Writing Plans

Write the smallest plan that makes the authorized work executable and verifiable. Prefer outcomes and contracts over a script for every keystroke. Do not write the implementation twice in a plan and then ask another model to transcribe it.

For a small change, a short task list is sufficient. For substantial work, save a plan at docs/superpowers/plans/YYYY-MM-DD-topic.md, or the project's preferred location. Include:

• Goal, scope, and accepted design or requirements.
• Files or components affected and dependencies between tasks.
• Required behavior, interface contracts, edge cases, and acceptance criteria.
• Meaningful verification and its independent oracle; commands when known.
• Ownership only where work is actually delegated.
• Optionally a "## 実装時の差異" section where deviations from the design are recorded during implementation (when there is no spec file).

Provide exact code only when an API contract or non-obvious constraint requires it. Avoid mandatory 2-5 minute steps, full implementation listings, and a separate commit or review for each task. Tasks should produce coherent, independently testable results.

Start only after the brainstorming Design Gate has passed: the design was reviewed and the user approved it. Self-check coverage, ordering, and contradictions. For a substantial plan, request one plan review: reserve it with node .claude/hooks/review-budget.cjs begin --phase plan --roles document-reviewer (.codex in a Codex-only project) and give the reviewer plan-document-reviewer-prompt.md. After the plan review, start implementation without asking the user again. If the plan review limit is reached, do not ask: record the remaining findings and how you handle them, and start implementation (documents/development/development-policy.md §1.0 "承認後の進め方"). Only when the user has already approved exceeding it, run extend yourself with the same review-budget script used to reserve the review (.claude/hooks/…, .codex/hooks/…, or npx --no ai-dev-helm review-budget (where the CLI is not installed, npx -y @crearize/ai-dev-helm@<version in .ai-dev-helm.json> review-budget)): extend --phase plan --rounds <1-3> --reason "<the user's approval, quoted>"; never ask the user to run commands, inspect state, or reset state.

A plan records the decisions the implementer cannot make alone, not a transcript of the code. For each input class or failure mode the spec implies but no task's tests exercise, add a line and the test that pins it to the owning task (Review Focus).

Handoff integrity:

• Quote the spec verbatim, preserving invisible characters.
• For a pure move, verify that the content before and after the move matches verbatim.

The parent may implement directly. Delegate only useful independent bounded work, accounting for briefing, rediscovery, and integration cost as well as model cost. Use executing-plans for continuation or subagent-driven-development when delegation is chosen. Do not ask the user to select an execution mechanism already authorized by the task.

The final change still passes the project's independent quality-check and relevant verification.