AI code review: 5 failure patterns outside PR #17 and PR #13 diffs

The author's starting point was a simple observation: AI review bots produce more comments than a team can act on, so the real question becomes which comments to trust. The two pull requests used here were deliberately small. PR #17 changed only a few lines in the CI matrix, and PR #13 swapped a Dockerfile base image and added specs across two files. Even at that size, five distinct patterns appeared, which is why the author argues the issue is not repository complexity.

The patterns share a common shape. In the CI matrix case, removing the Ruby 3.5 entry would have replaced a failing check with a required check that never reports, because the job name is tied to the matrix entry and the repository ruleset requires Ruby 3.4, Ruby 3.5, and CodeQL. The author notes that reading the diff cannot reveal this, and that confirming it requires access to the ruleset itself. In the RuboCop case, the review's conclusion that rake default would fail was correct, but its stated cause was not; the author reproduced the run and found 152 offenses under RuboCop 1.92.2. These examples illustrate why the author's final checklist begins with where the evidence is fixed, not with whether the comment sounds plausible.

The author also describes turning the method on itself. AI review flagged that the original draft had attributed all five patterns to a single dependency-update pull request, when they came from two separate PRs, and that a quoted diff did not match git show output. The author corrected those points and notes that three of the first five review comments pointed out patterns the author was warning about.