Agents Md

AGENTS.md

Overview

Produce or harden a project AGENTS.md: concise, actionable, harness-neutral instructions that coding agents load as always-on project context. Optimize for agent success, not human onboarding (that stays in README.md).

This skill targets the open AGENTS.md format (plain Markdown, no required fields). Do not confuse it with GitHub Copilot custom-agent personas under .github/agents/ (frontmatter + persona files). Lessons about commands, examples, and boundaries still apply; personas do not.

Modes

Pick one:

• Create — no usable AGENTS.md (or user asked to scaffold).
• Review — audit an existing file against the quality bar; report gaps without rewriting unless asked.
• Improve — rewrite or edit so the file meets the quality bar.
• Nest — add or fix package/subdir AGENTS.md files in a monorepo.

If ambiguous, ask once: create, review, improve, or nest?

Core Rules

• Prefer one root AGENTS.md. Nest only when a subtree has different commands, boundaries, or conventions.
• Required: every root AGENTS.md must include the Communication (ASD-STE100) block from references/communication.md. Paste it as-is. Nested files skip it unless they override writing rules.
• Every other line must earn its keep. Litmus: Would removing this cause the agent to make a mistake it would not otherwise make? If no, delete it.
• Target ≤150 lines for a root file; 30–50 is enough for small repos. Split or index when larger.
• Put exact executable commands early (with flags/filters the project actually uses). Agents will re-run these.
• Prefer pointers and an orientation table over pasting architecture essays, full style guides, or README clones.
• Document what agents cannot infer: package manager, odd layouts, silent invariants, forbidden paths, security gotchas, and “ask first” changes.
• Use concrete examples (short snippets or canonical paths) instead of vague style prose.
• State boundaries in three tiers when useful: Always / Ask first / Never.
• Stay harness-neutral: no dependence on one product’s tool names, UI, or hidden behavior. Optional capabilities (plan mode, approvals, subagents) may be mentioned generically with a manual fallback.
• Never put secrets, tokens, private keys, or real credentials in AGENTS.md.
• Treat the file as living docs: when the user reports a repeated agent mistake, encode the fix here (or in a linked doc).
• Prefer the project’s real package manager and scripts from package.json / Makefile / justfile / CI — do not invent commands.
• Prefer bun / bunx in this skill’s examples only when the target repo already uses Bun; otherwise match the repo.

Quality Bar (must pass)

A great AGENTS.md covers these when they apply (omit empty sections; do not pad). Communication is not optional on the root file.

1. Stack — languages, frameworks, key libs with versions when non-obvious.
2. Commands — install, dev, test, lint/typecheck, build; scoped variants for monorepos.
3. Communication — the ASD-STE100 block from references/communication.md (required on root files).
4. Layout — where source, tests, configs, and generated output live (only if non-standard).
5. Conventions — project-specific style/architecture rules agents get wrong; link to lint/format otherwise.
6. Testing — how to run focused tests, what “done” means, whether to add tests unprompted.
7. Git / PR — commit/PR norms that differ from defaults (title format, required checks).
8. Boundaries — always / ask first / never (secrets, vendor dirs, prod config, schema, deps).
9. Index — links or a table to deeper docs (architecture, ADRs, security model) instead of inlining them.

Full checklist: references/quality-bar.md.

Workflow

1. Inspect the repo

Before writing:

• Read existing AGENTS.md, README.md, CONTRIBUTING.md, CI workflows, and package scripts.
• Note package manager, test runner, lint/format tools, monorepo layout, and unsafe areas.
• Skim for prior agent instruction files (CLAUDE.md, .cursor/rules, GEMINI.md, etc.) to migrate unique content — do not blindly duplicate.

2. Draft or edit

• Use references/template.md as a starting shape; delete unused sections.
• Always paste the Communication block from references/communication.md into the root file.
• Put high-frequency commands near the top. Put Communication next.
• Encode only project-specific constraints; strip generic advice agents already know. Do not strip Communication.
• For monorepos, follow references/nesting.md.

3. Validate

Check against:

• Quality bar above and references/quality-bar.md.
• Communication block present and unmodified: references/communication.md.
• Anti-patterns in references/anti-patterns.md.
• Good/bad patterns in references/examples.md.

Confirm every linked path exists. Confirm commands match scripts/CI. Confirm no secrets.

4. Deliver

• Create / Improve / Nest: write the file(s) at the agreed path(s), usually repo-root AGENTS.md.
• Review: return a short report — passes, failures, suggested edits — without rewriting unless asked.
• End with a brief note: line count, sections present, and any intentional omissions.

Conflict and Precedence Notes

Document these for maintainers when relevant (do not lecture inside every file):

• Closest nested AGENTS.md to the edited files wins over parent guidance when agents resolve conflicts that way; user chat always overrides files.
• Some agents concatenate root → cwd (later/more specific wins). Keep nested files additive and explicit about overrides.
• Soft size caps exist in some tools (for example Codex’s project-doc byte limit). Prefer splitting/nesting over one huge root file.
• For Claude Code compatibility when needed, a thin CLAUDE.md that includes @AGENTS.md (plus any Claude-only extras) avoids duplicating the whole guide.

Out of Scope

• Replacing path-scoped IDE rule systems (for example .cursor/rules with globs) when the user wants fine-grained, file-pattern rules — AGENTS.md is the simple, portable default.
• Authoring Copilot custom-agent persona packs under .github/agents/ unless the user explicitly asks for those.
• Copying the entire human README into AGENTS.md.