UI はローカライズ済みです。スキルの紹介と説明文書は選択した言語へ AI が自動翻訳し、コード・コマンド・パス・専門用語は原文のままです。原作者の文書が優先されます。
スキルを探す
生産性・共同作業 · SKILL.md

ask-then-build

ask-then-build

Clarify a feature, change, or refactor through questions, then write a build prompt for another agent. Use for /ask-then-build or requests to ask questions before producing an implementation prompt.

Independent GitHub snapshot · Automated classification · Original documentation governs

スキルの説明

GitHub

name: ask-then-build description: 'Clarify a feature, change, or refactor through questions, then write a build prompt for another agent. Use for /ask-then-build or requests to ask questions before producing an implementation prompt.'

Turn a feature idea into a build prompt for another agent, in two phases.

Phase 1 — Questions

  1. Identify the 3-6 most non-obvious open questions about the feature or change: edge cases, where it lives in the UI, failure behavior, scope boundaries, how it interacts with existing rules.

  2. Ask them ONE at a time. For each: the question, top options A-D, your preferred pick with a one-line reason. Then stop and wait. Use this exact layout. Put a blank line between EVERY block. Markdown renderers (bb included) collapse single newlines into one paragraph, so options written back-to-back render as one run-on blob.

    Question title?

    One line of context, if needed.

    A. First option.

    B. Second option.

    C. Third option.

    D. Fourth option.

    My pick: A. One-line reason.

    Never number questions as "1 of N" unless you truly know N. If the count is open, keep it open and keep asking until you have the context you need.

  3. When the user answers, record the decision immediately — update the repo's docs (product requirements, ADR, or README) if the project has them.

  4. If the user overrides an earlier documented decision, update the docs right away and say what was superseded.

Phase 2 — Prompt

After the last answer, deliver ONE concise paragraph prompt for another agent. It must include, in this order:

  1. Read-first files: the authoritative docs (AGENTS.md, requirements, ADRs).
  2. What to build: numbered implementation steps, concrete file-level where useful.
  3. How to validate: build, lint, and a manual check with real data.
  4. Rules: don't commit, report back with files changed.

Style

  • Very concise, plain English, short sentences.
  • Never bundle questions. Never write the prompt before all answers are in.
  • Never put two options on one line. Always a blank line between A, B, C, D and before "My pick".
  • Keep the prompt to a single paragraph — if it needs two, the scope is too big; say so.
保存したスキル