What problem does it solve? Implementation plans often contain hidden flaws, missed edge cases, or over-engineered designs that only surface during coding. This Skill sends your plan to an independent external model (Codex) for review, verifies each finding against the actual codebase, and iteratively fixes the plan until it is approved. ## Core Features & Use Cases - Iterative cross-model review loop: Sends the plan plus project rules (CLAUDE.md, .claude/rules) to Codex in headless mode, receives structured JSON findings, and repeats until approval or the iteration limit (default 5). - Verified findings, not blind trust: Every Codex issue is checked against the real code and classified as CONFIRMED (fixed in the plan) or FALSE POSITIVE (rejected with arguments sent back to the reviewer). - Over-engineering detection and confidence threshold: The reviewer flags unnecessary abstractions and scope creep, and only reports issues with confidence >= 75, filtering out stylistic nitpicks. - Use Case: Before implementing a large refactoring, run the skill on your plan file; it produces a corrected plan, a final report, and a full audit log in .codex-plan-review/. ## Quick Start Ask the AI to run /codex-plan-review on your plan file, for example: review my implementation plan at docs/refactor-plan.md with at most 5 iterations.
