The human review that belongs in an AI-assisted workflow.
Start with the expected behavior.
Before reviewing an implementation, agree on what the user should be able to do. Walk through the important journey from beginning to end. A page can look finished while its most important action still fails.
Question the assumptions.
Check where information comes from, what a user is allowed to see, and what happens when data is missing. Generated code can make plausible assumptions that do not match the business. Someone has to notice and resolve them.
Try the failure path.
What happens when an integration is unavailable, an input is incomplete, or a session expires? A useful product explains what happened and gives the user a way forward. Review those states alongside the happy path.
Review the experience, too.
Use the product on a small screen and with a keyboard. Read the labels out loud. Check whether a person can understand the page without knowing how it was built. Technical correctness and a usable experience need separate attention.
Make release ownership explicit.
Agree on who approves the work, how the release is checked, and what the handover includes. Record known limitations and next steps. AI tools can support the process; they do not remove the need for someone to own the outcome.
