Run the prompt audit twice because one pass of /doctor prompt-audit can miss findings. The command (Claude Code 2.1.283+, alias /checkup prompt-audit) audits the instruction files Claude Code loads - CLAUDE.md, skills, subagent prompts, slash commands - for wording written for older models, such as rigid length caps, and also reports references to files or commands that don’t exist and files that contradict each other. A model does the reading, so the first report is not guaranteed complete.
Rule
- Run
/doctor prompt-auditand save the raw report before changing anything - it is your only before snapshot. - Apply the findings the repo itself can verify first: a path or command that no longer exists. These are the safest fixes.
- Run it again after the fixes. The re-scan sees the cleaned-up config and can surface what the first pass missed.
On this site’s config the first pass found 14 findings, including hard output-length caps in three subagent prompts. After the fixes, the re-scan flagged a fifteenth: the same kind of cap, ~50 lines, in a fourth subagent.
Not a CI gate
It needs the Claude Code CLI and a model call, and it suggests edits rather than blocking a merge. For rules that must hold every time, use a hook over prose; use the audit to find prose that has gone stale.
Full before/after: Claude 5.5 prompt audit: the bad lines weren’t there.