菜单
Runs six parallel Task subagents to critique a development plan (feedback only). Use when checking whether a plan is sound before implementation; no code or plan edits unless the user asks afterward.
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
| name | review-dev-plan |
| description | Runs six parallel Task subagents to critique a development plan (feedback only). Use when checking whether a plan is sound before implementation; no code or plan edits unless the user asks afterward. |
| disable-model-invocation | false |
/review-dev-plan)SUBAGENT_MODEL=composer-2.5
Every Task subagent: model = slug above (must be valid for your Cursor build; edit the value only).
Feedback only — do not change code, rules, or the plan unless the user asks afterward. Reply with a short synthesis in chat.
Resolve in order:
documentation/jobs/**/DEVELOPMENT_PLAN.md (state assumption).IMPLEMENTATION_PLAN.md or plan.md under documentation/jobs/ (state assumption).Self-contained prompts each; they do not see this chat. Each: orient (e.g. ARCHITECTURE.md, documentation/DOC_INDEX.md, documentation/DOC_APP_VISION.md, relevant .cursor/rules/, paths/READMEs the plan touches) → read the plan (path or pasted text) → structured feedback (gaps, risks, sequencing, validation, boundaries, checklist quality).
The Industry precedent agent must read .agents/skills/pattern-review/references/rubric.md and alert-template.md (same as pattern-review).
Perspectives (6):
| # | Role |
|---|---|
| 2 | You pick — distinct lenses (e.g. delivery, tests, security/ops, UX/API); not rebel, scale/mod, or industry precedent. |
| 1 | Rebel (fixed) — when a clean-slate / different approach is better; say so plainly + high-level alternative. No incremental polish only. |
| 1 | Scalability / modularity (fixed) — scale (load, cost, ops, data growth, failure modes, ceilings) and evolution (coupling, layers, ports, swappable deps, extension points, lock-in, future-you debt). |
| 1 | Industry precedent (fixed) — compare the plan to common product/engineering practice using .agents/skills/pattern-review/references/rubric.md: pick relevant aspects for this plan, name precedents, flag material divergence and missing § Pattern & precedent content. |
Short description per Task (e.g. “— rebel”, “— scale/mod”, “— industry precedent”).
Synthesize: agreements, must-fix vs nice-to-have, items ≥2 agents flagged. Do not rewrite the plan unless asked. Fold in user constraints (scope, timebox, risks) if given.
Allowed plan edits without full rewrite: Update only the Plan review row in Summary and rows in Decisions made when the user accepts the critique (or record waiver in Decisions made). Do not edit phase steps, Pattern & precedent body, or other sections unless the user explicitly asks.
Update the plan’s Plan review row to Done <date> when the user accepts the critique (or record waiver in Decisions made).
Next: .agents/skills/implement/SKILL.md when Plan review is Done and gates pass — or back to plan if the plan must change materially.
| Skill | Role |
|---|---|
review-dev-plan | Qualitative multi-lens plan critique (includes industry precedent) |
validate | Parallel repo rule subagents (.cursor/rules/) — JSON findings |
For Complexity M/L or rule-heavy plans, run both when appropriate.
validate — rule-shaped plan-review vs impl-reviewpattern-review — industry precedent procedure and rubric SSOTplan — creates DEVELOPMENT_PLAN.mdfeature — product/requirements processimplement — executes the planBundles all uncommitted work from multiple Cursor agent threads on the same branch and checkout into one finish commit, then pushes. Use when the user runs /bundle-ship, /finish-all, or asks to finish and push everything across threads on one checkout.
Interview the user until scope edges and interaction boundaries are aligned and the decision tree is resolved. Focuses on where the feature stops, what it will NOT do or touch, and how it meets other functionality. Chat-only alignment — authors no artifact. Use to stress-test scope/boundaries before feature or plan, or when the user says "grill me".
Picks the next action and skill, then invokes it (read that SKILL.md and execute the workflow in the same turn). Mid-thread: continue the logical next step on active work. Idle or new: fall back to gates and app-tasks.json backlog. Maps situations to project, user, and plugin skills; resolves overlaps. Product context SSOT: documentation/DOC_APP_VISION.md. Use when the user types /router, asks which skill to use, starts ambiguous work, or routes slash commands to SKILL.md files. Bare /router is next-action dispatch (not finish).
Challenges and simplifies a concrete feature to reduce steps, clicks, and cognitive load (flow and code modes by default). Use when the user targets a workflow and asks whether it is overbuilt, or runs /challenge.
Systematically audit the codebase for consolidation, abstraction, and redundancy removal opportunities across features. Use when the user asks to find duplication, consolidate code, audit for redundancy, reduce repetition, or identify shared patterns that should be unified.
Structured scientific debugging for runtime bugs, regressions, flaky behavior, frontend/backend failures, Supabase/RLS/Auth/Storage issues, and user-reported errors. Use when the user invokes /debug, asks to debug, reports an error, or describes inconsistent behavior.