What problem does it solve? Research work needs rigorous, adversarial review before submission, but manual review cycles are slow and self-review lacks independence. This Skill runs an autonomous review → fix → re-review loop with an independent reviewer from a different model family, stopping only when a policy-approved positive assessment is reached or the round limit is hit. ## Core Features & Use Cases - Autonomous Review Loop: Iterates review, fix implementation, and re-review up to 4 rounds, stopping when score >= 6/10 and verdict is "ready" or "almost". - Multiple Reviewer Backends: Routes to Copilot CLI native rubber-duck subagent, Codex MCP, manual review MCP, or compatibility copilot --agent mode, with enforced opposite-family reviewer selection for independence. - Adjustable Difficulty: Medium (default MCP review), hard (reviewer memory plus debate protocol), and nightmare (reviewer reads the repo directly via codex exec with adversarial verification). - State Persistence & Recovery: Writes REVIEW_STATE.json after each round so the loop survives context compaction and can resume mid-run. - Use Case: Before submitting a NeurIPS paper draft, run the loop so an adversarial cross-family reviewer reads your claims, code, and raw results each round while the executor implements the minimum fixes until the work passes. ## Quick Start Run the auto-review-loop skill on my current paper draft and experiment results to autonomously review and fix weaknesses until it is ready for submission.
