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。
开发与测试

clarify

Turn a vague request, voice dump, rough plan, or half-written prompt into a self-contained prompt that a fresh agent can execute. Use when asked to clarify, refine, improve, tighten, or stress-test an agent task. Resolve safe defaults directly. Ask one question at a time only when the answer would materially change the work.

owainlewis · 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 owainlewis/agent-skills --skill clarify

技能说明


name: clarify description: "Turn a vague request, voice dump, rough plan, or half-written prompt into a self-contained prompt that a fresh agent can execute. Use when asked to clarify, refine, improve, tighten, or stress-test an agent task. Resolve safe defaults directly. Ask one question at a time only when the answer would materially change the work."

Clarify

Turn a rough request into a prompt another agent can execute.

Choose a mode

  • Refine directly when the goal is clear and missing details have safe defaults.
  • Interview when one unresolved choice would materially change behaviour, data, security, compatibility, cost, output, or proof.
  • Execute the clarified prompt only when the user explicitly asks to run it or says "just do it."

Process

  1. Read the request and any referenced material.
  2. For repository work, inspect discoverable context such as AGENTS.md, CLAUDE.md, README.md, the file tree, and relevant code before asking questions.
  3. Extract the goal, current state, inputs, outputs, constraints, scope boundaries, failure behaviour, and success checks.
  4. Resolve unambiguous references and project conventions from the available context.
  5. Use a sensible assumption when the user can change it later without reworking the task.
  6. Ask one question when no safe assumption exists. After the answer, check whether another important decision remains.
  7. Include all context needed to execute the prompt. Do not refer to "our discussion" or information outside the prompt.

Ask useful questions

Ask about decisions such as:

  • which input, file, directory, system, or audience is in scope
  • what output must exist, where it belongs, and whether it may overwrite anything
  • required behaviour for bad input, missing data, network failure, or partial success
  • tradeoffs that change product behaviour, security, compatibility, cost, or architecture
  • the observable result or check that proves completion

Do not ask about cosmetic choices, conventions the repository answers, details already supplied, or preferences with an obvious low-risk default.

Use this shape:

Question: <one decision>
Recommendation: <answer> because <short reason>.

Write the final prompt

Include only what changes execution:

  • goal
  • relevant context and current state
  • exact inputs and outputs
  • in-scope and out-of-scope work
  • constraints and safe assumptions
  • required failure behaviour
  • success criteria and verification

Prefer positive instructions. Resolve contradictions or state which instruction wins. Use numbered steps only when order matters. Add [NEEDS: <detail>] only when the missing value cannot be safely assumed and the user has asked you to stop interviewing.

Return:

Final prompt:

<self-contained prompt>

Assumptions: <only material defaults; omit when empty>
Gaps: <only unresolved [NEEDS: ...] items; omit when empty>

Do not solve the prompt, explain the rewrite, or offer a menu of next actions unless the user asks.