菜单
Autonomous work mode — the user is going to sleep or will be unavailable, and cannot respond to questions or confirmations during the task. This skill is invoked explicitly via /sleep only. Do not auto-trigger based on conversational cues. Use this skill alongside whatever other skills the task requires.
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
Read third-party package source code via the `opensrc` CLI when investigating internal implementations of dependencies. ALWAYS invoke this skill before reading library/package source — including phrasings like "how does <pkg> work internally", "I want to read the source of <pkg>", "investigate the implementation of <pkg>", "let me check what <pkg> does under the hood", "git clone <repo> to read it", "look inside <pkg>", "<pkg> の実装を確認したい", "<pkg> のソースコードを調査", "<pkg> の中身を見たい", or any task that requires reading code from an npm / PyPI / crates.io package or a public GitHub/GitLab/Bitbucket repo. Use this skill instead of `git clone`, instead of asking the user to paste source, and instead of guessing from memory. Also use it when fetching only the path matters, e.g. setting up a search target for `rg` / `find` / `cat` across a dependency.
Enforce uncompromising implementation standards — no ad-hoc fixes, no shortcuts, no scope reduction. Use this skill whenever the user invokes /rigorous before giving an implementation task. This skill sets the mindset for the entire task: always pursue the architecturally correct solution regardless of change volume or complexity. Also use when the user says things like "ちゃんとやって", "しっかり実装して", "妥協しないで", "手を抜かないで", or any indication they want rigorous, no-compromise implementation.
Create a git commit for work done in the current session. Use this skill whenever the user invokes /commit or asks to commit their changes. This skill handles file selection (session-changed files, staged files, or explicitly specified files), commit message generation based on the repository's existing style, and outputs a summary of the commit. MUST be used for any commit operation — never commit without following this workflow.
Fetch the latest CLAUDE.md best practices from the official documentation. Use this skill before creating, updating, reviewing, or improving any CLAUDE.md file. Trigger on requests like "I want to write a CLAUDE.md", "review my CLAUDE.md", "what are the CLAUDE.md best practices?", or any work involving CLAUDE.md files.
Review the current session's work and evaluate whether any CLAUDE.md / AGENTS.md files in the project need updating. Only update when genuinely necessary. Use at the end of a session or at a natural breakpoint. Trigger when the user invokes /session-recap, or says things like "review docs", "do CLAUDE.md files need updating?", "reflect on what we did", "recap the session", or wants to check documentation consistency after completing work. This is not a mechanical codebase scan — it evaluates update needs based on the context of what was done during the session.
Scan the codebase to identify refactoring candidates — code that works correctly but would benefit from structural improvement. Use this skill when the user asks to find refactoring targets, code smells, cleanup opportunities, structural improvements, or wants to improve code maintainability. Also use when the user says things like "what should we clean up", "find messy code", "technical debt scan", or "what needs refactoring". This is NOT for bug hunting or security audits — it focuses purely on code structure, readability, and maintainability.
| name | sleep |
| description | Autonomous work mode — the user is going to sleep or will be unavailable, and cannot respond to questions or confirmations during the task. This skill is invoked explicitly via /sleep only. Do not auto-trigger based on conversational cues. Use this skill alongside whatever other skills the task requires. |
The user is about to be unavailable. They're trusting you to finish the job without them.
The user cannot answer questions, approve tool calls, or make decisions until they come back. Every confirmation prompt, clarifying question, or "should I proceed?" is a dead end — nobody is there to respond. Your job is to complete the task fully and autonomously.
Never ask for confirmation. Don't pause for approval. Don't ask "should I continue?" or "does this look right?" — just keep going.
Make your own judgment calls. When you hit a decision point (naming, architecture, implementation approach, which files to touch, how to handle an edge case), pick what you believe is the best option and move forward. You were chosen for this because your judgment is good — use it confidently.
Finish the entire task. Don't stop halfway with a status update and wait for feedback. Complete everything the user asked for, end to end.
If something blocks you, work around it. A tool permission gets denied? Try another approach. A test fails? Fix it. An ambiguity in the requirements? Make the most reasonable interpretation and go with it. Only stop if the situation is truly unrecoverable (e.g., missing credentials that you have no way to obtain).
Leave a clear summary when done. When the work is finished, write a concise summary of what you did, what decisions you made and why, and anything the user should review when they wake up. This is your handoff — make it useful.
Think of it like this: a colleague handed you a task at the end of the day and said "I'll check on it in the morning." You wouldn't email them at 2 AM asking which variable name they prefer. You'd make the call, do good work, and leave a note on their desk.
