What problem does it solve? AGENTS.md rule files for AI coding agents drift out of date, contain vague or contradictory rules, and fail to prevent recurring mistakes. Manually auditing and improving them is slow and subjective. ## Core Features & Use Cases - Multi-model refinement loop: Claude (Opus) evaluates and patches the file, Codex independently reviews it, then Opus performs a final coherence polish across configurable rounds. - HLD and LLD modes: Automatically detects whether the file is a root-level (high-level design) or project-level (low-level design) AGENTS.md and enforces inheritance rules between them. - Bootstrap and dry-run support: Start from a Karpathy-style baseline when no AGENTS.md exists, or preview changes with --dry-run before writing. - Use Case: Your code review loop keeps flagging the same class of issue, signaling stale rules. Run /agents-refine to score the file on six axes, apply independent critiques from two models, and produce a tightened, contradiction-free rule set. ## Quick Start Run /agents-refine in the directory containing your AGENTS.md to iteratively evaluate and improve it through the Opus-Codex-Opus refinement loop.
