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。
设计与 UI

intent-discovery

Use when the user wants something new built or existing behaviour changed, such as a feature, component, page, API, module or project, and no approved design for it exists yet in this conversation or in docs/plans

NoobyGains · 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 NoobyGains/godmode --skill intent-discovery

技能说明


name: intent-discovery description: "Use when the user wants something new built or existing behaviour changed, such as a feature, component, page, API, module or project, and no approved design for it exists yet in this conversation or in docs/plans"

Intent Discovery

Turn a request into a design the user has approved, with a specification precise enough that a capable builder could plan from it without asking again. You build what you believe the user wants, and that belief is usually thinner than it feels. Find out what they want and why, one question at a time, and get sign-off on a written design before anything is built.

The Prime Directive

NOTHING GETS BUILT UNTIL THE USER APPROVES THE DESIGN

No exceptions. No workarounds. No shortcuts.

Building includes writing product code, scaffolding, installing dependencies, creating projects and invoking any implementation skill. Reading and exploring are fine.

When to Use

Required:

  • A new feature, component, page, endpoint, module or project.
  • A change to how existing behaviour works.
  • A redesign or re-architecture.
  • The user picked option A or B in godmode:rationale.

Skip:

  • A typo, a dictated config value, a rename the user specified exactly.
  • A command the user gave verbatim.
  • A bug: use godmode:fault-diagnosis.
  • The user says to skip the questions. Then state any significant choice you are making (platform, language, storage) in one plain line and proceed through the normal pipeline.

The Protocol

Create one task per step and track them.

  1. Survey the ground. Read the relevant files, docs and recent commits. Note what exists, what conventions hold, and what the user has already decided. Evidence: a short list of what you found.

  2. Size the work. Say the size in plain words; the user may override it.

    SizeDesign lives inNext
    Quick, clear taskthe request itselfdo it directly
    Small changea design summary in chatgodmode:task-planning or direct work
    Projectdocs/plans/YYYY-MM-DD-<topic>-design.mdgodmode:task-planning

    If the request is really several independent projects, say so, agree an order, and run this skill on one at a time.

  3. Ask, one question per message, starting from what you already know. When the user has said what they want (a detailed request, a spec, a plan), open by echoing it back in your own words in about 200 words: what they want, who it is for and why, plus what the survey found that bears on it (how the relevant parts work today, what the change touches, any constraint, conflict or cheaper path they may not have seen). Invite corrections, then ask the one question whose answer would most change what gets built. When they have not said much, ask one open question in their own terms about the biggest thing the request leaves unsettled, and say what the answer would change. Build each later question on their last answer or on something you found. Cover purpose, users, success, constraints, non-goals and what they already have, but never ask what the project or the conversation already answers. Offer a short menu only when the user is stuck; a guess they can correct often beats a question. Stop asking when another answer would not change the design. If this skill started on its own rather than at the user's request, end the first message with a one-line offer to just get going instead; if they take it, state any significant choice you are making in one line and proceed.

  4. Look for prior art. Invoke godmode:reference-first for proven implementations and godmode:codebase-conventions for in-repo precedent. Invoke godmode:grounding for every library, API, version or platform rule the design will rest on. Evidence: the sources you will build on, with links.

  5. Propose 2-3 approaches. Label them A, B, C, lead with your recommendation, and give each its trade-offs in two or three lines. Cite the references that inspired each. The user chooses; never auto-pick.

  6. Play the design back in sections. Describe each part (what it is for, how it behaves, what it touches) in up to 200-300 words, mark your own guesses as guesses, and ask what is wrong or missing before the next section. Cover architecture, components, data flow, error handling and testing.

  7. Write the specification section (template below), scaled to the work.

  8. Write the design document to docs/plans/YYYY-MM-DD-<topic>-design.md and commit it. For a small change, the chat summary is the design.

  9. Check it as the builder would (projects only; a small change goes straight to the user). Tell the user a reviewer reads the document first and you may come back once with questions. Dispatch a fresh subagent with the builder-check prompt and the document path, or read the document in that role yourself when dispatch is unavailable. Sort what comes back into three piles: answered already by the conversation or the project (write the answer into the document); minor, where a sensible builder could settle it without changing what gets built (settle it and mark it as your call); theirs, where only the user can answer and the answer changes what gets built (bring these in one message, most important first). Update the document with their answers and hand it over with the list of calls you made. That is the only round of questions the check produces.

  10. Get approval on the full document. Approval covers what the user saw; a section you have not shown is not approved.

  11. Hand off to godmode:task-planning. That is the only next skill.

digraph intent {
    "Survey" -> "Size";
    "Size" -> "Do it directly" [label="quick task"];
    "Size" -> "Questions, one at a time" [label="small or project"];
    "Questions, one at a time" -> "Prior art + grounding";
    "Prior art + grounding" -> "2-3 approaches";
    "2-3 approaches" -> "Sectioned playback";
 
阅读完整来源 ↗