You told the candidate they can use any tool they want. They have Claude open. Now you are watching the recording.
What do you actually look for?
The wrong question
Most reviewers ask: did they use Claude too much?
That is the wrong question. There is no "too much" if the output is correct and the candidate owns it. The right question is whether they were directing the process, or along for the ride.
Four things to watch in the recording
How they prompt. A candidate who describes the problem clearly before asking for code understands what they need. One who pastes the task description directly and hits enter does not know what they are asking for yet.
Whether they read the output. Watch the cursor. Did they scroll through the response? Stop at a specific line? Edit before running? Or did they paste the whole thing and hope?
Whether they push back. Claude is wrong regularly, especially on project-specific context. A strong candidate catches it. They run the output, see it does not work, and tell Claude what it got wrong.
Whether they own the result. Can they explain the final code without Claude open? If not, they shipped someone else's work.
The two-interview method
One of the clearest ways to measure AI judgment directly: run the same task twice for the same candidate. One session with Claude available, one without.
The gap tells you the skill floor. A candidate who performs roughly the same in both rounds is genuinely capable. One who falls apart without Claude is a different hire, and you need to know that before day one.
EasyEnv lets you set up both sessions on the same role and compare the recordings side by side. A senior reviewer can get through both in fifteen minutes.
What does your current process tell you about how a candidate uses AI, versus whether they can work at all?