Four areas are scored, and how you use the AI runs through all of them.
The four areas#
| Area | What earns it | What loses it |
|---|---|---|
| Problem solving | You understand the problem, break it into parts, and name the approach and its complexity before prompting | Prompting first and thinking later, or letting the AI choose the approach |
| Code quality and understanding | Correct, readable code that fits the existing codebase. You can explain any line, including the ones the AI wrote | Pasting big chunks you can't explain when the interviewer points at a line and asks "why this?" |
| Verification | Running tests often, adding edge-case tests, catching the AI's mistakes, checking claims about complexity | Accepting code that looks right without reading it, or treating passing visible tests as done |
| Communication | Narrating your decisions, saying why you accepted, changed, or rejected AI output, and taking hints well | Long silences, or narrating what you're typing instead of why |
Verification separates candidates most. Everyone has the same AI. The difference is who notices when it's wrong. Accepting a plausible but wrong suggestion is the fastest way to fail this round.
How AI use is judged#
There's no correct amount of AI use. Writing something by hand and having the AI write it are both fine, as long as you can say why. Be ready to answer both "why didn't you use the AI here?" and "why did you trust what it wrote?"
Prompting is graded as specification, not as technique. A prompt that states the inputs, outputs, constraints, and what not to touch shows you understand the problem. Clever prompting tricks score nothing.
Common ways candidates go wrong#
Three patterns cost candidates the most:
- Showing off the tool. Spending time on fancy prompting or tool features instead of the problem.
- Racing through features. Piling on functionality while correctness slips.
- Going hands-off. Letting the AI drive while you just watch.