菜单
Bootstrap or update AGENTS.md and CLAUDE.md for a repo — consolidate agent instructions into README.md and make the agent files delegate to it.
用 Codex 或 Claude 帮你安装 复制这段 Prompt，粘贴到 Codex、Claude 或其他助手里，让它检查 Skill 页面并帮你完成安装。
基于 SOC 职业分类
| name | init-agent-docs |
| description | Bootstrap or update AGENTS.md and CLAUDE.md for a repo — consolidate agent instructions into README.md and make the agent files delegate to it. |
You MUST act as a repo setup engineer whose job is to ensure agent instruction files (AGENTS.md, CLAUDE.md) exist, are minimal, and delegate to README.md as the single source of truth for agent context.
Check what exists: look for AGENTS.md, CLAUDE.md, and README.md in
the repo root.
If neither AGENTS.md nor CLAUDE.md exists:
README.md has an ## Agent Instructions section (create a stub if
README.md is absent).AGENTS.md and CLAUDE.md containing only the delegation line
(see format below).If one or both already exist:
README.md under an ## Agent Instructions
section. Preserve all unique information; deduplicate across files.AGENTS.md and CLAUDE.md with only the delegation line.Both files must contain exactly this (substituting the actual relative path):
See [README.md](./README.md) for repo context, commands, and agent instructions.
No frontmatter, no additional content, no headings.
Place it near the end of README.md, before any license section. Use this structure:
## Agent Instructions
### Build & test
<commands to build, lint, test>
### Conventions
<code style, naming, tooling constraints>
### Repo-specific notes
<anything else agents need to know>
Only include subsections that have actual content. Do not add placeholder text.
## Agent Instructions
section.## Agent Instructions section, append/merge into
it — do not duplicate the heading.Run `headroom perf` and act on its recommendations — flag long/unstable conversations, surface uncompressed stale reads, and publish eligible TOIN patterns
Critique React/TypeScript frontend code for correctness, security, performance, and idiomatic patterns
Execute a plan artifact's work orders by delegating each to Claude or Codex at the cheapest sufficient model tier, reviewing every result, and bouncing blocked items back to plan
Turn one scoped task or Linear issue into an implementation plan artifact of work orders, ready for `implement` to execute — no code written here
Decompose a vague goal into a prioritized, estimated roadmap and push it to Linear as epics/issues — product/principal-engineer altitude, no code
Generate atomic git commit messages following trunk-based development practices