| name | agents-md-steward |
| description | Maintain AGENTS.md, CLAUDE.md, CODEX.md, and similar repo instruction networks. Use when asked to create, edit, audit, prune, harmonize, compress, or route agent-facing guidance; when guidance files feel bloated, narrative, stale, duplicated, or too generic; when deciding whether a rule belongs in root instructions, nested owner instructions, a skill, a spec, an issue, ordinary docs, or a memory store; or when the user asks to scrutinize opportunities for new instruction files without defaulting to adding more files. |
AGENTS.md Steward
Goal
Keep agent instruction networks concise, current, scoped, and behavior-changing.
Prefer pruning, routing, and clearer ownership over adding another rule.
Boundary
Do not use this skill for ordinary code edits, test fixes, reviews, or docs
work unless agent-facing instruction files are being created, edited, audited,
or routed.
First Move
- Inspect the active instruction surfaces before editing:
- root
AGENTS.md, CLAUDE.md, CODEX.md, or equivalent;
- nested files that apply to the changed paths;
- existing skills or scripts that may already own the workflow.
- Name the intended behavior change in one sentence.
- Decide the smallest durable home:
- root instructions for repo-wide invariants;
- nested instructions for path-local ownership, boundaries, commands, or gotchas;
- a skill for reusable workflows or conditional procedures;
- specs/issues/docs for temporary planning, narrative, history, or context.
Edit Rules
- Use exact owner, path, command, and boundary names. Replace generic quality
words like "thorough", "detailed", "clean", "robust", or "easy to read" with
observable checks, or delete the sentence.
- Keep only instructions that change agent behavior. Remove generic advice that
competent coding agents already know.
- Prefer attractors over long prohibition lists. Use repellors only for common,
costly failure modes.
- Remove narration, provenance, experiment history, old roadmap language,
session recap, and model/tool anecdotes from instruction files.
- Keep root instructions short enough to scan. Move local detail down to nested
instruction files or into skills.
- Do not duplicate the same rule across root, nested files, and skills. Keep the
broad invariant at root and the operational procedure in the owner surface.
- Do not promote memory or learnings entries into root or nested instruction
files automatically. Each promotion must pass classification and earn its
always-loaded cost.
- Do not create a new nested instruction file unless a path has distinct rules
that future agents cannot reliably infer from code and existing guidance.
- When a refactor, test, command, or owner-local doc would remove the need for a
rule, prefer that over growing the instruction network.
Classification
Classify proposed guidance before writing:
| Kind | Home |
|---|
| Repo-wide invariant | Root instruction file |
| Path ownership, public surface, commands, local gotcha | Nearest nested instruction file |
| Repeatable workflow, decision routing, external reference procedure | Skill |
| Temporary plan, roadmap, implementation state | Spec or issue |
| Historical explanation or rationale | Ordinary docs or decision record, only if durable |
| Durable personal learning or recurring correction, not repo-invariant | Memory store, only with explicit user authorization and dedupe against existing entries |
| Generic advice, recap, agent shorthand, stale process | Delete |
New File Gate
Suggest a new AGENTS.md only when at least one condition holds:
- repeated wrong-path edits or imports show the current network does not route
agents correctly;
- a subtree has a distinct public surface, command menu, or safety invariant;
- local behavior is non-obvious and expensive to rediscover;
- the rule would be too noisy at root and too important to leave implicit.
If none hold, improve the existing owner file, skill, test, or code boundary.
Validation
For guidance-only edits, run:
git diff --check
Also run targeted stale-language searches for:
- removed filenames, specs, commands, or modules;
- generic no-op quality words such as "thorough", "detailed", "clean",
"robust", "easy to read", and "best practices";
- agent shorthand, model names, experiment/provenance terms, and old roadmap
labels;
- duplicate copies of newly routed rules.
Report what moved, what was deleted, and why the remaining guidance earns its
space.
References
When shaping or auditing an instruction network, apply the shared principles in
references/source-principles.md and these primary sources:
Use those sources to keep guidance tactical, project-specific, and free of
generic advice, historical narrative, and behavior attractors that do not match
repo intent.