands can hide behind passing checklists.