multi-model-review

name	multi-model-review
description	Parallel multi-model quality review pattern — launch diverse AI reviewers simultaneously, collect findings, deduplicate, and fix. Use when reviewing articles, code, specs, proposals, or any deliverable that needs quality assurance. Trigger phrases include "dual review", "multi-model review", "parallel review", "quality review", "cross-model check", "review with multiple models".

Multi-Model Review Skill

The multi-model review pattern ensures quality by running the same deliverable through multiple AI models in parallel, collecting independent findings, and resolving them systematically. This eliminates single-model blind spots and catches hallucinations, factual errors, and quality issues.

When to Use

Scenario	Minimum Models	Recommended
Blog or newsletter articles	2	Claude Opus + GPT Codex
Tier 3 code changes	3	Claude Sonnet + GPT 5.2 + Claude Opus
Tier 4 safety-critical changes	4	3 standard + 1 safety-focused
Client proposals	2	Any 2 diverse models
Agent/skill definitions	2	Claude Opus + GPT 5.4

Core Pattern

Step 1: Prepare the Review Prompt

Write a single review prompt that clearly specifies:

• What the deliverable IS (article, code, spec, proposal)
• What to review FOR (accuracy, quality, safety, compliance, style)
• What format to report in (severity levels, location, issue, fix)
• Context needed (spec reference, style guide, brand rules)

Step 2: Launch Reviews in Parallel

CRITICAL: All review agents must be launched SIMULTANEOUSLY via the task tool. Never run them sequentially — that wastes time and provides no independence benefit.

# Launch ALL at the same time in a single response:

task(
  agent_type: "general-purpose",
  model: "claude-opus-4.6",
  name: "review-opus",
  prompt: "[REVIEW PROMPT + FULL DELIVERABLE]"
)

task(
  agent_type: "general-purpose", 
  model: "gpt-5.2-codex",
  name: "review-codex",
  prompt: "[SAME REVIEW PROMPT + FULL DELIVERABLE]"
)

# Optional 3rd reviewer for Tier 3+:
task(
  agent_type: "code-review",
  model: "claude-sonnet-4",
  name: "review-sonnet",
  prompt: "[SAME REVIEW PROMPT + FULL DELIVERABLE]"
)

Step 3: Collect and Deduplicate Findings

After all reviewers complete:

1. Collect all findings from each reviewer
2. Categorize by severity:
   • 🔴 Must fix — factual errors, security issues, broken logic, brand violations
   • 🟡 Should fix — quality issues, unclear language, missing context
   • 🟢 Nice to have — style preferences, minor improvements
3. Cross-reference: Issues flagged by 2+ models are HIGH CONFIDENCE real problems
4. Deduplicate: Same issue reported by multiple models = one fix
5. Discard noise: Single-model 🟢 issues that contradict the style/spec can be ignored

Step 4: Apply Fixes

1. Address ALL 🔴 must-fix issues — no exceptions
2. Address 🟡 should-fix issues unless they conflict with established style/spec
3. Consider 🟢 nice-to-have issues if they're quick wins
4. For each fix, note which reviewer(s) caught it

Step 5: Verify (Optional)

For critical deliverables (Tier 4), run a final pass:

• Single reviewer (any model) confirms all 🔴 fixes were applied correctly
• Checks that fixes didn't introduce new issues

Review Prompt Template

Adapt this template for your specific use case:

You are a [ROLE] reviewing a [DELIVERABLE_TYPE] for [CONTEXT].

Review for:
1. [DIMENSION 1] — [what to check, what good looks like]
2. [DIMENSION 2] — [what to check, what good looks like]
3. [DIMENSION 3] — [what to check, what good looks like]
4. [DIMENSION 4] — [what to check, what good looks like]
5. [DIMENSION 5] — [what to check, what good looks like]

For each issue found, provide:
- SEVERITY: 🔴 Must fix / 🟡 Should fix / 🟢 Nice to have
- LOCATION: Where in the deliverable
- ISSUE: What's wrong
- FIX: Specific suggestion

If the deliverable is clean, say so. Don't invent issues.

[REFERENCE MATERIAL — spec, style guide, brand rules if applicable]

Deliverable to review:
[FULL CONTENT]

Model Selection Strategy

Diversity matters more than "best" model

• Different model families catch different issues
• Claude excels at nuance, consistency, and factual grounding
• GPT excels at logical structure, code accuracy, and completeness
• Mixing families (Claude + GPT) gives better coverage than 2 of the same family

Current recommended pairings (as of May 2026)

Use Case	Model 1	Model 2	Model 3 (optional)
Blog articles	claude-opus-4.6	gpt-5.2-codex	—
Code changes	claude-sonnet-4	gpt-5.2	claude-opus-4.6
Architecture specs	claude-opus-4.6	gpt-5.4	claude-sonnet-4
Safety review	claude-opus-4.7	gpt-5.5	claude-opus-4.6

⚠️ Model Deprecation Awareness

• Check current model availability before launching reviews
• GPT-5.2 / GPT-5.2-Codex deprecated June 1, 2026 — migrate to GPT-5.4 after that date
• When a model is deprecated, swap to the next recommended model in the same family

Agent Type Selection

Review Scope	agent_type	Why
Content/articles	general-purpose	Full toolset for fact-checking, link verification
Code changes	code-review	Specialized for code analysis, won't modify
Specs/architecture	general-purpose	Can access codebase for context
Quick check	explore	Lightweight, faster, cheaper (Haiku)

Blog-Specific Review Dimensions

For long-form technical articles, always check these 5 dimensions:

1. FACTUAL ACCURACY — Are claims supported by linked sources? Any hallucinated facts?
2. SOURCE QUALITY — Are linked sources credible and current? Any dead or irrelevant links?
3. WRITING QUALITY — Is the voice consistent (first-person, opinionated, conversational)?
4. STRUCTURE — Does it flow logically? Is the hook strong? Is the closing substantive?
5. TECHNICAL ACCURACY — Are code examples correct? Are tool/version references accurate?

Code-Specific Review Dimensions

For code changes (Tier 3 development standard):

1. SPEC COMPLIANCE — Does the implementation match the spec in data/specs/?
2. BUGS — Logic errors, off-by-one, null handling, race conditions
3. SECURITY — Injection, auth bypass, data leaks, PII exposure
4. REGRESSIONS — Does it break existing behavior or contracts?
5. QUALITY — Naming, structure, test coverage, error handling

Anti-Patterns

• ❌ Running reviews sequentially instead of in parallel (wastes time, no benefit)
• ❌ Using only one model (misses model-specific blind spots)
• ❌ Using the same model family twice (e.g., two Claude variants — low diversity)
• ❌ Ignoring 🔴 findings from any reviewer (must be addressed or explicitly justified)
• ❌ Launching reviews without the full deliverable in the prompt (partial context = bad reviews)
• ❌ Reviewing incomplete/draft work (finish first, review second)
• ❌ Skipping review for "simple" changes that turn out to be Tier 2+ scope

Integration with Development Standards

This skill implements the Review phase of the Tier 2-4 development pipeline:

• Tier 2: 1-2 model review (blog-writer default)
• Tier 3: 3+ model review with model diversity (mandatory)
• Tier 4: 3+ model review + dedicated safety reviewer

The orchestrating agent manages the pipeline; this skill defines HOW the review phase works.