The 30 second answer
I treat the AI as a fast pair, not the author. Before each prompt I say what I'm asking for and what I expect back, then I read the generated code line by line and check it against the requirements. I write and run tests for every part, and when the output is wrong I explain why and fix it. The interviewer is watching my judgement, not my typing speed.
The reported candidate was unsure how the round was scored, but finished every part by doing three things consistently. They explained the intent behind each prompt and what they expected the tool to produce. They reviewed the generated output instead of accepting it. They wrote and ran tests to verify correctness. Their warning was not to trust the agent blindly, because it makes mistakes.
A useful habit is to keep prompts small and scoped to one requirement at a time, so each generated change is easy to review and easy to test. Name edge cases yourself rather than waiting for the tool to find them, and when the tool is wrong, say what was wrong and why. That spoken reasoning is the evidence of engineering judgement the interviewer can observe.