OpenAI and Anthropic spend an hour on one project you built. Here's that hour.

15 of 34 AI companies, OpenAI and Anthropic among them, give a round to one project you built. What the round is, how the hour goes, and the questions to expect.

Close-up of the gears inside a mechanical watch movement
Photo: Joey Zhou on Unsplash
  • 15 of the 34 AI companies we map spend a round on one project you built, Anthropic, OpenAI, SpaceXAI (formerly xAI), Google DeepMind, Cursor and Microsoft AI among them. Big tech and scaled startups mostly don’t: 5 of 14.
  • It is usually an hour: 6 of the 8 with a known length run 60 minutes.
  • 8 of 15 open with you presenting; the rest are a conversation from the first minute. Both end in follow-ups.
  • The follow-ups ask about you, not the diagram: your decisions, your share of the work, the numbers, and what you would change.

15 of 34 AI companies give one round to one project you built

A project deep dive is an interview about something you already built. You pick the project. The interviewer knows nothing about it except what you tell them, and spends the round finding out how much of it you really understand and how much of it was really yours.

Companies call it different things: project presentation, project deep dive, project walkthrough, agent presentation, take-home review, résumé deep dive, tech talk. At Anthropic you present a project you owned end to end. At OpenAI you present one or two projects you led. At Cognition it is a conversation about a system you built, where it broke and what you would change.

6 of 8 run an hour

AI company deep dives by length, where it is known (n=8).

  1. 60 minutes6 loops
  2. 45 minutes1 loop
  3. 30 minutes1 loop

An hour on one project is long. Describing the system takes a small part of it. The rest is follow-ups, and that is where the round is decided.

The hour has two parts: 8 of 15 start with your presentation

First, the story. Where you present, plan on about 20 minutes, as at Anthropic: the problem, why it mattered, the design, and what happened. Where it is a conversation, the interviewer asks you to describe the project and you have a few minutes before the questions start.

Then, the follow-ups. The interviewer picks a thread and pulls it. A good interviewer keeps going until you run out of answers, because that point tells them where your knowledge ends. Nobody expects you to know everything. They do expect you to know the parts you said were yours.

The questions to expect, and what each one checks

6 of these 7 were asked, in this order, in one candidate’s project deep dive at an AI company this fall. We added the simplification question; be ready for it too.

  1. What were the technical constraints? Whether you know why the obvious design didn't work for this system.
  2. What was the most challenging technical part? A decision you made between real options, and what it cost. Not a problem you ran into.
  3. How big was the team, and how long did it take? Your share of the work, so they can weigh everything you said before.
  4. What was the business impact? Numbers. What shipped, what moved, and what you did about what didn't.
  5. How do you know it works? How you tested it and how you measure success, edge cases included.
  6. If you had to simplify the system, which component would you remove? Whether you know which parts carry their weight, and what the system would lose without each one.
  7. If you could redo one thing, what would it be? One real moment where a choice hurt, why it made sense at the time, and what you changed after it.

Notice what isn’t there: “walk me through the architecture again.” You covered that in your walkthrough. The rest of the hour asks for decisions, numbers and judgment. If your answer to the hardest part is a problem you ran into, say what you chose between, what you picked, and what it cost you.

It is not system design, and it is not behavioral

In a system design round the interviewer gives you a prompt and you design something new on the spot. In a deep dive the system already exists and you are the only expert in the room, so a vague answer is easy to spot. In a behavioral round you tell stories about people and conflict. A deep dive is technical, and the story only works when the details hold up.

Some loops mix them. Sierra splits the hour between your take-home agent and a technical topic, and at Hugging Face you defend your take-home the way a maintainer would review it.

What this means for your prep: a diagram and a 5 to 7 minute walkthrough

Prepare one walkthrough of the high-level architecture that takes 5 to 7 minutes, then have each of these ready for the follow-ups.

  1. Draw the architecture diagram. One page: the main components, how data moves between them, and the part you built marked. You will draw it again live, so practice drawing it, not only showing it.
  2. Why this project. The problem, who had it, and why it was worth solving then. One or two sentences before any boxes.
  3. Your own contribution. List what you did yourself, apart from what the team did. Say "I" for your part and "the team" for the rest.
  4. The impact. What shipped and the numbers it moved: users, latency, cost, revenue, time saved. Lead with them.
  5. The technical challenge. The hardest part, told as a decision: the options, the one you picked, why, and what it cost.
  6. A design decision others pushed back on. Who disagreed and why, how you settled it (data, a prototype, a compromise), and whether you were right.
  7. What you would remove, and what you would redo. The one component you would cut to simplify the system, and what it would cost. The one decision you would make differently, and why it made sense back then.
  8. Say it out loud. Answer the 7 questions above, each in under 90 seconds, then stop and offer more detail.

For a full week of prep, day by day, read how to go deep in the project round. For which companies run the round and what each grades, see the AI labs that grill you on one project.

See every reported question for your loop, by round, with when it was last asked.

Pick your company

What the labs asked this week

A weekly email of what OpenAI, Anthropic and the other top AI labs asked engineers, round by round.