AgentSkillsCN

review-plan

对实施计划进行多LLM验证。支持--spec模式以验证需求,也支持默认模式以评估技术可行性。重点关注80/20高影响力问题。

SKILL.md
--- frontmatter
name: review-plan
description: Multi-LLM validation for implementation plans. Supports --spec mode for requirements validation and default mode for technical feasibility. Focuses on 80/20 high-impact issues.

Plan Review

Validate implementation plans using multiple LLMs before writing code. Focuses on 80/20 analysis - identifies the critical few issues that matter vs the trivial many that don't.

Two modes:

  • Default: Technical feasibility review (can this be built?)
  • --spec: Requirements validation review (are we building the right thing?)

Related:

  • Use /create-plan to generate plans
  • Use /multi-review for code/architecture decisions

Usage

code
/review-plan <plan-file>                           # Technical review (default)
/review-plan <plan-file> --spec                    # Requirements validation
/review-plan <plan-file> --req <requirements>      # Compare against original request
/review-plan <plan-file> --context <relevant-files> # Include codebase context
/review-plan <plan-file> --quick                   # Fast single-model check (Gemini only)

Two Review Modes

Default Mode: Technical Feasibility

Question: Can this plan be implemented?

ValidatesQuestions
FeasibilityDo required APIs exist? Dependencies available?
RisksData integrity? Security? Breaking changes?
AmbiguityCan developer implement without guessing?
ScopeOver-engineered? What can be cut?

Spec Mode (--spec): Requirements Validation

Question: Are we building the right thing?

ValidatesQuestions
CompletenessAll user requirements captured?
ClarityRequirements testable and unambiguous?
ScopeOut-of-scope clearly defined?
AcceptanceCriteria specific and verifiable?
StakeholderCould non-technical person approve this?

Use --spec FIRST, then default mode:

bash
/review-plan plan.md --spec    # Step 1: Requirements OK?
/review-plan plan.md           # Step 2: Technical OK?

Core Philosophy: 80/20 Prioritization

Not all issues are equal. This skill explicitly separates:

CategoryCriteriaAction
MUST FIXBlocks implementation, causes data loss, security risk, or will definitely failFix before coding
SHOULD FIXImproves quality but plan works without itAddress if time permits
DEFERValid concern but not for this iterationTrack for future
SKIPOver-engineering, premature optimization, or gold-platingIgnore

Technical Mode Blockers

A blocker IS:

  • Missing critical step that makes implementation impossible
  • Data integrity risk (orphaned records, corruption)
  • Security vulnerability (injection, auth bypass)
  • Undefined contract that blocks API integration
  • Ambiguity that requires guessing during implementation

A blocker is NOT:

  • Missing documentation
  • Imperfect error messages
  • Edge cases that affect <1% of users
  • "Best practice" that isn't required

Spec Mode Blockers

A blocker IS:

  • Missing user requirement from original request
  • Acceptance criteria that can't be tested
  • No "out of scope" section (scope creep risk)
  • Contradictory requirements
  • Requirements that stakeholder hasn't agreed to

A blocker is NOT:

  • Missing nice-to-have features
  • Imperfect wording
  • Missing implementation details (that's for technical review)

Models

See .claude/docs/multi-llm-review.md for model selection, quota limits, and fallback logic.

Plan Agents (Optional)

After multi-LLM review, run domain-specific agents for deeper analysis.

MUST RUN (Every Plan)

AgentCatches
design-flaw-finderLogical flaws, missing steps, impossible sequences
simplicity-reviewerOver-engineering, YAGNI violations

Architecture (Complex Plans)

AgentWhen to Use
system-design-reviewerMulti-component features
separation-of-concerns-reviewerLayer violations
client-server-alignmentFrontend+backend changes
permission-design-auditorPermission model changes
type-design-analyzerNew type hierarchies

Domain-Specific

AgentWhen to Use
api-contract-reviewerAPI endpoint changes
rest-api-expertREST API design
database-architecture-reviewerSchema changes
ux-design-reviewerUI/UX changes
edge-case-ux-analystError/empty/loading states
confluence-alignment-reviewerDocument/pages features
threat-modelerSecurity-sensitive features
arch-system-designLarge-scale architecture
repo-architectRepository structure

Note: These are PLAN agents. Do NOT use code agents (go-backend, xss-reviewer, etc.) on plans.

Full Agent Reference

For complete agent listing (~140 agents), see .claude/agents/AGENT_REGISTRY.md.

Workflow

Step 1: Gather Context

  1. Read the plan file
  2. If --req provided, read original requirements
  3. If --context provided, read relevant codebase files
  4. Detect mode (--spec or default)

Step 2: Run Reviews in Parallel

Launch all models from .claude/docs/multi-llm-review.md simultaneously (single message, multiple tool calls). This includes Codex, Gemini, AND seq-server — do NOT skip any.

Quick mode (--quick): Use only Gemini.

Step 3: Synthesize with 80/20 Filter

  1. MUST FIX - Only issues where 2+ models agree AND meets blocker criteria
  2. SHOULD FIX - Single-model findings that are valid but not blocking
  3. DEFER - Valid concerns for future iterations
  4. SKIP - Reject over-engineering suggestions

Be ruthless. Most "warnings" from LLMs are nice-to-haves.

