improve

Strong models are good at understanding a codebase and deciding what to fix, but running them through every edit is expensive. This skill splits the work. The strong model audits and writes the spec, and it never changes your source code. The plans are the output.

A run maps the repo and finds its build, test and lint commands. Parallel subagents then audit nine areas, including correctness, security, performance, tests and docs. The main model re-reads every cited file and line to drop false positives. You get a ranked table and choose which findings become plans.

Each plan is a markdown file written for a weak executor. It holds file paths, current code, repo conventions, a verification command per step, and STOP conditions for when the code doesn't match. Any agent or human can pick it up. The skill can also dispatch a cheaper model in a git worktree and review its diff, but merging stays with you.