菜单
Turns an agreed spec or conversation into an implementation plan of self-contained tasks ready for subagent delegation. Distinct from Claude Code's built-in plan mode. Use when the user asks to "make an implementation plan", "break this into tasks", "計画を立てて", or before multi-task implementation work.
用 Codex 或 Claude 帮你安装 复制这段 Prompt，粘贴到 Codex、Claude 或其他助手里，让它检查 Skill 页面并帮你完成安装。
基于 SOC 职业分类
A relentless one-question-at-a-time interview that sharpens a plan, decision, or idea until shared understanding is reached — not limited to software. Use when the user wants requirements or a decision clarified before acting, wants to stress-test their thinking, or asks to "grill me", "詰めて", "brainstorm this".
Executes an implementation plan or agreed work by orchestrating fresh subagents — one implementer per task, a reviewer between tasks — while the main model plans, briefs, and evaluates without writing code itself. Use when the user asks to "implement the plan", "start implementation", "実装して", or hands over a plan or spec to execute.
A feedback-loop-first discipline for diagnosing hard bugs and performance regressions. Use when the user reports something broken, throwing, failing, flaky, or slow, or asks to "diagnose", "debug this", "原因を調べて".
Builds a clearly-marked throwaway prototype to answer a design question before committing to an implementation. Use when the user wants to sanity-check whether logic or a state model feels right, explore what a UI should look like, or asks to "prototype this", "試作して".
Creates an Architecture Decision Record for a decision that is hard to reverse, surprising without context, and the result of a real trade-off. Use when such a decision has just been made, or the user asks to "record this decision", "ADRにして", or mentions decision records.
Reviews the diff since a fixed point along two axes run as parallel subagents — Spec (does the change match what was asked?) and Standards (repo conventions plus a code-smell baseline) — and reports them side by side. Use when the user asks to review a branch, work-in-progress changes, a diff, or "review since X".
| name | write-plan |
| description | Turns an agreed spec or conversation into an implementation plan of self-contained tasks ready for subagent delegation. Distinct from Claude Code's built-in plan mode. Use when the user asks to "make an implementation plan", "break this into tasks", "計画を立てて", or before multi-task implementation work. |
Turn an agreed spec or conversation into a plan a fresh subagent can execute one task at a time, with no other context.
State project-wide requirements once, at the top of the plan — versions, naming conventions, style rules, anything that applies to every task. Every task's requirements implicitly include this section; don't repeat it per task.
A task is self-contained when a fresh subagent, given only that task's text, has everything it needs to do the work and know when it's done. That's the only sizing rule — no fixed step count or timing.
Each task states:
Every task needs the actual content an implementer needs to act on. These are plan failures:
Before finalizing, check the plan against the spec: does every requirement map to a task? Do names and signatures used in later tasks match what earlier tasks define? Fix inline rather than re-reviewing.
Check .claude/engineering.local.md for a plans: key in its frontmatter — a path, or none to skip saving. If the file doesn't exist, ask once where (or whether) to save this plan, accept "don't save" as an answer, and mention /setup-engineering for making the choice permanent. Never write plans into CLAUDE.md or AGENTS.md.
Once the plan is saved, implement is a candidate for executing it.
