Behavioural Interview Framework¶

Theme: How you tell the work | Seniority: Senior → Staff

Interviewers are not collecting autobiographies. They are testing whether you can see a system of people and machines, make a decision under constraints, and leave the system better than you found it.

Why answers fail¶

A typical senior candidate walks in with “I led the migration.” Ten minutes later the interviewer still does not know: what was at risk, what you personally decided, what number moved, or what you would do again.

The fix is not more adjectives. It is a spine.

STAR + Reflection¶

Classic STAR stops at the press release. Senior loops add constraints, the decision, and what changed in you or the org.

flowchart LR
    C[Context\nwho, when, stakes] --> P[Problem\nwhat was broken or at risk]
    P --> K[Constraints\ntime, people, risk, politics]
    K --> D[Decision\nwhat you chose, what you rejected]
    D --> A[Action\nwhat YOU did]
    A --> R[Measured result]
    R --> W[What changed\nin the system]
    W --> L[What I learned]
    style D fill:#1565c0,color:#fff
    style R fill:#e65100,color:#fff
    style L fill:#1b5e20,color:#fff

Quantifying impact¶

You do not need a finance partner. You need a denominator.

Latency:     p99 1.8s → 240ms after the change, 3 weeks of data
Reliability: 12 Sev-2s / quarter → 2, same traffic shape
Cost:        $74k/mo Kafka → $51k/mo (−31%) at constant produce rate
People:      on-call pages 40/week → 9/week; MTTR 70m → 22m
Delivery:    lead time 14d → 5d; rollback rate 18% → 4%
Risk:        blast radius 100% of checkouts → 8% canary

If you truly have no number: rate, time, or blast radius. “I wrote a design doc” is not a result. “Two teams unblocked; API shipped in 3 weeks instead of a quarter” is.

One story, three altitudes¶

Same incident — a checkout timeout during a sale. The facts do not change. The unit of work does.

Influence ladder¶

You do not start every disagreement at “all-hands memo.”

flowchart BT
    L1[1. Fix it in your own diff]
    L2[2. Pair / 1:1 with the owner]
    L3[3. Written proposal + numbers]
    L4[4. Design review / RFC]
    L5[5. Cross-team working group]
    L6[6. Escalate with a decision request]
    L1 --> L2 --> L3 --> L4 --> L5 --> L6

Interviewers listen for rung choice. Jumping to 6 on a library bikeshed is a red flag. Staying on 1 while three teams ship the wrong protocol is also a red flag.

Worked example: Technical Disagreement.

Incident flow (tell it in this order)¶

When the prompt is “tell me about an outage,” do not start with the root cause. Start with detection, or they cannot judge your judgment.

flowchart LR
    D[Detect] --> S[Stabilize]
    S --> C[Coordinate]
    C --> M[Communicate]
    M --> G[Diagnose]
    G --> R[Recover]
    R --> P[Postmortem]
    P --> V[Prevention]

Full worked example: Production Incident.

Story inventory (build this before the loop)¶

Have six stories you can retell at junior / senior / staff altitude. Not twelve shallow ones.

1. Production incident you drove
2. Technical disagreement you influenced
3. You were wrong, and changed
4. Delivery under an ugly constraint (date, headcount, legacy)
5. Mentorship or raising the bar (hiring, review, on-call)
6. Ambiguous problem you framed (no ticket, you defined success)

For each, write the eight beats on one page. Rehearse the 90-second version and the 5-minute version.

Question bank — what they are actually scoring¶

Answer hygiene¶

Do

• First person for decisions: “I chose X because constraint Y.”
• Name the alternative you rejected.
• Put the number next to the decision, not in a slide at the end.
• Stop talking when the eight beats are done.

Do not

• Narrate a 14-person “we” with no you.
• Trash a previous employer or a named colleague.
• Claim the entire company’s revenue.
• Use a story you cannot defend under “what did the dashboard show?”

Full story — eight beats (senior)¶

Use this as a template, not a script. Swap in your numbers.

1. Context. Payments platform, 40 engineers, I owned the authorize path. Peak 120 TPS, SLO 300ms p99.
2. Problem. After we tripled retries to “fix” a flaky PSP, double-charges appeared for 0.4% of orders over a weekend.
3. Constraints. Could not take a week for a new ledger; support was already refunding by hand; legal wanted a written control by Monday.
4. Decision. Idempotency key at the PSP, not a full exactly-once rewrite. Rejected “just turn off retries” (availability) and “rebuild the ledger” (time).
5. Action. I spec’d the key format, sat with two service owners to plumb it, wrote the reconcile job against PSP exports, and paused the retry increase myself.
6. Result. Zero duplicates in the next 30 days; authorize success +11% once retries were safe; refund queue cleared in a week.
7. What changed. Lint rule: any outbound money call without an idempotency key fails CI. Runbook for “duplicate charge” now starts at the reconcile table.
8. Learned. Retries are an amplifier. I now ask “what is the idempotency token?” before I ask “how many retries?”

Staff altitude on the same story: you convene PSP + order + support, fund the lint org-wide, and measure duplicate rate as a company SLO — not a ticket on your team.

Red flags (theirs and yours)¶

If you hear a vague prompt — “tell me about a challenge” — pick the story yourself and announce the lens: “I’ll do this as an incident with detect → prevent” or “as a cross-team influence story.” That is a senior move.

90-second skeleton (memorize)¶

Context + problem + stake. “Q3, payments, 2% of auths double-charged after a retry change — trust and money.” Constraint + decision. “Could not dual-write for a week; I chose idempotency keys at the PSP boundary over a full outbox rewrite.” Action (two verbs). “I shipped the key + a reconcile job; I paused the retry increase.” Result. “Zero duplicates in 30 days; retry success +11%.” Changed + learned. “Idempotency is now a lint rule on any outbound money call. I learned retries without keys are a feature for creating money.”

Then shut up. Let them pick a beat to deepen.

Self-Assessment¶

• [ ] Six stories, each with a number and a rejected alternative
• [ ] One story I can tell at junior, senior, and staff altitude
• [ ] One story where I was wrong
• [ ] I can walk detect → prevent without jumping to the bug
• [ ] I know which influence rung I used, and why not one higher