Module 5 · Lesson 1 · Interview Craft · 26 min

The First 5 Minutes: How Staff Candidates Open Differently

The first 5 minutes set the floor for your final score. The Opening Triplet is the three-move choreography Staff candidates execute by reflex: name a framework, surface the highest-leverage constraint, commit under uncertainty. Done in 5 minutes, those three moves score three of the seven signals from Lesson 1.1 before the design even begins.

Watch any Staff candidate who consistently scores at level and you'll see the same choreography in the first 5 minutes. They name a framework within 15 seconds, ask the one highest-leverage constraint question within 2 minutes, and make their first explicit commitment within 4 minutes. By minute 5, before any architecture is drawn, the interviewer has formed an impression of 'this candidate operates at level.' That impression doesn't reverse easily for the remaining 40 minutes; it sets the floor.

Watch any Senior candidate at the edge of L6 and you'll see the inverse. They start with clarification questions that don't telegraph a framework, they ask QPS before asking the constraint that actually shapes the design, they describe options without committing. By minute 5 the interviewer is still uncertain about the candidate's level — and the remaining 40 minutes have to lift the candidate out of that uncertainty, which is much harder than starting at the right altitude.

The Opening Triplet is the choreography that closes the gap. It is not a script; it is three specific moves with specific timing that, executed deliberately, signal level to the interviewer in the window where impressions are formed. The content varies by prompt; the structure is the same.

Framework

The Opening Triplet

The first 5 minutes of a Staff design interview decide the floor of your final score. Three moves, executed in the first 90 seconds, separate Senior candidates from Staff candidates in the interviewer's notes before any technical content has been discussed. The Opening Triplet is the structured choreography of those three moves: a framework name, a constraint surfacing, and a commitment under uncertainty. Each move is small. Together they buy the interviewer's patience for the next 40 minutes and frame everything that follows.

1
Move 1 — Name a framework (the first 15 seconds)
Within the first 15 seconds of the interviewer finishing the prompt, name the framework you'll use. 'Before I sketch anything, I want to walk through CLARO — constraints, latency, access, read/write, objective.' This buys you patience to spend 5-7 minutes not drawing. Without the framework name, the same 5-7 minutes of clarification feels like stalling.
2
Move 2 — Surface the highest-leverage constraint (~minute 2)
Within the first 2 minutes, ask the one question whose answer reshapes the entire system. For content moderation, it's pre-publish vs post-publish. For RAG, it's how often the corpus updates. For coding assistants, it's privacy tenancy. Surfacing the highest-leverage constraint first signals you understand which design space the prompt actually lives in.
3
Move 3 — Commit to a position with the trade-off named (~minute 4)
Within the first 4 minutes, make your first explicit commitment with the trade-off named. 'I'll plan for a 170ms inference budget on the sync path — that's tight for any large model, so I'm planning for a cascaded classifier. If the SLA tightens further, the cascade becomes a single distilled model.' The combination of commitment plus the conditional revisits-rule is the Signal 1 + Signal 2 hit from Lesson 1.1, executed early and visible.
4
The cumulative effect
By minute 5, the interviewer has graded three of the seven signals from Lesson 1.1's rubric. They have signaled patience, identified the design space, and made a defensible commitment. The remaining 40 minutes are run with the interviewer's confidence in their pocket. Without the Triplet, the same technical content scores lower because the interviewer is still uncertain about candidate level when the design starts.

When to use

Run the Opening Triplet at the start of every Staff design interview. The choreography is the same regardless of the prompt; only the content of each move varies. The Triplet is also the right structure for the first 5 minutes of an architecture review at work, a tech-spec presentation, or any context where you're being evaluated for technical leadership in compressed time.

