adversarial-review

name	adversarial-review
description	Cross-model code review. Sends code to a DIFFERENT model (GPT/Gemini) for independent review. Reviewer must be on a different model than the code author. Use for critical PRs and security-sensitive code.
disable-model-invocation	true
model	sonnet
effort	medium

Adversarial Cross-Model Review

Principle

Same model has same blind spots when writing AND reviewing. Different model catches what the author's model missed.

Process

1. Identify code changes for review
2. Select DIFFERENT model than what wrote the code
3. Send code + context to external model
4. Collect review findings
5. Compare with internal review
6. Address all unique findings from external model

Rules

• Reviewer MUST be different model
• Claude wrote code -> GPT reviews (or vice versa)
• Include sufficient context for reviewer
• Don't dismiss external findings without investigation