The problem
Code written fast with an AI pair often gets reviewed by nobody. I wanted a second pass that is consistent, quick, and safe to run anywhere.
What I did
- Defined a subagent with one job: scan files for readability, performance, and best-practice issues.
- Limited it to read-only tools (Read, Grep, Glob), so it can suggest changes but never make them.
- Required the same shape for every finding: the problem, the current code, and an improved version.
- Wrote its description so Claude Code knows to reach for it right after code is written or changed.
Results
Small by design. Eight lines of configuration change how every review goes, and because it can't edit, I can run it on anything.
What I'd do differently
Give it a severity ranking and permission to say "no issues." Without one, a reviewer always finds something, and small nits crowd out real bugs.
Artifacts
The file above is the whole thing, verbatim. Below, it reviewing the script that builds this page: 14 findings, ordered by severity, and no files touched.
