Multi-AI Review

Consensus-driven code and document review using multiple AI models.

A Claude Code skill that orchestrates parallel reviews from any combination of AI CLIs (Codex, Gemini, Aider, Ollama, etc.), synthesizes findings with confidence scoring, and iterates until consensus approval.

Version

v1.0.0

Why Multi-Model Review?

Every AI model has blind spots. A single model might miss a security vulnerability that another catches instantly. Multi-AI Review solves this by:

• Redundant coverage — Every model reviews every domain independently
• Consensus scoring — Findings agreed upon by 2+ models are high-confidence signals
• Conflict resolution — When models disagree, Claude arbitrates with full codebase context
• Iterative improvement — Fix-and-re-review loops continue until consensus approval

What It Reviews

Review Type	Domains	How to Trigger
Code Review	Security, Bugs, Robustness, Performance, Architecture, Scalability, UX, Maintainability	review this code or /review
Security Audit	Deep security analysis with OWASP checklist	review with a security focus
UX Review	Accessibility (WCAG 2.1 AA), responsive design, component compliance	review for ux and accessibility
Document Review	Completeness, accuracy, clarity, consistency, product thinking	review the spec at specs/029/plan.md
Pre-Merge Gate	All code domains + regression checks	review as a pre-merge gate

Quick Start

1. Install the Skill

Copy the multi-ai-review folder to your Claude Code skills directory:

# Option 1: Download and copy
# Download the multi-ai-review folder and copy it to your Claude Code skills directory:
cp -r multi-ai-review ~/.claude/skills/

# Option 2: Clone from GitHub
# git clone <repo-url>
# cp -r multi-ai-review ~/.claude/skills/

The skill registers as /review. The folder name stays multi-ai-review.

File structure:

~/.claude/skills/multi-ai-review/
  SKILL.md
  config.md
  references/
    domains.md
    synthesis.md
    prompts.md

2. Verify External CLIs

At least one external AI CLI is needed. Claude always participates as a reviewer.

# Check what's available
codex --version    # OpenAI Codex CLI
gemini --version   # Google Gemini CLI

3. Run a Review

Natural language works best. All of these trigger the skill:

/review                                          — Review uncommitted changes (standard profile)
review this code                                 — Same as above, natural language
review src/auth/login.ts for security issues     — Specific file, security focus
review the spec at docs/plan.md                  — Document review
do a deep review of the authentication changes   — Deep profile, scoped to auth
run a quick review                               — Fast single-model pass

Claude interprets your intent from natural language — no special syntax needed. If the intent is ambiguous, Claude will ask which profile and scope you want.

Review Profiles

Profile	Speed	Coverage	Best For
quick	Fast	2 domains, 1 model	Small changes
standard	Medium	All 8 domains, 2 models	Default for most work
deep	Slow	All domains, all models, 2-3 passes	Critical features
security	Medium	Security-focused, all models	Auth, APIs, data handling
ux	Medium	UX + accessibility, all models	UI components
doc	Medium	8 doc domains, 2 models	Specs, plans, PRDs
pre-merge	Medium	All domains + regression checks	Final merge gate

How It Works

Phase 1: Scope     → Detect changes, build file list, gather context
Phase 2: Review    → Run external models IN PARALLEL (Codex, Gemini, etc.)
Phase 3: Self-Review → Claude reviews independently with full codebase context
Phase 4: Synthesize → Cross-reference findings, assign confidence levels
Phase 5: Report    → Generate consensus report with matrix and scores
Phase 6: Iterate   → Fix issues, re-review, repeat until consensus
Phase 7: Certify   → Final certification when all models approve

Consensus Matrix Example

| Domain         | Claude | Codex | Gemini | Consensus | Spread |
|----------------|--------|-------|--------|-----------|--------|
| Security       | 9/10   | 8/10  | 9/10   | 8.7/10    | 1      |
| Bugs & Logic   | 8/10   | 9/10  | 8/10   | 8.3/10    | 1      |
| Performance    | 7/10   | 7/10  | 9/10   | 7.7/10    | 2      |
| Architecture   | 9/10   | 9/10  | 8/10   | 8.7/10    | 1      |

Spread > 2 is flagged — it means one model saw something others missed.

Confidence Levels

Level	Meaning	Action
HIGH	2+ models found the same issue	Fix immediately
MEDIUM	1 model found it	Verify, likely valid
REVIEW	Models disagree	Claude arbitrates

Adding Custom Models

Any CLI that accepts a prompt and returns text works. See config.md for full details.

Name: ollama-qwen
Verify: ollama list | grep qwen2.5-coder
Invoke: echo "{PROMPT}" | ollama run qwen2.5-coder:32b
Workspace access: No
Timeout: 600s

Supported out of the box: Codex CLI, Gemini CLI

Community examples: Aider, Ollama, sgpt, custom scripts

File Structure

multi-ai-review/
├── SKILL.md                    # Main orchestration (loaded on trigger)
├── config.md                   # Model registry + profile definitions
├── references/
│   ├── domains.md              # 16 review domains (8 code + 8 doc)
│   ├── synthesis.md            # Consensus algorithm + report template
│   └── prompts.md              # Prompt templates for external models
├── README.md                   # This file
└── CHANGELOG.md                # Version history

Progressive disclosure: Only SKILL.md loads initially. Reference files load on-demand based on the current review phase, keeping context window usage efficient.

Design Principles

1. Redundancy over division — Every model reviews everything. No splitting work.
2. Confidence from consensus — Multi-model agreement is the strongest signal.
3. Structured output — Same format from every model enables automated synthesis.
4. Progressive depth — Quick for small changes, deep for critical code.
5. Project-aware — Reads your CLAUDE.md for project-specific standards.
6. Graceful degradation — Works with 1 model, better with 2, best with 3+.
7. Iterative — Fix-and-re-review until consensus, max 3 iterations.

Customization

Project-Specific Standards

Add to your CLAUDE.md:

## Multi-AI Review Defaults
- Default profile: standard
- Default models: codex, gemini
- Always include domains: security
- Extra criteria: Must follow our API versioning pattern

Override on Invocation

review with the deep profile using codex and gemini, run 3 passes
review only the security and performance domains, critical and important only
review but skip the scalability and maintainability domains

Requirements

• Claude Code (any version with skills support)
• At least one external AI CLI (Codex, Gemini, Aider, Ollama, etc.)
• External CLIs must be authenticated and working independently

License

MIT — use freely, contribute improvements.

Contributing

1. Fork this repository
2. Add your improvement (new domain criteria, model support, profiles)
3. Test with the skill itself: review your changes as a doc review on your changes
4. Submit a PR

The best contributions are battle-tested improvements from real review sessions.