Whiteboard Performance & Panic Protocol

The other reference sheets train what you know. This one trains what you do while someone is watching you think — the part that decides whiteboard, live-coding, and "talk me through it" rounds. It is written for people who know more than they can show under observation, and who panic.

Core premise: these rounds do not score a correct answer. They score observable signals of how you work. Panic makes you go silent and chase the answer; the fix is to run a loud, visible process you can execute even while scared. The process is a checklist. You can run a checklist with shaking hands.

This sheet is the method. The day-by-day calendar and daily loop live in sheet 11.

1. What is actually being scored

An interviewer leaves the room grading five signals, roughly in this order:

Problem intake — did you clarify before coding? (constraints, examples, edge cases)
Communication — could they follow your thinking out loud? Silence = lost signal.
Collaboration — did you take hints, ask, and adjust with them (a coworker), or perform at them?
Correctness under your own test — did you trace your code on an example and find your own bugs?
Complexity awareness — do you know what your solution costs and what's better?

Note what is not on the list: solving it fast, solving it alone, solving it without hints, an optimal answer on the first try. A candidate who narrates a brute force, states its cost, and improves it with the interviewer outscores a silent candidate who types the optimal answer. Your job is to make your thinking visible, not to be a genius in silence.

Reframe to carry in: "I am not being tested. I am pair-programming with a colleague on a problem neither of us has open-sourced yet." That is literally the thing they are hiring for.

2. The loop: CLARP (say each header out loud)

A five-phase spine. Announce the phase you're in ("Okay — clarifying first"). Announcing the phase is the communication signal, and it buys you thinking time.

Phase	Out loud	On the board
C — Clarify	"Let me make sure I have the problem. Input is…, output is…, and can I assume…?"	Write the signature + 1 example I/O + constraints
L — Lay out	"Brute force first so we have something correct: … That's O(n²). I think we can do better with …"	Write the approach in 3–4 plain-English bullets before any code
A — Attack	"I'll implement the hash-map version. Let me talk as I go."	Code the bullets, top to bottom
R — Run	"Let me trace this on `[2,7,11]`, target 9."	Walk the example line by line, updating variable values on the board
P — Polish	"Time is O(n), space O(n). Edge cases: empty input, no solution, duplicates."	Note complexity + the edge cases you handled

The order is load-bearing. Never skip C and L to start coding — jumping to code is the single most common panic tell and it forfeits signals 1 and 2. Two minutes of clarifying out loud is never wasted.

3. Panic first-aid (when your mind goes blank)

Grounded in evidence-based performance-anxiety techniques (diaphragmatic breathing; "expressive writing" to clear the blank; narrative over abstraction):

Board-dump on arrival (expressive writing). Before the problem, in the first 30 seconds, write your templates in a corner of the board: the CLARP headers, and 2–3 pattern skeletons you rehearse (binary-search bounds; BFS queue; hash-map counting). This externalizes memory so panic can't erase it — the formulas are already on the wall.
One physiological breath. A slow exhale-longer-than-inhale breath (in 4, out 6) once, deliberately. It is the fastest way to drop the adrenaline spike enough to speak. Practice it so it's automatic, not a thing you forget.
Narrate the stuck, don't hide it. Say the true sentence: "I'm blanking for a second — let me re-read the problem and think out loud." This reads as composure, not failure, and it re-opens collaboration (often the interviewer nudges).
Shrink the problem. Solve n = 1, then n = 2, by hand on the board. Almost every algorithm is visible in the tiny case. This is the "narrative" move — a concrete story instead of abstract numbers.
Fall back to correct-and-slow. A working brute force you can state the cost of beats an optimal solution you can't finish. Ship it, then optimize.

Rule: a stuck moment narrated is a passing moment; a stuck moment gone silent is a failing one. The panic is not the problem — going quiet is.

4. Asking questions & working synchronously (the collaboration signal)

These are the exact scripts. Use them verbatim until they're yours.

