AgentSkillsCN

cross-model-review

在提交前使用另一款LLM（Codex、Gemini）对计划与实施进行QA

SKILL.md
--- frontmatter
name: cross-model-review
description: Use another LLM (Codex, Gemini) to QA plans and implementations before committing
trigger: When starting multi-phase feature work or before merging significant PRs

Cross-Model Review Workflow

Use a second LLM (Codex CLI or Gemini) to review plans and implementations. This catches blind spots that a single model might miss.

When to Use

  • •Multi-phase features (3+ phases)
  • •High-risk changes (auth, payments, database migrations)
  • •Before merging large PRs (10+ files)
  • •When you want a "second opinion" on architecture

4-Step Workflow

code
┌─────────────────────────────────────────────────────────────┐
│  STEP 1: PLAN                              Claude Code      │
│  ─────────────                             Plan Mode        │
│  Open Claude Code in plan mode (Terminal 1).                │
│  Claude interviews you via AskUserQuestion.                 │
│  Produces a phased plan with test gates.                    │
│                                                             │
│  Output: .claude/plans/{feature-name}.md                    │
│                                                             │
│                          ▼                                  │
│                                                             │
│  STEP 2: QA REVIEW                         Codex/Gemini     │
│  ──────────────────                        Terminal 2       │
│  Open Codex CLI or Gemini in another terminal.              │
│  Prompt: "Review this plan against the actual codebase.     │
│           Insert 'Phase 2.5' if steps are missing.          │
│           Add findings but never rewrite original phases."  │
│                                                             │
│  Output: .claude/plans/{feature-name}.md (with findings)    │
│                                                             │
│                          ▼                                  │
│                                                             │
│  STEP 3: IMPLEMENT                         Claude Code      │
│  ──────────────────                        New Session      │
│  Start fresh Claude Code session (Terminal 1).              │
│  Implement phase-by-phase with test gates at each phase.    │
│                                                             │
│                          ▼                                  │
│                                                             │
│  STEP 4: VERIFY                            Codex/Gemini     │
│  ────────────────                          Terminal 2       │
│  Prompt: "Verify implementation against the plan.           │
│           Flag any deviations or missing requirements."     │
│                                                             │
└─────────────────────────────────────────────────────────────┘

Prompts for Codex/Gemini

Plan Review Prompt

code
Review this implementation plan against the actual codebase:

[paste plan content]

Instructions:
1. Check if referenced files/functions exist
2. Check if the proposed approach matches existing patterns
3. Insert intermediate phases ("Phase 2.5") where steps are missing
4. Add "Codex Finding:" or "Gemini Finding:" headings for issues
5. DO NOT rewrite the original phases - only add to them

Implementation Review Prompt

code
Verify this implementation against the original plan:

PLAN: [paste plan]
IMPLEMENTATION: [paste git diff or file changes]

Check for:
1. Missing requirements from the plan
2. Deviations from the agreed approach
3. Edge cases not handled
4. Tests that should exist but don't

Tips

  • •Keep terminals side-by-side: Claude Code in Terminal 1, Codex/Gemini in Terminal 2
  • •Don't over-use: Simple features don't need cross-model review
  • •Trust the findings: If the second model flags something, investigate it
  • •Document findings: Keep the annotated plan for future reference

Source

Pattern extracted from shanraisshan/claude-code-best-practice