The natural way to open a CriticCode submission is to scroll straight to the written answers, read them carefully, and then glance at the AI transcript at the bottom of the page for about fifteen seconds before closing the tab. The instinct makes sense. The written answer looks like the real deliverable, so it gets the attention. But the transcript is usually the richer part of the artefact. The written answer is the output of the thinking. The transcript is closer to the thinking itself.

Here's a way to scan an AI conversation quickly without losing the signal.

Start at the first question

The candidate's very first message to the AI tells you more than almost anything else in the submission. You're looking for one thing: did they spend a sentence establishing context before asking, or did they paste the problem in and ask the model to solve it?

"Here's the prompt, can you implement it?" is a junior move. Not because using AI is junior, but because outsourcing the framing is. The candidate has just handed the AI the hardest part of the job (understanding what's actually being asked) and accepted whatever the model decides is relevant.

"I'm working on X. I want to ignore Y for now because it isn't load-bearing. Help me think through Z." is a senior move. The candidate has already done the framing. They're using the AI as a thinking partner, not as a judge.

You can tell most of what you need to know from the first two messages. Read them carefully.

Then scan for pushback

Scroll through the conversation looking for any moment the candidate disagrees with the AI. Not necessarily "no, you're wrong". Even a "interesting, but I think in this case…" counts. What you're checking is whether this person takes AI output as an authority or as a draft.

If the candidate never pushes back, you should be uneasy. Either the AI happened to be right about everything (unlikely in a non-trivial problem) or the candidate didn't notice the parts where it was wrong.

If the candidate pushes back frequently but the pushback is vague ("I don't think that's quite right", then immediately accepts the next thing the AI says), that's a weaker signal than pushback with reasoning ("That would break if the write volume scaled. We can't assume single-writer.").

The presence, frequency, and specificity of the pushback is arguably the cleanest proxy for engineering judgement in the whole artefact.

Look for where they stopped

The last message in the conversation tells you how the candidate decided they had enough information to answer. This is easy to overlook because you're tired by the time you get there, but it's worth fifteen seconds.

A candidate who stopped because they got to a genuinely sharp question and decided to go think about it themselves is demonstrating something you want. A candidate who stopped because the AI produced a confident-sounding summary that matched the shape of the expected answer has demonstrated the opposite. They stopped looking when the search felt complete, not when the thinking was.

Weight the chat against the written answer

On a submission where the written answer is polished but the chat is thin, you're looking at a candidate who wrote the answer first and used the AI mostly to tighten it. That's not a bad signal by itself, but it should make you lean harder on the interview to test whether the answer was theirs to begin with.

On a submission where the chat is rich and the written answer is rougher, you're looking at a candidate who was still thinking out loud when they submitted. That's often a stronger hire than it looks on the page. The thinking is the signal. Don't penalise the person whose draft didn't get as many polishing passes.

One practical rule

Before your follow-up interview, pick one exchange from the transcript (one specific question-and-answer pair) and start the conversation there. "I noticed you pushed back on the suggestion to use a queue here. Walk me through what you were worried about." That opening does more work than ten generic questions, and it lets the candidate know you actually read the thing.

The whole reason we include the transcript in CriticCode is that it's the only part of the submission that can't be rehearsed. Use it.