Live Coding Interview — Prep Guide

The Compass

The interviewer is trying to answer one question: "Would I want to pair-program with this person?" Every behavior should make the answer yes. If a prep rule tells you to say something that would make you annoying to pair with — ignore the rule.

Scripted = annoying. Silent = opaque. Narrating the obvious = condescending. Drawing instead of building = slow. The sweet spot: talk like you're thinking alongside someone you respect.

---

T-Minus 10 Minutes

Before you join the call:

• Permissions pre-allowed — no "Allow?" dialog during the interview
• Repo already warm — opened, indexed, AI has read the codebase
• CLAUDE.md or .cursorrules in place with project context
• Parallel agents: OFF — serial is easier to narrate
• Fresh context — /clear or new session, no stale conversation
• Drawing tool open in a separate window
• Phone timer visible — for the 10-minute narration budget

Minute-Zero Line

Say this in your first breath:

"Walk me through what you'd want this prototype to do, then I'll sketch a thin slice, ask a couple questions, and we'll build from there."

This does four things at once: sets collaboration tone, scopes the task, buys you 30 seconds to breathe, and positions you as the one driving the session.

---

Prep 0: The First Five Minutes

Before you write a line of code:

1. Thin slice in their domain

"So we're building a [what] that takes [input] and returns [output]. I'll start with the smallest version that works end-to-end, then add layers."

2. Two scope questions

Pick two (not one, not three):

• "What's the biggest constraint — time, correctness, or UX?"
• "Are we optimizing for prototype speed or production readiness?"
• "What's the one thing that would make this prototype a win for you?"

3. One stack-bridge line

"I'll build this in [your stack] because [one reason tied to their constraint]."

You now have scope, constraints, and a stack justification. The interviewer has already seen you ask before you build, connect tech choices to business constraints, and drive the session.

---

The Six Action Mantras

When your brain is full and you can't remember principles, fall back to these. Each is ≤10 words — they work when you're stressed and non-native.

#	Mantra	Covers
1	"Done with X. Next: Y, because Z."	Fork narration + intent. The unsheddable one.
2	"Screen shows what. I'm saying why."	Invisible/visible boundary
3	"One thing I checked: [specific]."	Code review
4	"Typo." / "I've seen this." / "Let me think." / "External."	Debug triage
5	"Let me sketch — three moving parts."	Diagram trigger
6	"What would you poke at first?"	The landing line

Overload Hierarchy

When you can only do one thing:

1. Layer 1 (always possible): "Done with X. Next: Y, because Z."
2. Layer 2 (if you have capacity): Name one rejected alternative.
3. Layer 3 (if you have more): Self-test — "most important decision, did I say why?"
4. Layer 4 (rehearsal only): The full principles below.

Layer 1 is non-negotiable. If you only say "Done with X. Next: Y, because Z" after every unit of work, you are 80% of the way to a strong interview.

---

The Five Principles — R³·D²

Three R's, then two D's: Recover, Reveal, Review · Debug, Draw. Say "R-three-D-two" like the droid and the five principles come back in order, with their mantras attached:

	Principle	Mantra	Hook
R1	Make reasoning Recoverable	"Done with X. Next: Y, because Z."	Breadcrumb trail
R2	Reveal what the screen won't remember	"Screen shows what. I'm saying why."	Director's commentary
R3	Review what matters	"One thing I checked: [specific]."	Customs officer opens ONE suitcase
D1	Debug like you're pairing	"Typo." / "I've seen this." / "Let me think." / "External."	ER triage nurse
D2	Draw to think	"Let me sketch — three moving parts."	Napkin, not blueprint

Each hook is a picture, not a rule — recall the picture and the behavior follows.

---

Principle 1: Make Your Reasoning Recoverable

Goal: Someone who walked out for 5 minutes could reconstruct every consequential decision and why — just from what you said.

Mantra: "Done with X. Next: Y, because Z."

Hook (R1): Breadcrumb trail — a colleague who stepped out for five minutes can retrace every fork from your words alone.

You're failing if the interviewer asks "why did you do X?" — you're hoarding reasoning. You're also failing if you chain "next I'll do X, next I'll do Y" without ever naming a rejected alternative. That's a todo list, not reasoning.

When interrupted: Answer their question first. Only recover context if you genuinely lost it — no formulas.

When asked "why did you do that?": Answer the question directly. No meta-apology. Then make the next two decisions more vocal. The content proves you learned; the apology proves you read a prep guide.

---

Principle 2: Say What the Screen Won't Remember

Goal: The code stays on screen. Your reasoning doesn't. Verbalize the invisible: rejected alternatives, assumptions, intent.

Mantra: "Screen shows what. I'm saying why."

Hook (R2): Director's commentary — the screen is the film; you're the track that says why the scene exists.

You're failing if:

• You read your prompt text aloud (visible)
• You explain basics a domain expert already knows
• You invent rejected alternatives you never considered (worse than silence)
• You don't have one: "I only know one way — open to better ones." Honest. Power move.

---

Principle 3: Review What Matters

Goal: Prove critical reading without a line-by-line audit.

Mantra: "One thing I checked: [specific]."

Hook (R3): Customs officer — don't open every suitcase; open the risky one and say which and why.

For small generated diffs (≤50 lines): Per-file verdict. "Model looks right — composite key. Service is light on errors — noted."

