After using Claude to write code for a long time, there will be a bunch of description files in the project: CLAUDE.md, AGENTS.md, various rules, skills, custom commands and sub-agents. It was useful when I was writing it, but no one can tell how much of it will still be correct six months later.
Claude Code's newly added /doctor prompt-audit does this. Once you hit it, it will go through all these files and pick three types of questions:
One is the rules written for the old model. For example, the hard constraints that were originally added to prevent an old version from making mistakes are not needed in the new model, but are actually holding it back. The second is that files or commands that no longer exist are referenced. The third is that several documents are fighting with each other.
Finally, I will give you a report with suggestions for modifications. According to the Anthropic documentation, a word of the file will not be changed until you explicitly ask it to be changed.
You can also check only a certain directory, followed by the path: /doctor prompt-audit .claude/skills/deploy.
Requires Claude Code 2.1.283 or newer. Lydia Hallie, who works on Claude Code at Anthropic, came out on Thursday to remind everyone to run it through their warehouses.
This kind of "configuration rot" is a real problem, and the more serious people use it, the worse it becomes. More and more rules will only be added, and few people will go back to delete them. Over time, the model will have to read a large section of outdated instructions every time. It's not a waste to spend a minute running once.