Code health skill for Claude Code

The code health skill is AR Labs' prototype skill for Claude Code, Anthropic's command-line coding agent (an AI agent that writes code). A skill is packaged instructions, with any scripts, that an agent loads only when a task needs them; this one guides an agent through a structured code health inspection, for anyone maintaining code that agents wrote. It was built at the turn of 2026 in response to Steve Yegge's December 2025 article "Six New Tips for Better Coding With Agents". The article argues that coding agents tend to pile up code without refactoring it, so a large share of development time should go to regular code health passes, and that reviewing several times, each pass broader than the last (his "Rule of Five"), catches problems a single review misses.

The skill sets out five review passes of increasing depth and scrutiny, covering bugs, code smells, security problems and architectural issues, and writes findings to a markdown report ranked from P0 (most urgent) to P4 with an overall health score. Alongside the method, we built a small Python decision engine: it detects the languages, versions and frameworks a project uses and whether security-sensitive code is present (authentication, cryptography, SQL), and from that chooses how far the reviewing model can rely on its own knowledge and when it should check external documentation. Any security-sensitive code moves the final passes to the strictest strategy.

This is a methodology and a prototype, not a working analyser. The skill's instructions and the report format are usable as a prompt for a human-led review. The Python inspector's own findings are simulated placeholders, so it does not analyse code by itself. The project has not been developed further since early 2026.