The decomposition round (Palantir-style)

The signature FDE round — invented at Palantir, now copied across the industry — hands you a vague, often non-technical problem ('design technology to help low-vision elderly people cook,' 'reduce a city's 911 response times') and watches how you think for 45–60 minutes. Candidate reports consistently call it the biggest filter in FDE loops, and the #1 rejection is the same everywhere: jumping to a solution.

The drill (narrate every step)

1. Clarify + scope: who is this for, what does success mean numerically, what's explicitly out of scope?
2. Users & stakeholders: who uses it daily, who sponsors it, who can block it? (The elderly-cooking question is scored on user empathy, not tech.)
3. Map the data & constraints: what data exists, how dirty is it, where does it live, what can't change?
4. Decompose into subproblems sequenced by risk × value — find the natural seams and say why you'd attack them in that order.
5. Walking skeleton: propose the thinnest end-to-end slice that proves value in week one — then iterate.
6. Failure modes + measurement: where will it break, and how will you know? Volunteer this unprompted.

• Narrate continuously — interviewers read silence as stuck. Think out loud even while you're structuring.
• Expect injected constraints mid-round ('now assume half the users have no smartphone'). Treat them as gifts: restate, adapt one component, keep moving.
• State explicit assumptions with numbers ('let's say 200 dispatchers, 5k calls/day') instead of waiting for permission.