Worked example

Prompt: 'Design a fraud detection system.' Triplet executed: Move 1, '~12 seconds in: I want to run CLARO before I sketch — five minutes on constraints, latency, access, read/write, objective.' Move 2, '~2 minutes in: The highest-leverage question is whether this is inline with payment auth or post-auth — those have completely different SLAs and architectures.' Move 3, '~4 minutes in: Inline at 50 ms p99 forces a distilled-student model with co-located features; I'll commit to that architecture, with the teacher model serving the async escalation path. If the inline SLA relaxes, we can revisit and use the teacher directly.' Five minutes; three signals scored; remaining 40 minutes done with the interviewer's confidence.

Calibration ladder

The interviewer says: 'Design a system that summarizes long meetings into action items.' Your first 30 seconds.

Opening-move probe. The interviewer is watching for the framework name in the first 15 seconds.

L4 · Mid

Sure. So we'd need to record the meeting, transcribe it, then summarize it with an LLM, and probably some kind of UI to show the action items. What scale are we thinking?

Missed: Started drawing in words. No framework, no constraint surfacing. The interviewer's first impression: implementation-focused, not architecture-focused.

L5 · Senior

Before I jump in, let me clarify a few things. What's the scale — how many meetings per day? Are these video calls or audio only? Do we have transcription already, or are we building that too? What's the SLA — should action items appear within seconds of the meeting ending, or is overnight fine?

Missed: Asked the right kind of questions but didn't name a framework or identify the highest-leverage one. Will get to a good design slowly.

L6 · Staff

Before I sketch, I want to run through five things — constraints, latency, access patterns, read/write shape, and objective. Should take about five minutes. The single highest-leverage question I want to ask first: is this real-time summary surfaced during the meeting, or post-meeting only? Those are completely different systems with completely different SLAs.

Missed: Named the framework, identified the highest-leverage constraint. Strong opening. Missing only the explicit commitment-promise.

L7 · Principal

Before I sketch, I'm going to run CLARO — five minutes on constraints, latency, access, read/write, objective. Two questions to anchor the constraints first: real-time-during-meeting versus post-meeting only, because those are two completely different products with two different SLAs and two different privacy postures. And: is the action-item extraction the primary value or is the full summary the primary value? Those map to different model architectures. Once those are anchored I'll be able to commit to a system shape inside the next two minutes.

What scored L7

Named the framework AND identified two constraint questions that anchor the design space AND telegraphed the commitment ('I'll be able to commit to a system shape inside the next two minutes'). The commitment-promise itself is a signal — it tells the interviewer the candidate's pace and confidence before any commitment is actually made. By minute 1, the interviewer has three of the seven signals from Lesson 1.1 already scored. The remaining 44 minutes are run at level.

Simulated interview

The same prompt — 'design a long-meeting summarizer' — handled by a candidate executing the Opening Triplet deliberately. Three turns, five minutes elapsed, no architecture drawn yet.

Interviewer

Design a system that summarizes long meetings into action items. Whatever scale you want to assume.

Candidate

Before I sketch I'm going to run through five things — constraints, latency, access patterns, read/write shape, and the objective. About five minutes. The most important question I want to anchor first: is this surfaced during the meeting in real time, or post-meeting only? Those are completely different products.

Interviewer

Post-meeting only. The action items appear within a few minutes of the meeting ending, surfaced in the meeting recap.

Candidate

Got it — post-meeting async, few-minute SLA. That means I can use a richer model and a streaming pipeline, not an inline LLM call. I'll commit to that architecture and we can revisit if the SLA tightens. Now I want to ask about scale, multi-tenancy posture, and what the primary measurement is — action-item recall, summary faithfulness, or both. Two more minutes on these, then I'll draw.

Outcome

Three turns, ~2 minutes, framework named, constraint anchored, commitment made, pace telegraphed. The interviewer has formed an at-level impression before any architecture is on the board. The Triplet is the choreography that produces this impression deliberately rather than hoping it emerges from technical depth alone.

Pattern recognition

When you see

You feel pressure to start drawing in the first 60 seconds of a design interview.

