review-plan

name	review-plan
description	Review a pasted or freshly generated implementation, operation, research, or rollout plan as a pre-execution gate without executing it; return a concise human summary, prioritized plan-review findings, and a copy-ready prompt for another agent, with conditional routing for UI/frontend, superpowers, and subagent-suitable work.

Review Plan

Review the user's pasted or freshly generated plan and produce a human-readable review plus a prompt they can paste into a fresh agent. Treat this as a pre-execution gate: review happens after planning and before implementation, ECS operations, deployment, migration, or other irreversible work. Do not implement the plan, edit files, run deployment, or create follow-up artifacts unless the user explicitly asks for a separate implementation task after the review.

Intended Placement

Use this skill when the user pastes a plan or a UI exposes a "Review Plan" action immediately below a generated plan. The normal sequence is:

1. Generate or paste the plan.
2. Run review-plan.
3. Revise the plan if the verdict is ready with fixes, blocked, or unclear.
4. Execute only after the user explicitly asks for execution after review.

If the user mentions deployment, ECS, production, data migration, destructive actions, or external side effects, call that out as a reason to keep the review strictly before execution. Do not treat post-execution validation as this skill's main purpose.

Workflow

1. Identify the plan type: implementation, operations/release, research, refactor, UI/frontend, superpowers plan, or mixed.
2. State a concise verdict for the human: ready, ready with fixes, blocked, or unclear.
3. Review the plan for blockers, risk, ambiguity, missing verification, sequencing problems, scope creep, and handoff gaps.
4. Apply conditional skill routing rules.
5. Generate a copy-ready prompt that a fresh agent can use with no surrounding conversation.
6. Keep the copy prompt in revise-only mode unless the verdict is ready and the user's latest request explicitly asks for execution after review.

Conditional Skill Routing

Include these recommendations in both the review findings and the copy prompt when they apply.

Condition	Required handling
UI/UX, visual design, frontend pages/components, responsive behavior, interaction states, accessibility, or visual QA	Require ui-ux-pro-max and frontend-design. Review UI/UX, accessibility, responsive behavior, interaction states, visual consistency, and verification coverage.
Three or more mostly independent implementation tasks with low shared-file or shared-boundary conflict, and each task has a clear verification path	Recommend subagent-driven-development for same-session execution.
Tightly coupled implementation, unclear architecture, one large shared file, or unresolved requirements	Do not recommend subagent-driven-development; ask for decomposition or clarification first.
Existing superpowers plan	Preserve the superpowers execution order, checklist style, and handoff conventions. Do not replace the plan's workflow; provide delta-style revision instructions.
Plan creation or major redesign is still needed	Recommend revising the plan before execution. Do not tell the next agent to execute immediately.
Deployment, ECS, production, migration, destructive operation, cost-incurring operation, or external side effect is present	Treat review as a pre-execution safety gate. Require explicit user approval before execution and include rollback/verification gaps as review findings when missing.

Review Standards

Prioritize findings by implementation impact:

• Blocker: likely prevents correct execution or creates unacceptable safety, data, security, or deployment risk.
• High: likely causes rework, user-visible failure, broken verification, or unreliable handoff.
• Medium: meaningful ambiguity, missing edge case, sequencing weakness, or maintainability risk.
• Low: polish, wording, or optional completeness improvement.

Prefer concrete fixes over abstract advice. If evidence is only in the pasted plan, cite it as plan text. If repository evidence is needed but unavailable, mark the item as Needs verification rather than inventing facts.

Output Contract

Return exactly these top-level sections unless the user asks for a different format.

## Summary

- Verdict:
- Biggest risks:
- Must-fix items:
- Ready status:

## Review Findings

### Blockers

- [Blocker] <finding and concrete fix>

### High / Medium Risks

- [High] <finding and concrete fix>

### Missing Verification

- [Medium] <finding and concrete fix>

### Sequencing / Scope

- [Low] <finding and concrete fix>

## Copy Prompt

```text
Role:
You are an agent revising the following plan before implementation.

Mode:
Revise only. Do not execute the plan unless the user explicitly asks you to execute after revision.

Required skills:
- <list required skills and why; include conditional UI/frontend/subagent/superpowers guidance when applicable>

Original plan:
<paste or summarize the original plan faithfully. Preserve essential details, commands, file paths, constraints, and acceptance criteria.>

Review findings to address:
<prioritized findings from this review>

Instructions:
- Address blockers first.
- Preserve the original scope unless a finding says the scope is unsafe or incomplete.
- Preserve superpowers conventions if this is a superpowers implementation plan.
- If UI/frontend work is present, include UI/UX, accessibility, responsive, interaction-state, and visual QA requirements.
- If subagent-driven-development is appropriate, decompose into independent tasks with disjoint ownership and verification steps.
- If execution is not yet safe, return an improved plan and open questions instead of implementation steps.

Expected output:
- Revised plan or execution-ready handoff.
- Verification checklist.
- Open questions or blockers.

Constraints:
- Do not ignore unresolved blockers.
- Do not add unrelated features or refactors.
- Keep the response copy-paste ready for the next agent.
```

If a category has no findings, write - None found. under that category. Keep the human-facing summary short; put detail in the findings and copy prompt.

Prompt Construction Rules

• Make the copy prompt self-contained: include the original plan content or a faithful compact restatement with all important commands, file paths, constraints, and acceptance criteria.
• Use Mode: Revise only by default. Use Mode: Revise, then execute only if explicitly requested only when the user's latest request clearly wants execution after revision.
• For generated-plan UI actions such as "Review Plan", assume the user asked for review only. Do not infer permission to execute from the presence of an implementation, deployment, or ECS plan.
• Do not claim that named skills can be forcibly activated in another runtime. Instead, instruct the next agent to use them.
• Keep the next-agent instructions imperative and testable.
• Do not include private chain-of-thought or hidden reasoning.

Common Failure Modes

Failure	Correction
Reviewing only prose quality	Review buildability, risk, verification, sequence, and handoff clarity.
Adding UI skills only to the copy prompt	Also include UI-specific review findings when UI/frontend content exists.
Over-recommending subagents	Recommend subagent-driven-development only when tasks are separable and verifiable.
Telling the next agent to execute unsafe plans	Keep default mode as revision-only until blockers are resolved.
Rewriting superpowers plans from scratch	Preserve their conventions and provide delta-style fixes.