菜单
Create a durable plan when explicitly requested, required by the High-risk tier, or needed for resumable dependency coordination; skip routine native planning.
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
Execute an existing plan when implementation is authorized and dependent outcomes need tracking; skip short routine checklists.
Mandatory newcomer-safe router for every conversation. Classify by action, impact, and named risk; add workflows only when their descriptions match.
Resolve ambiguous product or system work before implementation. Use when materially different interpretations would change scope, architecture, or consequential outcomes; skip precise tasks.
Complete an authorized commit, push, integration, PR, branch, or worktree action while preserving user state. Use after relevant verification when the user or repository requires delivery.
Evaluate incoming code-review feedback against repository evidence. Keep evaluation-only requests read-only; implement supported fixes only when authorized.
Independently review a defined change before integration when requested or when material security, data, concurrency, migration, contract, or cross-module risk warrants it; exclude incoming feedback.
| name | writing-plans |
| description | Create a durable plan when explicitly requested, required by the High-risk tier, or needed for resumable dependency coordination; skip routine native planning. |
Use a repository file only when the High-risk tier requires it or for cross-session survival and material dependency coordination. A native plan alone does not require this file.
AGENTS.md, requirements, relevant code, and repository documentation convention../superpowers/docs/plans/YYYY-MM-DD-HHMMSS-01-plan-<slug>.md; keep that timestamp and slug stable.# <Task title>
**Goal:** <one sentence>
**Why planning is required:** <material trigger>
**Acceptance:** <observable completion criteria>
### Outcome N: <result>
- Work: <requirements, named files/resources, state/data flow, and failure behavior>
- Risks/open questions: <material privacy, security, or blocking unknowns; omit when none>
- Verify: `<repo-specific command>`
Keep outcomes dependency-ordered and omit code, routine mechanics, predicted output, and artificial microsteps. Update the same plan only after a material requirement, scope, design, or verification-strategy change.
Create it before implementation. For destructive or consequential external-state work, put recovery, stop conditions, and pre-execution checks in acceptance. For production-data work, include target, approvals, dry-run counts, backup/restore evidence, integrity checks, batch/abort thresholds, and sensitive-output hygiene. For production deployment, include target revision, approvals, rollout, health evidence, rollback, and monitoring stop conditions.
If the current mode forbids writes, name the intended path and create it only after writes are authorized. A durable plan does not automatically require a log, review, worktree, or separate completion gate; select each independently.