Saved in this browser - sign in to keep your review list.
How should your speech become text?
Listening… your words appear above as you speak - tap Stop when you're done.
Recording · cr - tap Stop & transcribe when you're done.
Transcribing with AI…
Voice:
Keep going - a few more words and AI can grade it.
Interview lens
How interviewers actually use this question
Likely follow-ups
How small should one batch be, and how do you know it's ready to review?Small enough to read the complete diff in one pass, one concern per change, and focused checks passing before the next batch starts.
The output is wrong. How do you decide between re-prompting and changing the spec or context?Find the cause first, such as missing context, a vague goal or a wrong assumption. Fix that input instead of re-rolling the same prompt.
Which agent actions should need your approval before they run?Anything hard to undo or outward-facing, like deletes, migrations, pushes, deploys and access to secrets. Mention scoped permissions.
Is there any setting where a looser, accept-and-go style is fine?Throwaway prototypes and spikes where nothing ships. Say where the line is: once code is kept or shared, the full workflow applies.
What you can say
Prompt-and-pray is sending a vague request, accepting whatever large result comes back, and hoping it works, with no control over scope and no evidence that it's right.
The healthier version starts before the prompt. I write down the goal and the acceptance criteria, then add the relevant context, the constraints, and what's out of scope.
For anything nontrivial I ask for a plan first and correct its wrong assumptions before any code gets written.
Then I work in small batches: one reviewable change, a read of the complete diff, and focused checks before I continue, with approval required for risky commands.
When the result is wrong, I fix the spec or the context instead of re-rolling the same prompt.
The assistant drafts and explains, and I own direction, permissions, and what counts as done. It's slower per step than one big prompt, but I can defend every change I accept.
Weak answers to avoid
Stops at 'write a better prompt'Wording is one step. The interviewer wants the loop: acceptance criteria, a plan, small batches, a full diff read, and checks before continuing.
Accepts one large diff at the endNobody can review a huge diff properly. Say you work in small batches, read each complete diff, and run checks before starting the next one.
Trusts green checks without reading the diffChecks are one piece of evidence. Read the complete diff too, because passing checks won't flag changes nobody asked for.
Lets the agent own direction and permissionsThat gives control to the tool. The engineer sets direction, grants permissions, approves risky commands, and decides what evidence is enough to accept.
Interview lens
Likely follow-ups, what you can say, and the weak answers to avoid.
