Shipping & Review
You don't get to skip review because “Claude wrote it.” AI-assisted code goes through the same pull-request process as any other code — and because AI makes huge diffs cheap to produce, the discipline of small, readable, disclosed changes matters more than ever.
The review rules
- Read every line. Don't accept a diff you haven't read. If it's too big to read, it's too big to ship.
- Small PRs beat giant PRs. AI makes a 4,000-line PR tempting; don't. If it can't be reviewed in 30 minutes, it's two PRs. The biggest review failure mode is a change so large no human can meaningfully review it — AI doesn't make that easier, it makes it more dangerous.
- AI is the first reviewer, never the only one. Paste your diff into a fresh chat: “What could go wrong here? Edge cases?” It catches a surprising amount. A human still approves. Always.
- Sensitive paths get two humans. Auth, payments, deletion: two human reviewers minimum — and AI is not one of the two.
- Disclose the AI assistance. The PR description says what was AI-assisted and how it was reviewed. That's useful context for reviewers, not a confession.
What a good AI-assisted PR looks like
- Linked to a written spec
- Architecture decisions noted in the description
- Tests written, passing, and meaningful
- AI assistance disclosed
- Reviewed by a human who actually read the code
That's the bar we hold ourselves to in OIT — and the bar we'd love to see everywhere AI-assisted software ships at Notre Dame.
Working solo?
Working solo doesn't mean skipping review. Open a PR to yourself and read the diff cold, ideally the next morning; have the AI critique its own change in a fresh chat first. For anything with real users or Internal-and-above data, borrow a second set of eyes — office hours is a good place to find them.