→

Think

Resist. The pressure is the test. The Triplet buys the interviewer's patience for the 5-minute opening; without it, the 5 minutes feel like stalling.

Most candidates feel an implicit pressure to demonstrate competence by producing visible output (drawing, coding) quickly. In a Staff design interview, the opposite is true — Staff candidates demonstrate competence by spending the first 5 minutes structuring the problem before producing any output. The Triplet is the explicit choreography that makes that 5 minutes feel deliberate rather than slow.

Drill · 10 minutes

Practice this. Time yourself.

You have 10 minutes. For each of these three design prompts, write your first three turns (framework name, constraint surfacing, commitment) in the form they would actually come out of your mouth. Time each set to ~90 seconds elapsed. (a) 'Design a real-time fraud detection system.' (b) 'Design a RAG system over our company docs.' (c) 'Design an LLM-powered customer service agent.'

Self-assessment rubric

Dimension	Weak	Passing	Strong	Staff bar
Framework named in first 15 seconds	Did not name a framework.	Named CLARO or equivalent.	Named the framework with the time-promise ('about 5 minutes').	Named the framework AND signaled it's not stalling ('to make sure the architecture I draw fits, not the one I'm familiar with').
Highest-leverage constraint surfaced	Asked about QPS or scale first.	Asked one shape-of-design question (sync vs async, pre-publish vs post-publish).	Same plus: named why that question shapes the design ('those are completely different systems').	Same plus: identified the prompt-specific highest-leverage question correctly across all three drills (fraud: inline vs offline; RAG: corpus freshness or scope; CS agent: action-surface and reversibility from the Bounded Autonomy lesson).
Commitment with revisits-rule	Asked more questions instead of committing.	Made a commitment.	Commitment plus 'we can revisit if X' framing.	Same plus: telegraphed pace ('two more minutes on these, then I'll draw'). Demonstrated control of the interview's rhythm.

Reveal model solution

(a) Fraud detection. Move 1: 'Before I sketch, I want to walk through CLARO — constraints, latency, access, read/write, objective. About five minutes; I want to make sure the architecture I draw is the one that fits, not the one I'm familiar with.' Move 2: 'The single most important question: is this inline with payment authorization, or post-auth? Inline at sub-50 ms is a fundamentally different architecture from post-auth at seconds.' Move 3 (after answer): 'OK, inline. That forces a distilled-student model with co-located features and a sidecar feature store. I'll commit to that architecture. If the SLA relaxes, we can revisit and use a richer model directly. Two more minutes on the access pattern, the read/write shape, and the objective, then I'll draw.' (b) RAG over company docs. Move 1: 'Before I sketch, I want to walk through CLARO — about five minutes.' Move 2: 'Highest-leverage question: how often does the corpus update — daily, weekly, monthly — and what's the team's actual problem: retrieval (finding the right doc) or trust (the docs are stale and engineers don't trust answers)? Those are completely different systems.' Move 3 (after answer): 'Got it — corpus updates daily, primary issue is trust. That means the design isn't just a RAG pipeline; it's a corpus-confidence system with retrieval attached. I'll commit to that framing. Two more minutes on the access pattern, the scale, and the eval set, then I'll sketch.' (c) LLM-powered customer service agent. Move 1: 'Before I sketch, I want to walk through CLARO, but I also want to run through the Bounded Autonomy dimensions — action surface, step budget, reversibility tier, confidence threshold, recovery envelope. About six or seven minutes total.' Move 2: 'Highest-leverage question first: which of the agent's actions are reversible (look up an account) versus irreversible (issue a refund, escalate to a human, mark account closed)? The reversibility tier writes the architecture more than anything else.' Move 3 (after answer): 'OK — refunds and account changes are irreversible; everything else is reversible. I'll commit to a design where irreversible actions are human-gated for the first six months regardless of confidence, and reversible actions can be autonomous within the step budget. Five more minutes on the rest of the framework, then I'll draw.'

