fde
The FDE Decomposition Interview: A Framework and Four Worked Examples
How to answer a decomposition interview: a five-step framework, the rubric it is scored on, and four worked prompts including a weak and a strong answer side by side.

Prepare for the Forward Deployed Engineer loop
The usual way to fail a decomposition interview is to start drawing architecture in minute three. The prompt is vague on purpose; the interviewer wants to see you find the person who will use the output, check whether the data can support them, and scope something buildable in one to two weeks, then adapt when new facts arrive.
The framework and prompts here draw on our web research into FDE loops at eight companies, built from 23 public sources (candidate write-ups, interview guides, a Blind thread and Sierra's own engineering blog) covering 52 reported questions.
That description comes from reports of Palantir's version: a 60-minute round at a whiteboard with minimal code, where the candidate covers the end user, the data schema, the APIs and an MVP deliverable in about one to two weeks, and the interviewer injects constraints along the way (Aced). Databricks is reported to run a 60-minute decomposition round with a city-traffic prompt and an explicit request to define a KPI (Aced).
How the round is scored
The reported rubric for Palantir's decomposition round has five lines: structured breakdown, end-user anchoring, data reasoning, scoping to an MVP, and iteration with success metrics (Aced). Each step below exists to earn one of them.
| Step | What you say | Rubric line it earns |
|---|---|---|
| 1. Owner | Who acts on the output, and what decision changes? | End-user anchoring |
| 2. Data | What exists, who owns it, how fresh, what is broken? | Data reasoning |
| 3. MVP | The smallest thing that changes that decision in one to two weeks | Scoping |
| 4. Metric | One number with a baseline and an owner | Iteration and success metrics |
| 5. Risks and next phase | What could break, and what you build after the MVP works | Structured breakdown |
Spend roughly the first ten minutes on steps 1 and 2. The interviewer is reported to inject new constraints as you go (Aced), and a candidate who skipped the data questions has nowhere to put them.
Worked example: one view for a nurse
The prompt: a hospital has patient records in six disconnected systems; give a nurse one view. It appears in the techinterview.org FDE guide without a company attached. The hospital and all numbers below are made up for this example.
Owner: Which nurse, doing what? Say the interviewer answers: ward nurses at shift handover, who today spend the first part of each shift opening four of the six systems to rebuild each patient's picture. The decision that changes is what to check first for each patient.
Data: Six systems means six patient identifiers. Ask how a patient is matched across them. Suppose the answer is that two systems use the hospital's medical record number and the others use their own IDs, with name and date of birth as the only overlap. That is an entity-resolution problem, and it comes before any screen.
MVP: Join only the two systems that share the record number (say the admissions system and the medication system), for one ward, and show a handover sheet per patient: admission reason, current medications, anything given in the last eight hours. Read-only. No writes back to any clinical system.
Metric: Minutes from shift start to handover complete on that ward, baselined over the two weeks before launch, owned by the ward manager.
Risks: A wrong match shows one patient's medications on another's sheet, a patient-safety failure. So the MVP uses only exact record-number matches, and fuzzy matching for the other four systems waits for phase two with a human review queue.
The interviewer now adds: "Lab results are the thing nurses ask about most, and the lab system has no record number." A good answer does not force the lab system into week one. It says the lab feed goes into phase two with a matching review step, and that the handover sheet will show a clear "labs not yet linked" line so no nurse assumes there are none.
Three more prompts, sketched
A logistics client losing money on deliveries (Ophy). Owner: whoever controls routing and shifts. Data: margin per delivery broken down by slot, distance and failed attempts. The likely finding is that a subset of deliveries is losing money, so the MVP targets that subset. thita's free case study works this one fully.
Improve traffic in a city (Aced, Palantir; Aced, Databricks). Three possible users (control-room operators, planners, commuters) need different outputs. Pick operators, check sensor coverage before modelling, and ship congestion alerts on one well-covered corridor. The city traffic case study shows what happens when the busiest corridor turns out to have dead sensors.
Reduce 911 response times using call, traffic and ambulance GPS data (a prompt listed in Aced's FDE prep material with no company attached). The tempting answer routes ambulances city-wide. The narrower one asks the dispatch supervisor where time is lost: answering the call, assigning a unit, or driving. If assignment is slow because dispatchers cannot see which units are free, a live unit-status board is the MVP, and the metric is call-to-assignment time.
Weak and strong, side by side
Here is an illustrative script written for this page. Neither candidate is real.
Prompt: "A retailer wants to cut support-ticket volume." (A version of this appears in Educative's FDE course.)
Weak opening: "I'd build a RAG chatbot over the help centre, using a vector database and an LLM, and deflect tickets before they're created." It names a stack in the first sentence, assumes the tickets are answerable from documents, and has no owner or metric.
Strong opening: "Who owns ticket volume, and which ticket types cost the most? If I can see a month of tickets tagged by reason, I'd sort by count times handling time. If the top reason is 'where is my order', the fix might be a tracking link in the confirmation email, which ships in days and needs no model at all."
The strong candidate might still end up building a chatbot. They will build it for the ticket types where it fits, and they will know how to measure it.
When the interviewer adds a constraint
Every worked example above has a moment where the interviewer changes the facts. That moment is the test. Restate the new constraint, say which part of the plan it breaks, and change only that part. Defending the original plan costs you the iteration line, and so does throwing the whole plan out.
Practise each prompt above for 45 minutes, aloud, then ask someone to interrupt you at minute 25 with a fact that breaks your MVP. The FDE interview questions bank has more prompts; the FDE interview guide shows which companies run this round.
Start with the free FDE lessons
The FDE track on thita.ai opens with two free patterns: what a forward deployed engineer is and the decomposition framework, plus two worked case studies, improving city traffic with sensor data and a logistics client losing money on deliveries. The full track adds AI mock interviews for five FDE rounds: decomposition, customer discovery, applied-AI design, practical coding and behavioral.
Your next step
Prepare for the Forward Deployed Engineer loop
The FDE track walks through the rounds FDE loops use, decomposition, coding and customer scenarios, then lets you rehearse them out loud with an AI interviewer that asks follow-up questions.