What you'll learn: the round types a customer-facing AI engineering loop can contain, a six-step framework for a 35-minute design question, and a rubric to score your own answer.
At a glance
Figure 1. The answer framework: six steps and one loop back
How to read this: the solid path is the order you speak in, and the dotted arrow is where you restart when a requirement changes.
- A design question tests how you reason, not whether you remember one architecture.
- Spend the first five of the 35 minutes on questions that change the design.
- Every design needs a stated trade-off, a list of ways it fails, and a way to measure it.
Scenario
Composite scenario, not a real company. Lumen Industrial Supply is a fictional distributor whose sales engineers read datasheets, FAQs and compatibility notes to answer customers. A practice interviewer says: "Design an assistant that helps Lumen's sales engineers answer product questions." You get a whiteboard and 35 minutes. The question is underspecified on purpose: what you do with the gaps is assessed first. All numbers here are illustrative.
What you will be able to do
- Name the round types a customer-facing AI engineering loop may include.
- Answer an open design question in six ordered steps in 35 minutes.
- Tie each clarifying question to the design decision it changes.
- Score an answer with the rubric.
Prerequisites
- From a vague request to measurable outcomes: the customer-side version of clarifying.
- Passing identity and Limits, state and step recording.
- The AI Architect Mindset lesson.
Core ideas
How a loop is commonly organised
Employers differ in the number, order and content of rounds. Ask the recruiter for the real format.
Figure 2. A generic loop: five round types
How to read this: the boxes are areas to prepare, not any employer's process; rounds may be merged, repeated or reordered.
| Round type | What it probes | Prepare with |
|---|---|---|
| Screening | Fit, communication | Positioning |
| Technical depth | Python, HTTP, debugging | Python and API questions |
| Design or scenario | Reasoning with constraints | This module |
| Customer-facing | Explaining, scoping, saying no | FDE scenario questions |
| Behavioral | Judgement, ownership | Behavioral questions |
The six steps and their time boxes
Name the step you are on, so the interviewer can steer you.
| Step | Question you answer | Minutes | On the board |
|---|---|---|---|
| 1 Clarify | Who uses it, what is a wrong answer worth? | 5 | Three to five answers |
| 2 Constraints | What limits the design? | 3 | A list with numbers |
| 3 Design | What is the path of one request? | 12 | One labelled diagram |
| 4 Trade-offs | What did you choose, give up, and when would you switch? | 6 | Two decisions with reasons |
| 5 Failure modes | How does it break, and what does the user see? | 5 | Three failures and responses |
| 6 Measurement | How do you know it works? | 4 | Metrics and a threshold |
Why clarify first
Predict first. The interviewer says only: "Design an assistant that helps Lumen's sales engineers answer product questions." You may ask three questions before you draw. Which three change the design the most?
Answer
- Who can see which documents? Partner-only content makes access control part of retrieval.
- What does a wrong answer cost? A wrong spec in a quote is expensive, so refusing beats guessing.
- Where do the documents live and who owns them? This decides ingestion and freshness.
"Which model should I use?" can wait: it does not change the system's shape.
A clarifying question earns its place only if you can finish "if the answer is X, I design A, otherwise B".
Say your assumptions aloud
If the interviewer does not answer, state an assumption ("I will assume 30 users and 1,000 questions a day") and move on.
Failure modes and measurement are where answers are won
Say what the user sees when retrieval finds nothing, and which number shows quality dropped. Reserve nine of the 35 minutes for steps 5 and 6.
When the interviewer changes a requirement
Return to step 2, name what changed ("partner users now also need quotes"), and revise only the affected components.
Worked example
The sample prompt, answered in outline. Assumptions: 30 sales engineers, 40 questions each per working day, so 1,200 questions a day (illustrative).
| Step | One line of the answer |
|---|---|
| Clarify | Partner-only documents exist; a wrong spec is costly; documents are files in one repository |
| Constraints | 1,200 questions a day, answers in a few seconds, access by role, every claim traceable |
| Design | Ingest with metadata, retrieve with the access filter inside the query, answer only from retrieved text, cite sources |
| Trade-offs | Keyword search first for exact part numbers, vector search if wording differs; refuse rather than guess |
| Failure modes | Nothing found, two documents disagree, superseded datasheet, injected text in a document |
| Measurement | A hand-labelled question set with answerable, unanswerable and access-control cases; zero access leaks |
Score it with the rubric reused in the next three lessons.
| Criterion | Weak | Acceptable | Strong |
|---|---|---|---|
| Clarifying | Starts drawing at once | Generic questions | Questions tied to design decisions |
| Design fit | Components, no request path | Path ignores a constraint | One request traced end to end |
| Trade-offs | Names a tool, no reason | One reason | Default, cost, switch condition |
| Failure modes | Not mentioned | Lists failures | Failures with what the user sees |
| Measurement | "We will test it" | Names a metric | Metric, test set and threshold |
Score weak 0, acceptable 1, strong 2: five criteria, total out of 10. The outline scores Clarifying 2, Design fit 1 (no number for "a few seconds"), Trade-offs 1 (no reason for refusing), Failure modes 1 (no user-visible response) and Measurement 2: 7 of 10.
Practice task
Run the framework once, alone, with a timer.
- Set a 35-minute timer. Read the prompt once and skip the worked outline.
- Write the six headings and spend the stated minutes on each, drawing one diagram in step 3.
- Score all five criteria, 0 to 2 each, and write one sentence on the lowest.
- A day later, repeat with "Design a way for Lumen's customers to request a quote by email" and compare.
At the customer: an engineer asking "how would you build this?" is running the same test. Start with their data and users, then constraints, then components.
Exercise
Explain it back. In three sentences, explain to a colleague why you ask questions before drawing and how many minutes each step gets. A good explanation gives the reason (questions change the design), the six steps in order, and the time split: about one-seventh on clarifying, a quarter on steps 5 and 6.
Done when a colleague can name the six steps back after hearing you once.
Hint 1: start with the interviewer's problem
What does the interviewer miss if you just draw?
Hint 2: pick one concrete question
Take "who can see which documents?": what changes for public-only versus mixed users?
Common mistakes
- Drawing for ten minutes before asking anything.
- Listing products instead of a request path.
- Defending the first design after a requirement changes.
- Skipping failure modes because the design "should work".
Knowledge check
1. Why is the framework ordered clarify, constraints, design rather than design first?
The constraints decide the components. Asking first is cheaper than redrawing.
2. You asked a clarifying question and the interviewer says "your call". What do you do?
State an assumption aloud and continue. It gives the interviewer something to correct.
3. The interviewer adds "and partners must be able to draft quotes". Which step do you return to?
Step 2, constraints, because a new action changes who may do what. Then revise only the affected components.
4. Which two steps are easiest to skip, and what does each add?
Failure modes and measurement. One shows what the user sees when it breaks, the other how you would find out.
5. Does this page tell you how a specific company runs its interviews?
No. It describes round types in general terms. Ask the recruiter for the real process.
You can now
- Name five round types and what each probes.
- Run the six steps in order with a time box for each.
- Turn a clarifying question into a design decision.
- Score an answer on five criteria.
Further reading
- The C4 model: four diagram levels, a vocabulary for step 3.
- Anthropic, Building effective agents.
Last verified: 2026-11-02