Common failures

✗Forgot to name the framework. The 'before I sketch I'll run through five things' line is the load-bearing sentence; without it, the constraint surfacing feels like stalling.
✗Asked about QPS first. QPS is a useful question but not the highest-leverage one for these prompts.
✗Committed without the revisits-rule. The 'we can revisit if X' framing is what makes the commitment defensible — without it, the commitment can be picked apart.
✗Did not telegraph pace. 'Two more minutes on X, then I'll draw' is a pace-control move that tells the interviewer you're in command of the time budget.

Artifact · reference card

The Opening Triplet Wallet Card

Move 1 — Framework name (~15 seconds in)

Reflex sentence: 'Before I sketch I want to walk through [CLARO / TRACK / RGSA / etc.] — about [N] minutes.'
Why it works: Buys the interviewer's patience. The framework name signals you're structuring, not stalling.

Move 2 — Highest-leverage constraint (~minute 2)

Reflex sentence: 'Most important question first: [X versus Y]? Those are completely different systems.'
Prompt-specific: Fraud: inline vs offline. RAG: corpus freshness + trust. Moderation: pre-publish vs post-publish. Agent: reversibility tier. Coding: privacy tenancy.

Move 3 — Commitment with revisits-rule (~minute 4)

Reflex sentence: 'I'll commit to [architecture]. If [condition] changes, we revisit. [Pace telegraph: 'Two more minutes on X, then I'll draw.']'
Why it works: Signal 1 (commitment) + Signal 2 (trade-off) + pace control in one sentence.

The cumulative effect

By minute 5: Three of the seven signals from Lesson 1.1 are scored. The interviewer's floor estimate is at-level.
Remaining 40 minutes: Run with the interviewer's confidence. The technical content can be the same as a Senior candidate's; the score will be different because the framing was set at level.

Post-mortem · anonymized

Setup

L6 candidate at a top AI lab, fifth interview of a six-round on-site. The candidate had strong technical content across the loop, including a great Lesson 4.1-style recsys design in the previous round. In this round, the prompt was 'design a code-review agent that comments on PRs.'

What happened

The candidate started drawing immediately. Within 90 seconds the whiteboard had a GitHub webhook, a queue, a code-LLM, a comment poster, and a prompt template. The interviewer asked, 'what's the scope of the comments — style only, or correctness?' The candidate said 'both.' The interviewer asked, 'what's the action surface — are we just commenting, or can the agent suggest commits, or merge things?' The candidate said 'commenting only, for now.' The interviewer asked, 'is this autonomous or are comments human-reviewed?' The candidate paused. The architecture they had drawn assumed autonomous; rebuilding for human-reviewed required reorganizing the diagram. They lost ten minutes.

The moment

The interviewer's post-loop debrief identified that the previous round's recsys design had been excellent and this round's design 'didn't reach the same level.' The actual technical content was equivalent. The difference was that in the recsys round the candidate had run a clean Opening Triplet, and in this round they had skipped it. The choreography difference was visible in the interviewer's grading even though the technical knowledge was identical.

What they should have said

At the start of the code-review-agent round: 'Before I sketch, I want to run through CLARO and the Bounded Autonomy dimensions — about six minutes. Most important question first: what's the action surface — commenting only, commit suggestions, or merge actions? Reversibility tier writes the architecture more than anything else here.' That single 30-second move would have surfaced the autonomous-vs-human-reviewed question explicitly, anchored the design space, and let the candidate commit to the architecture before drawing any of the wrong system. The technical content was there; the choreography was missing.

Lesson

The Opening Triplet is the choreography Staff candidates execute consistently across interviews. Executing it once doesn't make you Staff; executing it consistently is what reads as Staff to interviewers across a loop. Practice the moves until they come without thought, and run them at the start of every design interview regardless of how confident you feel about the prompt. Confidence is exactly when candidates skip the Triplet — and exactly when they shouldn't.