How do you catch subtle errors in AI documentation before they confuse your team?

I spent hours checking a PR's actual logic recently but completely skipped reading the agent-generated setup notes. Turns out the doc had a wrong step order and it ended up burning half a day for a teammate trying to run the project locally.

Bad documentation feels way harder to catch than buggy code because everything looks fine until someone actually tries following the steps.

Are you guys auditing generated markdown files as strictly as code, or did you decide to ban AI docs altogether?