poteto
Entry router: match a task to one of seventeen playbooks (feature, bug-fix, refactoring, perf, research, ship, ...) and run its skill steps as a todo list. Use when the route is unclear.
- License 许可
- MIT
- 最近更新
- 2026-11-14
- Stars
- ★ 5
- Forks
- 0
01 前言 Front Matter
- name
- poteto
- version
- 1.0.0
- description
- Entry router: match a task to one of seventeen playbooks (feature, bug-fix, refactoring, perf, research, ship, ...) and run its skill steps as a todo list. Use when the route is unclear.
- argument-hint
- [task]
- compatibility
- Designed for Claude Code
02 资源文件 20
- 参考文档NOTICE.md3.3 KB
- 参考文档playbooks/ambition.md824 B
- 参考文档playbooks/autopilot.md5.6 KB
- 参考文档playbooks/babysit.md1.4 KB
- 参考文档playbooks/bug-fix.md630 B
- 参考文档playbooks/feature.md1.6 KB
- 参考文档playbooks/hillclimb.md3.8 KB
- 参考文档playbooks/investigation.md533 B
- 参考文档playbooks/orchestrate.md4.9 KB
- 参考文档playbooks/pause.md1.6 KB
- 参考文档playbooks/perf.md1.3 KB
- 参考文档playbooks/prototype.md1.0 KB
- 参考文档playbooks/refactoring.md1.2 KB
- 参考文档playbooks/research.md747 B
- 参考文档playbooks/retro.md353 B
- 参考文档playbooks/session-pickup.md623 B
- 参考文档playbooks/ship.md5.2 KB
- 参考文档playbooks/swarm.md1.4 KB
- 参考文档references/prior-art.md404 B
- 脚本scripts/bench.sh1.8 KB
03 SKILL.md 正文
Name the matched playbook first.
Ask the user only through the agent's question tool (Claude Code AskUserQuestion, Codex request_user_input, Pi ask_user_question, Gemini CLI ask_user, omp ask): ≤4 per call, recommended option first and labelled (Recommended); plain text only without one.
| Task | Playbook |
|---|---|
| Build or change behavior | playbooks/feature.md |
| Error, crash, failing test, regression | playbooks/bug-fix.md |
| Restructure code without changing behavior | playbooks/refactoring.md |
| Slow, latency, memory, throughput regression | playbooks/perf.md |
| Push one metric toward a target over many keep-or-revert attempts | playbooks/hillclimb.md |
| Decide a design by building throwaway variants | playbooks/prototype.md |
| Question, review, diagnosis, "how does X work"; nothing to change | playbooks/investigation.md |
| Web research, read sources, deep dive into a topic | playbooks/research.md |
| "Be ambitious", make a task deeper, "what would an expert do" | playbooks/ambition.md |
| Get PRs green, address review comments, "check on PR X" | playbooks/babysit.md |
| Land PRs, release, tag, publish, deploy, "is this worth a release" | playbooks/ship.md |
| Resume, "where was I", an active plan exists | playbooks/session-pickup.md |
| Pause, "stop here", save state to resume later | playbooks/pause.md |
| Learn from recent sessions, retrospective | playbooks/retro.md |
| One in-session wave of parallel sub-agents, swarm | playbooks/swarm.md |
| Multi-day program of many units or PRs, "orchestrate" | playbooks/orchestrate.md |
Carry a queue of PRs to verified (merge only in full), "autopilot" |
playbooks/autopilot.md |
Read the matched playbook. Open a todo list whose first items are its steps copied verbatim, in order. A step you choose not to do stays in the list as skip: <reason>; never drop it silently. No playbook fits → say so and ask which one applies.
Each step names a skill; that skill owns the procedure. Resolve it through the host's skill metadata. If a required skill is unavailable, report the missing dependency and stop.
Gated skill — retro — runs only from a playbook step here or when the user names it.
Defer to: the skill each playbook step names; git only on the owner's explicit request.
04 同仓库更多 Skills
read
Search the web and fetch URLs or PDFs as clean Markdown via a privacy-first cascade (local, gh, TinyFish/Firecrawl/Exa, opt-in proxy). Use to read, cite or save a URL. Not for repo files.
mono-write
Polish a research or learn draft (EN/VI) so it reads naturally, and write CHANGELOG entries and release notes from a commit list. Not for UI copy, marketing or tone changes.
librarian
GitHub code research via gh CLI, no cloning. Use to investigate external repos, find where a symbol is defined, or gather cited evidence and usage examples from GitHub.
git
Git operations with conventional commits. Use for staging, committing, pushing, PRs, merges. Auto-splits commits by type/scope. Security scans for secrets.
work
Execution orchestrator after `brainstorm` (and `to-plan` for high-risk). Runs phases wave-by-wave from locked artifacts, verifies each task, and hands off to `check`, `git`, or `handoff`.
wayfinder
Chart a Goal too fogged for one session as a decision map in the active plan, settle one decision per session, then hand it to brainstorm lock. Use when grilling shows the work spans sessions.