Lab L01 · Engagement
Decomposition Drill
- MVP set
- ≈ 25 min
- Interactive version coming
The customer’s campus is the canvas. Each branch of your tree lights up the building, team and system it touches. Unexplored areas stay dark.
What you’ll practise
- Turning a one-line executive mandate into a decomposition tree under time pressure
- Separating tested facts from assumptions you still have to check
- Picking a first slice that is small enough to ship and big enough to matter
In the book
Practice pack · Chapter 4
Web companion to Chapter 4 of Fieldcraft. Free, no login. Work through it with a timer and the chapter open.
Chapter 4 practice pack: Decomposition
This is the practice companion to Chapter 4 of FIELDCRAFT. The chapter teaches the method; this page is where you drill it. You’ll find the skip-test with model answers, a timed 30-minute drill on a new account (Fernhollow Freight), a minute budget for decomposing out loud, an exercise to run with a peer, and how all of this shows up in the FDE decomposition interview. The chapter’s own exercise (turning assumptions into questions) and the checklist stay in the book.
Quarrybrook Health Alliance (QHA), Kestrune Labs, Fernhollow Freight and all characters are fictional. Numbers from the QHA report are illustrative.
Skip-test, with model answers
Answer all three without looking, then check. If you can, skip to Field procedure in the chapter.
-
In “Use AI to cut prior-authorization turnaround from 9 days to 2”, which words state the outcome and which propose a means? The outcome is “cut prior-authorization turnaround from 9 days to 2”: a direction, a baseline of unknown definition and a target. “Use AI” is a proposed means. It returns in layer 6 (System), if at all.
-
A report gives the mean days a request spends in each status. Why can you add those to get mean turnaround, and why can you not do the same with medians? The mean of a sum is the sum of the means, so means add across sequential steps. Medians do not. At QHA, 54% of requests are never pended, so the median of the chase term C is 0. Sum the medians and the largest term in the budget disappears. Budget with means, then report the median and the 90th percentile (p90) beside them.
-
Name the four things every first slice has. What do you do with a high-value branch whose data you cannot reach this month? One named user group, one real task done end to end, real data, and a measured baseline. A high-value branch with unreachable data is a discovery priority, not a slice, unless the data becomes reachable at a contract milestone that has an owner and a date. Then it is a candidate with that milestone as a named dependency, and until the milestone lands you build only the parts that need no restricted data (Chapter 4, Table 4.3).
Decomposing live, out loud
With a customer in the room, or an interviewer, you do the same eight steps in front of people, and the hard part is time. The table below is a budget for a 30-minute pass. Say which layer you are on as you enter it, so the people listening can follow and correct you.
A 30-minute pass
| Minutes | Layer | What you say or ask |
|---|---|---|
| 0–3 | Outcome | Restate the ask as a number. Ask which statistic, from what event to what event, who owns it, and what limits apply |
| 3–7 | Users | “Who touches a request today, and how many of them are there?” |
| 7–13 | Workflows | Draw the steps as verbs. Ask where it waits longest |
| 13–17 | Data | For the two slowest steps: where does the input come from, and what does it look like? |
| 17–20 | Decisions | Mark the judgments. Say aloud which ones stay human |
| 20–23 | System | Hang two or three boxes on the branches. No vendor names |
| 23–26 | Risks | One risk per main branch |
| 26–30 | First slice | Name it, say why it is first, and name the first assumption you would test |
Three habits keep the pass on track. Write the outcome line where everyone can see it and point at it when the conversation drifts. Keep a “parking” box in the corner of the board: when someone names a solution early, write it there, say “we’ll come back to that in the system layer”, and return to the layer you were on. And ask one question per layer, not five; a live decomposition is a draft, and the dashed lines are where the rest of your questions go.
Exercise 1 · Timed: Fernhollow Freight (30 minutes, out loud)
Carla Benning, vice president of operations at Fernhollow Freight, sends one line: “Use AI to cut carrier-invoice disputes in half.” You know four things: Fernhollow receives about 14,000 carrier invoices a month; 9% are disputed; a dispute takes 23 days to resolve on average; an accounts-payable team of 11 handles them. Invoices arrive as electronic data feeds and PDFs, and rate agreements live partly in a contract system and partly in spreadsheets.
Set a timer, record yourself, and follow the 30-minute budget above. Produce one page:
- the outcome line, with your question about what “in half” means (count, dollars or days?);
- any constraint you would ask about (carrier contracts often set dispute windows);
- the tree, with dashed branches;
- a first slice with its reason;
- three assumptions, each with a test.
Rubric G4.1 (Appendix G). If you are on the interview reading path, do this twice, a few days apart.
Exercise 2 · Assumptions into questions
This exercise is printed in the chapter (Practice, rubric G4.2).
Exercise 3 · Defend a different slice
Swap your Fernhollow tree with a peer’s. Choose their first slice, then argue for a different one using Table 4.3 (the first-slice decision table) and an arithmetic budget for their outcome. You win if they change their mind or you change yours, with a reason. Rubric G4.3.
Interview angle
One interview-preparation guide describes the FDE decomposition round as an open-ended session on a deliberately vague, real-world problem, typically 45–60 minutes on a whiteboard or shared editor (Aced, formerly Exponent, “How to answer decomposition interview questions: the definitive guide (2026)”, https://www.aced.io/blog/how-to-answer-decomposition-interview-questions-the-definitive-guide-2026, accessed 2026-11-11). The same guide lists seven signals interviewers score, among them clarifying requirements before proposing solutions, making assumptions explicit, reasoning about trade-offs and considering the end user. The eight layers map onto most of those signals. Two need practice on top of the layers: naming edge cases and failure modes (push the risks layer past one row per branch), and communicating under pressure (the timed pass above).
Sample prompt: “A city wants to use AI to reduce emergency-room wait times. Go.”
A strong answer asks within the first 2 minutes which wait (door to triage, triage to clinician, or decision to admitted bed) and which statistic; asks whether any wait has a mandated limit; budgets the total into its terms before naming any technology; names roles with rough counts; says aloud which decisions stay with clinicians; and ends with one first slice, the arithmetic for what it can move, and the first assumption it would test. A weak answer starts with “I’d train a model to predict arrivals” and is still describing the model at minute 20.
Appendix E of the book has 30 decomposition prompts; Chapter 33 covers the full interview loop.