AGENTS.md Setup and Audit

• IS: wiring a repo so every agent tool in use reads the same rules, then auditing, scoring, refactoring, and writing the AGENTS.md / CLAUDE.md / CLAUDE.local.md files agents load at session start.
• IS NOT: authoring SKILL.md files (use agent-skills-creator), project docs or READMEs (use ghostwriter), or mining session history (use the external cadence-advise skill where installed; this skill audits the file as-is).

AGENTS.md files are execution contracts, not knowledge bases. Two tests catch the two ways a line fails.

• Dead weight: "Would removing this cause the agent to make a mistake?" If no, cut it; bloat makes agents ignore the rules that matter.
• Harmful precision: "Is this wrong on any plausible task in this repo?" A prohibition that is wrong one task in ten is still obeyed on that task, and the agent cannot tell that this is the exception. State the outcome you want and let the surrounding code pick the path. NEVER write comments becomes match the comment density of the file you are editing: shorter, no exception list to maintain, and correct in a densely commented file without being told.

Absolutes still earn their place for safety, data loss, format contracts, and rules this repo's agents have actually been observed to break.

AGENTS.md is the tool-agnostic source of truth. Claude Code's built-in agents-md mod supports it directly. Use AGENTS.md at the root and in scoped subdirectories, without CLAUDE.md wrappers or symlinks. The default claude-md-or-agents-md mode yields when project Claude instruction files exist on the root-to-working-directory path; see references/project-setup.md for migration, settings, and loader limits. Preserve unique instructions before removing old files. User-level and managed Claude files are separate from repository migration.

Choose a Mode

• Repo has no agent instructions, or targets a tool it is not wired for -> Setup: references/project-setup.md decides which files exist and which tool reads each one, then Writing From Scratch below fills the root file, then Audit scores it.
• A file exists and the question is quality -> Audit, below.
• A file exists and is over the Size Budget below, stale, or scored badly -> Refactor, references/refactor-workflow.md.

Size Budget

Instruction files load into every session, so each line is paid on every task, and they grow by accretion: every change adds a line and none removes one. This section owns the numbers; other files point here.

Codex reads root first, so half its ceiling leaves room for the nested files between root and the launch directory, which a full root crowds out silently. Measure in every mode, before and after editing:

# Lines per instruction file
find . \( -name AGENTS.md -o -name AGENTS.override.md \) -not -path '*/node_modules/*' -exec wc -l {} +

# Bytes on one launch path: run once per directory agents start in (root, and each package)
d=<launch-dir relative to root>; while :; do f=$d/AGENTS.override.md; [ -f "$f" ] || f=$d/AGENTS.md; [ -f "$f" ] && echo "$f"; [ "$d" = . ] || [ "$d" = / ] && break; d=$(dirname "$d"); done | xargs wc -c

# Project skill descriptions, per tool (Claude Code reads .claude/skills, Codex reads .agents/skills)
for d in .claude/skills .agents/skills; do [ -d "$d" ] || continue; printf '%s: ' "$d"; awk 'FNR==1{f=0;d=0} /^---$/{f++;next} f!=1{next} /^[A-Za-z_-]+:/{d=/^description:/} d' "$d"/*/SKILL.md | wc -c; done

The last command is the other always-loaded cost: every project skill's description sits in every session of the tool that reads that folder. Report it per tool beside the file sizes, and treat a set of project skills that duplicates root, or whose descriptions would claim most prompts, as bloat on the same footing as a long root file.

Holding the budget:

• At or over budget, adding a line means removing or moving one. The commit or PR that adds it names the line it replaced, or where that line moved.
• When a rule changes, rewrite it where it stands. A dated change note, an "Update:" paragraph, or a "previously" note leaves both versions in the file, and agents follow whichever they read first. History lives in git. A date or version that limits when a rule applies ("on Node 22 and later") stays.
• Over a ceiling, Refactor is the top finding: apply it when the request allows edits, otherwise propose it as the first diff.

Reference Files

Writing From Scratch

The content step of Setup, and the whole job when the repo is already wired and only the file is missing. Skip the audit. Gather real commands from the manifest (package.json, Makefile, CI config), pick a skeleton from references/templates.md, fill it with verified commands and known gotchas, then validate against references/quick-checklist.md before delivering.

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 and score
- [ ] Step 4: Report findings with score table
- [ ] Step 5: Propose minimal diffs
- [ ] Step 6: Validate changes
- [ ] Step 7: Apply and report before/after scores

Step 1: Discover files

find . \( -name "AGENTS.md" -o -name "AGENTS.override.md" -o -name "CLAUDE.md" -o -name "CLAUDE.local.md" \) -not -path "*/node_modules/*" 2>/dev/null | sort
ls -la CLAUDE.md .claude/rules .cursor/rules 2>/dev/null

Also check ~/.claude/CLAUDE.md and ~/.codex/AGENTS.md; both load in every repo. ls -la CLAUDE.md tells you whether it is a symlink, an @AGENTS.md pointer, or a second copy, and a copy is a finding on its own. For monorepos, include workspace-level files. Audit each level independently: root holds universal rules, child files hold directory-specific rules (see what each tool loads in references/project-setup.md).

Step 2: Select audit mode

• Quick (default): 12 checks from references/quick-checklist.md, target >= 10/12.
• Full: 49 checks from references/quality-criteria.md, target >= 91% of applicable points (grade A). Use when the quick audit fails, the file gates a high-risk repo, or full scoring is requested.

Step 3: Run audit and score

Score each root file independently; exclude N/A checks from the denominator.

Step 4: Report findings

Output a concise report before any 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 |

Every issue in the table must map to a Step 5 diff; no vague findings.

