agents-md-writer（特工 MD 作家）

AGENTS.md Writer

Edit AGENTS.md guidance to be short, specific, and executable. Prefer repository facts over generic advice. Default to editing existing instruction files; create new files only when no suitable file exists for the scope.

Workflow

1. Identify target tools and audit instruction sources.

• Detect which agentic tools must read the file (for example Codex, OpenCode, Claude Code, Cursor).
• Inspect repo-root and nested instruction files already in use (AGENTS.md and tool-specific alternatives).
• Read the matching tool docs for discovery/precedence behavior before editing.
• Read references/openai-codex-agents-md.md only when Codex behavior matters.

1. Prefer edit-in-place over scaffolding.

• Start from the closest existing instruction file and keep diffs minimal.
• Create a new file only when guidance is missing for a scope or precedence must change.

1. Collect project facts before writing.

• Read package manager scripts, CI workflows, linters/formatters, and test commands.
• Extract only commands and constraints that are true in the current repository.
• Add instructions only for recurring failures agents cannot resolve reliably from code discovery.

1. Draft the smallest useful change.

• Include only instructions an agent cannot infer quickly from code.
• Use imperative bullets with concrete commands and paths.
• Cover setup/build/test, code conventions, testing expectations, safety constraints, and change-management rules.
• Prefer positive direction over broad prohibitions when possible.
• If a behavior can be fixed in codebase structure/tooling, prefer that over adding instruction text.

1. Place overrides intentionally.

• Put shared defaults at repo root.
• Put subproject-specific guidance near the subtree it governs.
• Use AGENTS.override.md only when replacing same-directory AGENTS.md is intentional.
• Keep cross-tool guidance in common files; isolate tool-specific quirks to clearly labeled sections.

1. Verify and tighten.

• Remove duplicates from README or docs unless needed for execution.
• Remove vague guidance.
• Ensure each rule is observable, testable, or clearly enforceable.
• Validate behavior from repo root and a nested subdirectory with each target tool when possible.
• Remove architecture summaries and generated context that tools can infer from source in seconds.

Authoring Rules

• Prefer exact commands over abstract advice.
• Encode hard constraints: security boundaries, data policies, migration safeguards, approval gates.
• Keep architecture prose brief; link docs by path instead.
• Keep instructions concise to avoid truncation by byte limits.
• Resolve conflicts explicitly so nearest-file precedence is unambiguous.
• Keep guidance tool-agnostic unless a rule is truly tool-specific.
• Treat instruction text as a patch for repeated failures, not a full project explainer.
• Avoid "pink elephant" phrasing (for example broad "do not use X") unless required for safety/compliance.

Recommended Skeleton

# AGENTS.md

## Setup commands
- Install deps: `...`
- Start dev env: `...`

## Build and test
- Full checks: `...`
- Fast checks for touched package: `...`

## Code conventions
- Formatter/linter commands: `...`
- Framework or language constraints: `...`

## Change expectations
- Add or update tests for changed behavior.
- Update docs in `...` when public APIs change.

## Safety and approvals
- Do not ...
- Ask before ...

## Tool-specific notes (optional)
- Codex: ...
- OpenCode: ...

Debug Discovery

• Use each target tool's introspection or dry-run command to report active instruction files.
• Compare behavior from repo root and a nested subdirectory to confirm precedence.
• If source attribution is unclear, inspect tool logs/session traces when available.
• If instructions are missing, verify filename support, precedence rules, non-empty files, size limits, and restart requirements.

Resources

• references/openai-codex-agents-md.md: Codex-specific discovery order, override behavior, fallback filename config, limits, and verification commands.