I want to see what you do when you don't know
What I look for in AI-assisted interviews
AI has changed what a coding exercise tells me about a candidate. When a model can help produce the solution and explain it, a polished submission on its own doesn’t tell me how much the candidate understands. I want to see how they make decisions and work through questions about it.
I encourage candidates to use all the resources available to them. Read the docs, search, use AI. Then let’s talk through the work, including the parts you haven’t figured out yet.
For about an hour, one candidate and I went through his take-home exercise with an AI agent. He was asking questions, picking things up quickly, and jamming with AI to understand the problem. It felt like teamwork between the three of us.
The exercise was to build a CLI that called an AI model and extracted structured data from medical-history PDFs. We gave candidates a 24-hour window after presenting it.
The PDFs were pretty crappy and ambiguous. Each represented an occurrence in a patient’s history, and the candidate had to use their judgment to deal with the limitations of the source material. The output needed to be structured in FHIR, a standard for exchanging healthcare data.
There were what I’d call LeetCode-lite elements in it. Each clinical impression linked back to the previous one, forming a linked list through the patient’s history. It contained a summary of that occurrence and a summary of the history so far. Sort of like a ledger.
The documents were ingested one at a time, out of chronological order. The exercise specified an insertion algorithm that candidates were supposed to preserve.
When we discussed performance and production readiness, this candidate considered adding other insertion mechanisms while keeping the required one. As far as I remember, nobody else had explored that possibility. I hadn’t asked for alternatives. He recognized that the requirement to preserve one algorithm still left room for other ways of adding records. I loved that bit of lateral thinking.
His repository was also a delight to review. The documentation was thoughtful, it was easy to find my way around, and getting set up was straightforward. He’d even made a small UI to help make the solution understandable. I could trace his thinking by looking through what he’d put together.
What impressed me was how useful and thoughtful the additions were. Doing more wasn’t the point. These choices made it easier for me to understand his work. That’s what I appreciated about his taste and his empathy for the reviewer, who in this case was me.
In the discussion, he kept asking questions and working through the exercise with me and the agent. He grasped things quickly and cared about understanding what we were discussing. I came away feeling that he was using AI to support his thinking, not to replace it.
I could follow his reasoning in the repository and discuss alternatives with him. I could also participate as he worked to understand the problem. Those were the parts of the exercise that helped me form a view of him beyond the finished submission.
With AI available, I want an interview to give me that opportunity. I want to see how someone uses the help, makes sense of the answers, and works with another person when the path isn’t already familiar. The candidate’s questions are part of what I want to hear.
If you’re preparing for an interview where AI is welcome, use it to explore what you don’t know. Follow up on an explanation until you can explain the idea yourself. Be ready to discuss why you chose an approach and what you would change, and make that reasoning easy to follow in your submission.
Come ready to learn in the conversation, too. Ask questions about the parts you haven’t figured out, and let the interviewer work through them with you. Give them a chance to see what thinking through a problem with you is like.
This essay expands on my remarks in How to Scale High-Performing Engineering Teams in 2025, published by Tech Teams Today by Revelo on May 20, 2025. The full recording and its transcript are available on YouTube.