Prompt Templates

Technical Review Prompt (Default)

code
Review this implementation plan for readiness to code.

## The Plan
<plan>
[paste plan content]
</plan>

## Original Requirements (if provided)
<requirements>
[paste requirements]
</requirements>

## CRITICAL: Apply 80/20 Thinking

Focus on the 20% of issues that cause 80% of problems.

**A MUST FIX blocker is ONLY:**
- Missing step that makes implementation impossible
- Data integrity risk (orphaned records, corruption, loss)
- Security vulnerability (injection, auth bypass, SSRF)
- Undefined contract that blocks integration
- Ambiguity requiring guesswork during implementation

**NOT a blocker (put in DEFER or SKIP):**
- Missing docs, imperfect error messages
- Edge cases affecting <1% of users
- "Best practices" not strictly required
- Future-proofing for hypothetical scenarios

## Evaluate (priority order)

1. **Feasibility**: Do required APIs/functions exist?
2. **Risks**: Data integrity? Security? Breaking changes?
3. **Ambiguity**: Can developer implement without guessing?
4. **Scope**: What can be cut for MVP?

## Output

1. **MUST FIX** (0-3 max): What breaks? How to fix?
2. **SHOULD FIX** (0-5): Why it matters, why not blocking
3. **DEFER**: Valid for later
4. **SKIP**: Over-engineering to reject
5. **VERDICT**: READY / NEEDS WORK / MAJOR REVISION

Spec Review Prompt (--spec flag)

code
Review this plan's REQUIREMENTS for completeness and clarity.
DO NOT review technical implementation - only requirements.

## The Plan
<plan>
[paste plan content]
</plan>

## Original User Request (if provided)
<request>
[paste original request]
</request>

## CRITICAL: Requirements-Only Review

You are validating "are we building the right thing?" NOT "can we build it?"

**A MUST FIX blocker is ONLY:**
- User requirement from original request NOT captured in plan
- Acceptance criteria that cannot be tested/verified
- Missing "Out of Scope" section (scope creep risk)
- Contradictory or ambiguous requirements
- Unstated assumptions that could surprise stakeholders

**NOT a blocker (put in DEFER or SKIP):**
- Missing implementation details
- Technical approach concerns
- Nice-to-have features not in original request
- Imperfect wording that's still clear

## Evaluate

1. **Completeness**: Every requirement from original request captured?
2. **Testability**: Each requirement has verifiable acceptance criteria?
3. **Scope Boundaries**: "Out of Scope" section exists and is clear?
4. **Clarity**: Could a stakeholder approve without asking questions?
5. **Assumptions**: Are implicit assumptions made explicit?

## Output

1. **MUST FIX** (0-3 max): Missing/unclear requirements
2. **SHOULD FIX** (0-5): Improvements to clarity
3. **DEFER**: Nice-to-haves for future
4. **SKIP**: Scope creep suggestions to reject
5. **VERDICT**: READY / NEEDS WORK / MAJOR REVISION

Output Format

markdown
## Plan Review: [plan-name]
### Mode: Technical / Spec

### MUST FIX (Blockers)
| Issue | Found By | What Breaks | Fix |
|-------|----------|-------------|-----|
| [description] | o3-pro, gpt-5.2 | [why blocked] | [fix] |

*If empty: "None - plan is ready"*

### SHOULD FIX (Quality Improvements)
- [Issue] - [why it matters but isn't blocking]

### DEFER (Future Iterations)
- [Issue] - [why it can wait]

### SKIP (Rejected Suggestions)
- [Suggestion] - [why this is over-engineering/scope-creep]

### What's Good
- [Validated aspects]

---

### Verdict: READY / NEEDS WORK / MAJOR REVISION
### Confidence: HIGH/MEDIUM/LOW

Verdict Criteria

VerdictCriteria
READY0 MUST FIX items. Proceed.
NEEDS WORK1-2 MUST FIX items. Quick fixes needed.
MAJOR REVISION3+ MUST FIX or fundamental flaw. Rethink.

CLI Reference

See .claude/docs/multi-llm-review.md for CLI commands and quota fallback logic.

Examples

bash
# Requirements validation first
/review-plan implementation-plans/feature.md --spec

# Then technical review
/review-plan implementation-plans/feature.md

# With original user request for comparison
/review-plan implementation-plans/feature.md --spec --req "User asked for X with Y"

# Quick check
/review-plan implementation-plans/small-fix.md --quick

Integration with Workflow

code
User request
    ↓
/create-plan                    # Generate structured plan
    ↓
/review-plan plan.md --spec     # Validate requirements ← NEW
    ↓
/review-plan plan.md            # Validate technical approach
    ↓
Fix MUST FIX items
    ↓
User approval
    ↓
Implementation
    ↓
/review-code                    # Code review

Tips

  • Run --spec before default - No point validating technical if requirements are wrong
  • Be skeptical of LLM "blockers" - Most are actually SHOULD FIX or DEFER
  • Parallel execution: Run all model calls in single message
  • Quick check: Use --quick for initial pass
  • Trust judgment: If it feels like over-engineering, it probably is