Almost every hiring manager I've spoken to in the last year has told me, at some point in the conversation, that they want to hire engineers who can work with AI well. Almost none of them can point to a specific part of their interview process that would surface that ability. The want is real; the measurement is missing.
That gap isn't unusual in hiring — we often measure what's easy to measure, not what we actually care about. With AI the gap is wider than usual, because most processes weren't designed around the skill at all. The skill became important faster than most interview loops could adapt.
What doesn't count as evidence
It's worth ruling a couple of things out before looking for what to keep.
A candidate saying in an interview that they use AI daily doesn't count. Most candidates do; the claim is table-stakes and not separating. Asking "tell me about your AI workflow" will get you a polished, rehearsed answer that tells you nothing about what the person does on a Tuesday afternoon.
A candidate producing a clean artefact that was obviously AI-assisted doesn't count either. The artefact tells you nothing about the distribution of authorship: how much of it was theirs, how much was the tool's, and where the handoffs happened. You're measuring the polish of the output, which is downstream of the skill you actually care about.
What does count
The evidence you actually want looks like specific decisions the candidate made around the tool, which are hard to fake and easy to spot once you know what to look for.
Pushback moments. You want to see the candidate disagreeing with the AI's output at least once during a non-trivial exercise. Not stylistic nitpicks — actual reasoning that rejects part of what the tool produced in favour of something the candidate thought through. The specific content of the pushback is the signal.
Question quality. The phrasing of the candidate's prompts to the AI is a cleaner proxy for engineering judgement than most interview questions. A prompt that establishes context, names the constraint the candidate already noticed, and asks a specific question is very different from "here's the problem, solve it." You can tell a lot about a candidate from their first two messages.
Integration with their own thinking. You're looking for moments where the candidate holds information in their own head rather than round-tripping every decision through the AI. A candidate who names a trade-off before asking the model about it is showing you something the model can't do for them.
Override decisions. If the candidate ever writes down "the AI suggested X; I didn't do that because Y," you've got the cleanest evidence possible. Cheaper to generate than most teams think, and it almost always distinguishes the stronger hires.
What to redesign to get that evidence
If your current exercise doesn't produce any of the above, the shape is wrong, not the candidate. Two small changes tend to matter more than any individual tweak.
First, make the exercise longer than the AI's comfort zone. A problem that fits cleanly inside the AI's first attempt won't surface override moments, because the AI will seem right. The exercise needs enough depth that the candidate has to catch the tool in at least one mistake — or at least in one place where the tool's output was plausible but wrong for the actual constraints.
Second, collect the reasoning, not just the artefact. That can mean asking the candidate to write down where they disagreed with the tool. It can also mean capturing the full AI transcript so you can read it yourself. Reading the transcript usually takes less time than grading the artefact, and it produces richer signal by a wide margin.
At CriticCode this is exactly the shape: the candidate works with an in-page AI, writes their reasoning in their own words, and the hiring manager opens the submission with the whole package in view. The point isn't the tool specifically — it's that unless your process collects the conversation, you're measuring output without authorship.
The audit question
If you want a single question to run against your current process: if a candidate was exceptional at working with AI, would your current interview produce evidence of that fact I could hand to someone else? If the honest answer is "probably not," you have your to-do list. The fix is rarely a new interview round. It's usually adjusting the one you already run so the behaviours you care about have somewhere to show up.