For large generations (>50 lines): Scan the riskiest file only. Say which file and why. "Checking the auth module — if that's wrong, nothing else matters."

For streaming: Narrate as it arrives. The narration IS the review.

You're failing if:

• "Looks good" with no specifics. Approval, not review.
• "Nothing jumps out" for the 5th time. Rubber-stamping.
• You miss something the interviewer finds: "I missed that — let me look."

---

Principle 4: Debug Like You're Pair-Programming

Goal: The interviewer should feel like they're debugging with you, not watching you debug.

Hook (D1): ER triage nurse — classify the patient before treating anyone.

Triage tier (say this, not a timer):

Tier	Words	Action
Typo	"Typo."	Fix instantly. Zero ceremony.
Known	"I've seen this."	"Usually [X]. Checking." ~30s.
Internal	"Let me think."	Logic/config error. Hypothesis → test.
External	"External."	Rate limit, API outage, quota. Check service BEFORE debugging code.

The External tier is critical: a rate limit or API outage can eat your entire round if you treat it as a logic bug.

For Internal bugs: Count hypotheses, not minutes. After 3 failed hypotheses, ask. "Three things ruled out. My best guess is [X]. Should I continue or move on?" If they say continue: broaden scope (hypotheses 4-5), then narrow (6-7). After hypothesis 7: shift to collaborative. "Here's what I've ruled out. Your guidance?"

You're failing if:

• "Typo." takes 15 seconds of narration. It's a typo. Fix it.
• You're watching the clock more than the bug.
• "My best guess is X" when you don't have one. Say so.
• "Continue or move on?" when you know you're close. "I'm close — one minute."
• Claiming to have "ruled out" things you barely checked. Say what you tried.

---

Principle 5: Draw to Think

Goal: A diagram is a thinking tool, not a deliverable. Draw when it helps YOU reason.

Mantra: "Let me sketch — three moving parts."

Hook (D2): Napkin, not blueprint — a messy sketch that clarifies beats a pretty one that decorates.

Draw when you can't hold the system in your head. If the tool fights you for more than 60 seconds, describe it verbally. If screen share is blurry, narrate everything — the diagram is your private scratchpad.

You're failing if:

• You draw something explainable in two sentences
• The diagram is prettier than it is useful
• You redraw for cleanliness — annotate instead

---

Prep 6: AI Tool Craft

Generation waits are free narration time

When the AI is generating, you're not typing — talk. "While it generates, my plan is [X]. The interesting decision coming up is [Y]."

The Rule of Two

If the first generation is wrong, you get ONE fix-prompt. If the second is still wrong, you write it yourself. "Two attempts — I'll write this one manually." This proves you can code without AI. Interviewers are specifically watching for this.

Interrupt bad generations early

"Wrong direction — canceling. Let me be more specific: [narrowed prompt]." Saves time and shows you're steering, not spectating.

Context hygiene

• /clear between major features
• Don't let stale narration from 10 turns ago poison current prompts
• If the AI references something obsolete, clear it

---

Prep 7: The Landing (T-15 to T-0)

At T-15: Feature Freeze

"No new features. Let me make sure what we have works end-to-end."

The Demo (2-3 minutes)

1. Happy path: "Here's the core flow working."
2. One deliberate failure: "Here's what happens when [most likely failure]." Pick the failure — don't discover it accidentally.
3. Hand it back: "What would you poke at first?"

This turns a demo into a conversation. The interviewer tells YOU what they care about — and you get one last chance to address it.

If you finish early (5+ minutes remaining)

Don't fill time with features. "I have time — want me to harden anything specific, or should I talk through what I'd add with more time?" If they say "whatever you think": pick ONE thing that would most impress THEM, based on what you learned in Prep 0.

---

Chaos Protocols

1. Task is too big: "This is a lot for 60 minutes — where should I focus?" This is the senior move. Junior devs try to build it all.

2. "Whatever you think": "I'll go with [X] — easy to change if we want [Y] later." Pick the reversible option. Move on.

3. Need to re-plan: "That changes the approach. Let me take 30 seconds to replan." Price the replan aloud. Then actually replan.

4. "Don't use AI for this part": "Got it." Close the AI tool. Don't argue. Prove you can code without it.

5. Unknown stack: "I don't know [their stack] well, but here's how I'd approach it in [mine]." Bridge to what you know. Patterns transfer.

---

Narration Budget: ≤10 Minutes Total

v4's narration overhead measured at ~30-37% of a 60-minute round. That's 18-22 minutes of talking about code vs. building code. Too much.

Hard budget: narrate no more than 10 minutes total. Track on your phone timer. If you're past 5 minutes by T-30, throttle to mantra-only: "Done with X. Next: Y."

---

Pre-Interview Rehearsal

Record a 10-minute mock. Play back. Ask:

• Did I open with the minute-zero line and Prep 0? Or did I start coding immediately?
• Did I say "Done with X. Next: Y, because Z" after every unit of work?
• Did I triage errors before debugging? Or did a typo get a 3-minute investigation?
• Did I use generation waits as narration time? Or did I sit silent?
• Did I feature-freeze at T-15 and land the plane?
• Would I want to pair-program with this person?

---

This guide is the output of 15-team adversarial review across 5 rounds. v1 was 5 naive prep points. v5 stripped all trained phrases after interviewer-POV teams found them detectable as scripts. What remains: principles, mantras, anti-patterns, and a compass. Use your own words.