reflexion:critique

Multi-agent debate improves code reviews by spawning specialized judges for requirements, architecture, and quality. Each judge uses chain-of-verification to self-check findings, debates disagreements, and synthesizes a final consensus report with prioritized action items to reduce individual bias.

To run a multi-judge consensus critique on a pull request, request a review of the specified files or commit range. The skill coordinates parallel judge agents to analyze correctness, debate findings, and generate a structured consensus report with scores and prioritized improvements.

Yes, you can use chain-of-verification to review design documents and commits. The multi-judge critique process is applicable to conversation-driven implementations, pull requests, and larger releases, generating evidence-backed critiques across various software engineering artifacts.

The best way to identify architectural concerns and correctness gaps is using a multi-perspective review with specialized judges. This approach surfaces requirement alignment issues and actionable improvements while producing an evidence-based consensus critique with prioritized action items.

Yes, multi-agent consensus building works for reviewing both small features and larger releases. The critique process orchestrates independent judges to evaluate quality and correctness across varying scopes, synthesizing reasonable disagreements into a unified, actionable report.