The FDE Interview
All 12 cards in this deck
The nine round archetypes and what each one actually scores — decomposition, learning, incremental coding, the defended take-home. Answer each question before you open it — then run the review above, which schedules every card with spaced repetition so it comes back right before you would forget it.
What does the decomposition round actually score?
Your clarifying questions, before you are allowed to propose anything. Jumping to a solution is the fastest fail — the round measures whether you find the real problem, not whether you can design.
What is the learning round?
You are handed docs for something unfamiliar — a bespoke library, an internal API — and asked to become useful fast. Scored on reading strategy and honest questions, not recall.
How do you read unfamiliar docs under time pressure?
Shape before detail; hunt invariants ("always", "never", "only", "before") rather than signatures; go looking deliberately for the line that breaks the obvious model; state your model out loud; name the gaps.
What is the incremental coding round?
A small problem that gains a new constraint every few minutes, each stage invalidating the last shape. It measures whether your first version can absorb the next requirement — and whether you refactor calmly.
What is the weakest possible answer in an FDE loop?
A confident wrong answer. It tells the interviewer you will do the same thing at a customer, where nobody is checking your work. "I have not understood the failure semantics yet" scores higher.
What is a defended take-home?
A take-home you then have to explain and modify live. The defence, not the submission, is the signal — they change a requirement and watch what your design does.
Why do FDE loops include a values or judgement round?
Because the job puts you alone in front of a customer with commercial pressure. At frontier labs it fails more candidates than the coding does.
What does the customer-conversation round test?
Whether you can deliver bad news, push back on scope, and keep the relationship stronger. Saying no to a request and having the customer thank you is the top-of-band answer.
How do you answer "how accurate is it?" in an interview or a demo?
Not with a number. Establish the ceiling: how often do two of their own experts agree on the same case? Then propose measuring it together and setting the target relative to that.
What is the single highest-value question in discovery?
"Walk me through what happens today, step by step." It routinely reveals that the stated problem is already solved and the real work is somewhere else entirely.
What separates a 3-signal question from a 0-signal one?
Whether the answer changes what you build. "How do you feel about AI?" has no wrong answer, so it returns nothing and costs a slot.
What do interviewers write on a learning-round scorecard?
Strategy (skimmed for shape vs read linearly), invariants (extracted rules vs memorised API), gotcha (found the buried line), honesty (flagged gaps unprompted), application (reasoned about an unmentioned case).