Step 5: Propose minimal diffs

In priority order:

1. Fix broken or stale commands; bugs, not style.
2. Remove generic, duplicate, or obsolete guidance, restatements of what the harness already does, and facts auto-memory owns.
3. Diff each project skill (.claude/skills/, .agents/skills/, skills/) against the root file. Where both state the same fact or procedure, keep one copy (the skill for a procedure, root for what every task needs) and leave a one-line pointer in the other.
4. Rewrite blanket prohibitions as the outcome they were protecting; keep the absolute only where the harmful-precision test clears it.
5. Move detail needed in fewer than ~30% of tasks to a location that loads on demand: a nested AGENTS.md in the directory it concerns, a path-scoped .claude/rules/*.md, or a skill (not an @import; see Gotchas).
6. Add emphasis ("IMPORTANT:", "YOU MUST") only on critical rules agents skip, one line at a time.

For an audit request, propose the diffs. A request to improve, refactor, or write the file already authorizes those edits; apply them and report the rationale.

Step 6: Validate changes

1. Smoke-run core commands (dev, test, build, lint/typecheck) where the environment allows; otherwise verify the script exists in the manifest and note the limitation.
2. Check every linked and @imported path resolves. In a Claude Code session, /context lists the memory files that actually loaded; a file absent from that list is not loaded, whatever the tree looks like.
3. Confirm no contradictory rules remain across levels (home, root, child), against installed skills (.claude/skills/, .agents/skills/, ~/.claude/skills/), or against harness defaults. Where the overlap is deliberate, the file must say who wins, so the agent is told precedence instead of arbitrating it every task.
4. Issues found: revise, then validate again. Never proceed on "looks right".

Step 7: Apply and report

Apply approved edits, re-score with the same checklist, report before/after scores and the Size Budget measurements. Per future PR, add at most one new gotcha, only if it prevented or fixed a real mistake.

Make Drift a Gate

Step 6 proves the file once; the next rename breaks it again and nothing notices. When the repo has a check script, add one that fails when an instruction file (every AGENTS.md, any review rubric, every project SKILL.md) names something the repo no longer has:

1. Every <package-manager> run <script> names a script in the root manifest, or in the workspace a --filter names.
2. Every relative Markdown link resolves.
3. Every backtick path names a tracked or non-ignored file or directory, relative to the root or to the instruction file's own directory.
4. Every backtick camelCase or PascalCase identifier appears somewhere in source.

The identifier check proves existence, not visibility: a documented middleware made private or moved behind another API still passes. Say so in the script's header so nobody reads a pass as more than it is. If the repo keeps eval tasks with setup patches (references/repo-evals.md), the same gate checks each patch still applies to HEAD. Run it in the pre-push or CI check, not as advice.

Gotchas

• A project CLAUDE.md, .claude/CLAUDE.md, or CLAUDE.local.md can suppress the default AGENTS.md fallback across the project. Removing only the root wrapper may not fix loading. Check the built-in mod and /config Project instructions, then verify loaded files.
• @import moves text, not cost, and reaches Claude Code only. Imported files expand into context at launch (a 400-line file split into five imports still loads 400 lines), Codex and Cursor see the line as plain text, an import inside a code span or fence never loads, chains stop at four hops, and an external import from AGENTS.md needs an approval the mod cannot raise. Each failure is silent, so a rule every tool must obey stays inline.
• Nested files do not load the same way per tool. Claude Code loads a subdirectory's file when it reads files there; Codex concatenates only the files on the path from repo root to the launch directory. A rule that lives only in packages/api/AGENTS.md is invisible to a Codex session started at the root and to any Claude Code task that never opens that subtree. Universal rules go in root.
• Project-specific commands in ~/.claude/CLAUDE.md or ~/.codex/AGENTS.md load in every repo, so one project's pnpm dev becomes noise or a wrong command everywhere else.
• A committed AGENTS.override.md replaces AGENTS.md in that directory for Codex, so one checked in by accident silently swaps the rule set while the AGENTS.md everyone edits looks correct.
• Audit CLAUDE.local.md only for broken commands and contradictions with the shared file; it's gitignored personal config, and it exists only in the worktree that created it.
• Don't strip emphasis markers (IMPORTANT, YOU MUST) during a density cut; they exist because plain phrasing was already ignored once. When many lines carry them, none stands out, so the fix for a new skipped rule is emphasis on that one line, not another pass over the file.
• Content auto-memory owns (user preferences, personal feedback, evolving project status) collects in CLAUDE.md from the old #-hotkey habit. It loads every session, isn't repo knowledge, and drifts silently because nothing in the codebase contradicts it. Cut it to memory or CLAUDE.local.md.
• A passing quick score doesn't prove commands run; stale commands hide behind checklist passes. Step 6 smoke-runs aren't optional.
• Don't rewrite a whole file when targeted diffs would pass; full rewrites destroy battle-tested wording and inflate review burden.

Related Skills

• agent-skills-creator: authoring and improving SKILL.md files (different format and rules).
• External cadence-advise skill where installed: proposes AGENTS.md/CLAUDE.md edits from observed session history; complements this skill's file-first audit.
• ghostwriter: human-facing documentation; AGENTS.md content that belongs in docs should move there.
• codebase-architecture (Harden mode): the rest of the repo an agent works in. A rule a linter can enforce belongs there as an exit code, not here as prose, and it owns the docs tree this file indexes.
• Claude Code's /doctor checkup: proposes trims for a checked-in CLAUDE.md, cutting what Claude can derive from the codebase and migrating always-loaded procedures into skills and nested files. Complementary automated triage; it doesn't run the commands, so it never replaces Step 6.

Maintenance only: evals/evals.json contains regression scenarios for changes to this skill; it does not load during a user task.