Clarifying (before coding — always ask ≥2):

"Can I assume the input fits in memory / is already sorted / has no duplicates?"
"What should I return on an empty input or when there's no valid answer?"
"Are we optimizing for time, or is space tight here too?"

**Taking a hint (this is a positive signal — accept, don't defend):**

"That's a good nudge — let me follow it. So if I use a heap instead of sorting…"
Never argue with a hint. The interviewer is telling you the answer they want to see you reach with them.

Thinking out loud without rambling (narrate decisions, not keystrokes):

Good: "I'm choosing a hash map here so lookups are O(1) instead of scanning."
Skip: reading each character you type.

When you disagree or see a tradeoff (shows seniority):

"We could sort for simplicity, O(n log n), or use a heap for O(n log k) if k is small — which matters more here?"

Checking in (makes it synchronous):

"Does this approach match what you had in mind before I code it?"
"Want me to handle the duplicate case now or note it and move on?"

5. Self-review rubric (grade each rep 0–2)

The daily loop and calendar live in sheet 11. After each rep, watch two minutes of the tape — no more — and score yourself. The camera is the training stressor: it manufactures the "someone is watching" pressure you are desensitizing to.

Signal	0	1	2
Talked continuously	long silences	some gaps	narrated throughout
Clarified before coding	jumped to code	one question	≥2, wrote example first
Traced own code	didn't	partially	full trace, found issues
Stated complexity	forgot	time only	time + space + edge cases
Composure	visible spiral	recovered slowly	named it, breathed, continued

Target: not a perfect 10, just +1 on your weakest row each day. Two weeks of +1 is a transformed interview. Log one line per rep (just rep "…") and note ONE thing to fix next rep.

6. Get a human in the loop

The camera desensitizes you to being observed; a person desensitizes you to being interrupted and judged. Both are the real stressor, split in two.

A friendly practice partner (e.g. a family member willing to Zoom) is perfect early on: they don't need to know the CS. Their job is to watch you present and interrupt with "why did you do that?" — that alone trains composure and narration. Hand them the §5 rubric to grade you.
A technical mock (peer, or an AI interviewer) once the reflex exists: someone who can throw a real hint or a follow-up constraint mid-problem.
Rotate stressors so no single format stays scary.

7. Interview-morning version (the one-page panic card)

Board-dump templates + CLARP headers in the corner first thing.
One slow breath (in 4, out 6) before you touch the marker.
Clarify out loud — ask two questions before any code.
Brute force first, state its cost, then optimize.
Narrate decisions, not keystrokes. Silence = lost points.
Trace your code on a tiny example. Find your own bug.
Blanking? Say so, breathe, shrink to n=1. Stuck-and-talking passes.
It's a colleague and a problem, not a test and a judge.

8. CLARP ↔ REACTO ↔ UMPIRE crosswalk

Different bootcamps and platforms drill the same five-signal loop under different labels. If a candidate already thinks in REACTO or UMPIRE, don't make them relearn vocabulary mid-rep — map the words, keep the gates.

CLARP (this packet)	REACTO (Fullstack Academy)	UMPIRE (CodePath)
Clarify	Repeat, Examples	Understand
Lay out	Approach	Match, Plan
Attack	Code	Implement
Run	Test	Review
Polish	Optimize	Evaluate

All three gate on whether a signal was emitted, never on whether the mid-rep work is correct — a phase passes on "did they say it," not "is it right." Two imports are worth stealing regardless of which spine you run: REACTO splits Repeat from Examples into two separate gates, so restating the problem without a concrete example still fails visibly instead of sliding through (its Approach step is CLARP's Lay out under another name); UMPIRE's Match gate makes "name the pattern family" its own checkpoint, catching someone who restates and jumps straight to code without ever naming an approach.

Weak signal	Run
Composure / silence under pressure	CLARP
Clarify-before-coding weak, jumps straight to code	REACTO
Panel uses bootcamp vocabulary ("REACTO", "UMPIRE")	REACTO
Default	CLARP

9. Research backing

The web guide carries the evidence and video shelf: docs/guide/interview-practice-evidence.md.