跳至正文
OpenAgentSkill
浏览已开放 · 目录记录日期 2026-11-21 · 自动更新和账户服务暂时关闭。
浏览器与自动化

ai-plan

Work out what you're actually building by answering one question at a time, and turn every answer into a check — so you finish with a written standard for what "done and right" means, not just a plan. Trigger for "work out what we're building", "clarify what done means", "turn this into checks", "write the answer key". Not for deciding what the idea should be in the first place — use /ai-brainstorm. Not for running the build itself — use /ai-goal once the contract exists.

arcasilesgroup · 已有静态检查记录

安装与使用

运行命令前先阅读来源与依赖说明。复制命令不会自动安装或执行。

npx skills add arcasilesgroup/ai-engineering --skill ai-plan

技能说明


name: ai-plan description: >- Work out what you're actually building by answering one question at a time, and turn every answer into a check — so you finish with a written standard for what "done and right" means, not just a plan. Trigger for "work out what we're building", "clarify what done means", "turn this into checks", "write the answer key". Not for deciding what the idea should be in the first place — use /ai-brainstorm. Not for running the build itself — use /ai-goal once the contract exists. license: MIT

ai-plan

An idea has arrived, and the way from here to the finished thing isn't visible yet. This skill finds that way by naming where you're going, laying out the questions standing between here and there, and working them one at a time.

The output is not a plan for building. It's an answer key — a written standard for judging whether the finished thing came out right, including an honest list of what nobody has decided yet.

That distinction is the whole point. A plan says what to build. An answer key says how you'd know it came out wrong. If a reviewer — human or agent — is handed a plan and asked "is this good?", it has nothing to check against, so it invents a standard and approves whatever it sees. The answer key exists so nobody has to invent one.

Adapted from Matt Pocock's wayfinder skill (github.com/mattpocock/skills), stripped of the parts built for multi-week, multi-person efforts. The answer-key output is not part of his design and he has not endorsed it. His own guidance is to use a single conversation when the work fits in one — this deliberately runs the heavier interview on smaller work, because the thoroughness and the honesty about unknowns are what we're here for.

Plan, don't do

This skill decides. It does not build. Every question resolves into a decision plus a check, and the work is finished when nothing is left to decide before someone goes and builds the thing.

The pull to just start building is the signal you've reached the edge of the map and it's time to stop and hand off. This is absolute — there is no note, instruction, or exception that turns this into an execution skill. If the user wants it built, that's a separate session, after the answer key exists.

The file

Everything lives in one markdown file: .wayfinder/<slug>/MAP.md. No issue tracker, no ticket files, no dependency graph — one person, one sitting, one effort.

# MAP — <thing>

## Destination

<one or two lines: what reaching the end looks like>

## Open questions

<!-- ordered. This list is the running order — re-sort it when new questions arrive. -->

1. [ ] <question> — grilling
2. [ ] <question> — research
3. [ ] <question> — prototype

## Not yet specified

<!-- the fog: in-scope, but you can't phrase the question sharply yet -->

## Out of scope

<!-- ruled beyond the destination. Adding these makes the result worse. -->

## Answers

<!-- the detail. One section per answered question: the answer, the reasoning, and the check. -->

### <question>

**Answer:** <what was decided>

**Why:** <the reasoning — this is the primary source a reviewer reads when the summary isn't enough>

**Check:** <how you'd know if this came out wrong>
**Judged by:** run it | A/B pick
**Reference:** <a named, fetchable thing — or "—" when a check does the job>

Running order, not blocking

The Open questions list is ordered, and its order is the running order. Take the top one. When new questions arrive — and they will — put them where they belong in the list rather than appending them to the end. There are no blocking edges to wire; the list position says everything a dependency graph would.

The three question types

  • grilling — the default. A question that can be settled by talking it through with the user. Use commands/grill.md. The user answers; you never answer for them.

  • research — a fact outside this project is blocking a decision. You go and find out; the user isn't involved. Dispatch a background agent so the interview keeps moving, and point it at primary sources — the actual documentation, spec, or code — with a citation for every claim. Run these in parallel; they're the one question type that doesn't wait its turn.

    Two things follow. A found fact usually makes a run it check, because the fact is the check — that's most of why this type exists. And if the research doesn't settle it ("the docs don't say," "it depends how you configure it"), don't decide for the user: convert it into a grilling question and put it in the running order. If the honest finding is "you'd only know by running it," that's an Unknown.

  • prototype — "how should this look" or "how should this behave," which talking cannot settle. Use commands/prototype.md. Build something rough, react to it together. The rough thing then becomes the reference in the answer key — which is how you get a fetchable standard for something that has no famous example to point at.

Every answer produces a check

This is the one thing that makes the output an answer key instead of notes.

After the user settles a question, ask one more: "How would you know if this came out wrong?"

The answer to that is what goes in the bar. It takes one of exactly two forms:

Judged byUse it whenWhat the reviewer does
run itThere's a right answerRuns the thing and checks the outcome
A/B pickIt's a matter of taste or feelPuts it side by side with a named reference and picks one, blind

Never a score. Not "rate the checkout 1–10," not "assess whether it feels premium." Every line is binary — it passed or it didn't. If you find yourself wanting a third option, the check isn't specific enough yet; push it back into the conversation rather than softening the judgment.

A run it check has to name the observable outcome, not the topic. Whoever grades this will be a stranger in

阅读完整来源 ↗