| name | agents-md |
| description | Use before creating or editing any `AGENTS.md`. Defines what belongs there, how to keep it root-scoped, and how to keep agent guidance project-specific, actionable, and terse.
|
Use before any create/edit of AGENTS.md.
Mental Model
Write for an experienced principal engineer: setup done, real work starting.
AGENTS.md is an orientation layer, not a full map. It should help them choose the right starting point for the work in front of them, avoid repo-specific mistakes, and know how to verify the result. If a line would not change how they scope, execute, debug, review, document, or validate the work, cut it.
AGENTS.md is user-scoped operational memory, not agent-authored completeness. Start with the smallest useful instruction set. Expand only when the user asks or when a missing rule would clearly cause repeated repo-specific mistakes.
Hierarchy-aware scope
- Root/global
AGENTS.md: repo-wide invariants, routing cues, and cross-cutting verification only.
- Subtree
AGENTS.md: local conventions, workflow details, validation, and gotchas for that subtree.
- If a closer
AGENTS.md already owns a topic, do not duplicate it at root unless the user explicitly asks to promote it.
- For root-level references to non-standard directories, default to a directory map:
path — purpose.
Modes
- Patch: if asked to record or fix one thing, change only that.
- Bootstrap: if
AGENTS.md does not exist, create the smallest useful file for the request.
- Comprehensive: only do a broad/default pass when the user explicitly asks for a standard, full, or comprehensive file.
Coverage Areas
Cover only what the requested scope needs. For comprehensive passes, make sure the file answers:
- Where to start: source-of-truth paths, relevant subtrees, task routing cues.
- What can go wrong: forbidden actions, risky commands, confirmations, stop conditions.
- What "done" means: mandatory tests, checks, review steps.
- How this repo differs from defaults: ownership boundaries, invariants, non-obvious conventions, workflow traps.
- Where to go deeper: short conditional pointers for less-common domains.
Writing Rules
- Write instructions, not explanations.
- Default to sparse: prefer short bullets over prose; include only what changes where the agent starts, what it avoids, and how it knows it is done.
- When the user asks for "tight", "brief", "sparse", or "just reference it", default to
path — purpose bullets or one-clause instructions.
- Add extra clauses, examples, or commands only when they change execution, scope, or verification.
- Prefer imperative bullets, exact paths, concrete commands, explicit decision rules.
- Prefer task-based routing (
if touching X, read Y) over sequential reading lists.
- Match the file's compression level when it helps readability.
- It is fine, and often preferable, to add sparser/terser guidance to a denser file when that better fits the requested scope and improves scanability.
- Cut generic engineering advice, framework basics, temporary notes, duplicates.
- State defaults, ask points, and no-touch zones.
- Encode explicit user instructions and stable repo facts; do not turn one session into broad policy.
- Do not scan the repo broadly, run tests, or fill gaps just to make a new file feel complete.
- Remove stale or conflicting guidance in the same edit.
Quality Check
Before finishing, verify a strong new engineer could answer the questions relevant to the requested scope:
- For the work I am doing, where do I start?
- What must I run before I say "done"?
- What would cause damage or review comments?
- What repo-specific rule would I otherwise miss?
If a relevant answer is missing, add it or ask. Do not broaden file scope just to satisfy this checklist.
If you changed AGENTS.md, summarize what changed and why this scope fits the orientation-layer mental model.
