agent-init（Agent 初始化）

AGENTS.md Guide

Overview

Provide a repeatable workflow and quality bar for writing or updating AGENTS.md files so agents receive concrete, repo-specific instructions.

Trigger examples

• "Create an AGENTS.md for this repo and include exact test commands."
• "Replace our agent init command with a proper AGENTS.md file."
• "Audit our existing AGENTS.md for missing boundaries and commands."
• "Update AGENTS.md to reflect the current build and lint commands."
• "Create CLAUDE.md as a symlink to AGENTS.md for Claude Code."

Core Principles (condensed)

• Be concrete: prefer executable commands and code examples over prose.
• Put commands early and include flags/options.
• Use real code examples over explanations.
• Remove ambiguity: state boundaries and non-goals explicitly.
• Be specific about the tech stack; include versions and key dependencies.
• Define the agent's role/persona and how it fits the team.
• Cover the six core areas: purpose, code organization, run/test/validate, change management, dependencies/environment, and avoid list.
• Use three-tier boundaries: Always / Ask first / Never.
• Start small and iterate as the repo evolves.

Brainstorming Mode (divergent Q&A)

Use this mode when requirements are unclear or the repo structure will influence the document shape.

• Ask one question at a time. Prefer multiple choice when possible.
• Base questions on the actual repo shape (single app vs monorepo).
• Confirm which agents will consume the docs and their expectations.
• If Claude Code is in scope, plan a CLAUDE.md symlink to AGENTS.md.
• Ask for exact commands (with flags) and a real code example.
• Propose 1-2 AGENTS.md layouts (section order) with tradeoffs, then converge.
• Keep questions short and actionable, then confirm before drafting.
• Read references/agents-md-brainstorm.md for question banks and decision heuristics.

Workflow

1) Clarify scope and placement

• Confirm whether the task is to create new docs or revise existing ones.
• Default output is root AGENTS.md.
• If the user asks for a different location, confirm the exact path and proceed.
• If Claude Code is in scope, create CLAUDE.md as a symlink to AGENTS.md.
• Still maintain a single AGENTS.md (no per-agent files unless explicitly requested).
• Confirm which agents will consume the docs and align content with their workflows.
• If the repo uses a different location or filename, ask before proceeding.

2) Gather repo-specific facts

Collect only what will make the file actionable:

• Stack and versions (language/runtime/package manager).
• Key directories and ownership.
• Build, test, lint, and format commands.
• Environment variables, secrets, and local setup details.
• Git/PR expectations (branching, commit style, approvals).

2a) Shape the doc by repo structure

• If monorepo, add per-app or per-team sections inside AGENTS.md.
• If multiple roles are needed, use role-specific subsections inside AGENTS.md.
• If the repo is small, keep a single concise AGENTS.md.
• If a multi-file layout is requested, confirm the target paths before writing.

3) Draft with the six core areas

Write sections that map to the six core areas, keeping each section short and command-focused. Use the template if needed.

4) Add concrete commands and examples

• Prefer copy/paste commands over prose.
• Include example file paths and typical change patterns.
• Show expected outputs if ambiguity is likely.
• Include at least one real code example that matches repo style.

5) Add boundaries and safety

State what the agent must not do, what requires approval, and any data/credential constraints.

6) Review for clarity

• Eliminate vague statements ("run tests" -> specify the exact command).
• Ensure every requirement is tied to a command, path, or policy.
• Keep it short enough to scan quickly.

Output Requirements

Include, at minimum:

• Role/purpose of the agent and its scope.
• Repo map (key paths, entry points, ownership).
• Run/test/validate commands (copyable, placed early, include flags).
• Tech stack with versions and key dependencies.
• Change management rules (branching, commits, PR expectations).
• Dependencies and environment setup.
• Code style example (real snippet).
• Explicit avoid list and approval gates with Always / Ask first / Never.

Cross-agent compatibility

• Keep instructions in plain Markdown with explicit commands and paths.
• Avoid agent-specific fields or features unless the target agent requires them.
• Keep a single root AGENTS.md as the source of truth.
• If Claude Code is in scope, create CLAUDE.md as a symlink to AGENTS.md.

Multi-platform support

• Ask which agents are in scope (Codex, Claude, GitHub Copilot, Cursor, etc.).
• Use one shared AGENTS.md that works across those agents.
• Do not assume tool-specific directories; only update AGENTS.md unless the user asks otherwise.
• When showing examples, label tool-specific details as optional (not required).
• If Claude Code is in scope, add a CLAUDE.md symlink to AGENTS.md.
• If GitHub Copilot coding agent is in scope, consider adding .github/copilot-instructions.md and keep it minimal (ideally pointing to AGENTS.md as the source of truth).

References

• Read references/agents-md-template.md when drafting a new file.
• Read references/agents-md-example.md for a complete AGENTS.md example.
• Read references/agents-md-checklist.md when reviewing or auditing an existing file.
• Read references/agents-md-brainstorm.md when running a divergent Q&A gap analysis.