Multi-Model Review

Review PRs or code changes across multiple AI models to get diverse perspectives, then interactively work through findings with the user.

Capabilities

• Review PRs using GPT 5.2, Gemini 3 Pro, and Claude Opus 4.5 in parallel
• Generate individual model reviews and a synthesized summary
• Track issue status (applied, skipped, discussed) across all model reviews
• Present findings interactively for user decision

When to Use

• PR review requiring thorough analysis
• Skills or prompts where token efficiency matters
• Architectural changes benefiting from multiple perspectives
• Any code where catching edge cases is critical

Execution

Phase 1: Gather Context

Desired end state: Full understanding of what's being reviewed

Required context:

• PR number/URL or branch with changes
• Output location (user-specified or repo root)
• Any specific review focus areas (optional)

Gather:

• PR diff and changed files
• Related specs, plans, or reference materials
• Relevant context from the codebase

Phase 2: Parallel Model Reviews

Desired end state: Three independent reviews saved to output location

Models: GPT 5.2, Gemini 3 Pro, Claude Opus 4.5

Review prompt template:

You are reviewing [description of changes]. Review critically against [context].

## Changes Being Reviewed
[Include full content of changed files]

## Reference Context
[Specs, plans, related patterns]

## Review Task
Write a review covering:
1. **Correctness**: Do changes implement requirements? Any gaps?
2. **Pattern Comparison**: What patterns from [reference] could improve these? Missing concepts?
3. **Mistakes and Issues**: Bugs, inconsistencies, problematic patterns
4. **Improvement Suggestions**: Concrete, actionable improvements
5. **Token Efficiency**: Opportunities to reduce verbosity (for prompts/skills)
6. **[Domain-specific criteria]**

Write in markdown format suitable for saving to a file.

Output files:

• REVIEW-{MODEL}.md for each model
• Use task tool with model parameter for each review

Phase 3: Synthesis

Desired end state: Consolidated findings document identifying consensus and unique insights

Synthesis structure:

# Review Synthesis: [Subject]

**Date**: [date]
**Reviewers**: GPT-5.2, Gemini 3 Pro, Claude Opus 4.5
**PR/Changes**: [reference]

## Consensus Issues (All 3 Models Agree)
[Issues flagged by all models - highest priority]

## Partial Agreement (2 of 3 Models)
[Issues flagged by two models]

## Single-Model Insights
[Unique findings worth considering]

## Priority Actions
### Must Fix
[Critical issues]

### Should Fix
[High-value improvements]

### Consider
[Nice-to-haves]

Output file: REVIEW-SYNTHESIS.md

Phase 4: Interactive Resolution

Desired end state: All findings addressed (applied, skipped, or discussed)

For each model review (start with most comprehensive, typically Opus):

Present each finding to user in chat, not a selector:

## Finding #N: [Title]

**Issue**: [Description]

**Current**:
[Show current code/text]

**Proposed**:
[Show proposed change]

**My Opinion**:
- [Rationale for applying, skipping, or discussing]

---

**Your call**: Skip, discuss, or apply?

Track status:

• applied - Change made
• skipped - User chose not to apply
• discussed - Modified based on discussion, then applied or skipped

Cross-reference: When moving to next model's review, check if finding was already addressed:

• If same issue was applied → "Already addressed in Finding #N from [Model]"
• If same issue was skipped → "Previously skipped (Finding #N from [Model]). Revisit?"
• If similar but different angle → Present as new finding

Completion

Report back:

• Number of findings per model
• Applied vs skipped counts
• Summary of key changes made
• Any remaining items flagged for future consideration

Output Location

Ask user for output location. Options:

• Specific directory path
• Default: Repository root as reviews/[PR-or-branch-name]/

Quality Criteria

• Each model review should be independent (don't share results between model calls)
• Synthesis should identify true consensus vs coincidental overlap
• Interactive phase should be efficient—group related findings when possible
• Track cross-model duplicates to avoid re-presenting same issue