Many people can solve a problem silently and then fall apart when they have to talk through it. Interviewers can only score what they hear.
Recording yourself is unpleasant, and it's the fastest way to improve. Transcription tools now give you text with timestamps, which means you can measure things that were once only a feeling: long silences, filler words, and how often you give a reason.
Coditioning's guide describes the same balance from the other side: weak candidates go silent for long stretches, or talk constantly without adding insight.
- Transcript segment: A piece of text with a start and end time. Most speech-to-text tools can export these.
- Reason word: A word that signals why: because, so that, instead of, the risk is. A rough but useful marker.
- Monologue: A long stretch of talking with no pause for the interviewer to step in.
Steps:
- Record one timed round. Use your phone. Speak as if someone were there.
- Transcribe with timestamps. Any speech-to-text tool that exports segments will do.
- Run the numbers. Look at silences, reasons, fillers, and the longest monologue.
- Pick one fix. Change one habit per round, or you won't be able to tell what helped.
- Ask AI to grade the transcript. Numbers show patterns, and a reader can tell you what was unclear.
Treat the numbers as prompts: The limits in this script are starting points that we picked, and no study backs them. Use them to spot your own patterns and compare one week with the next.
Prompt, Grade my explanation:
Below is a transcript of me solving a coding problem out loud. Act as a strict interviewer. Don't comment on the code's correctness.
Find: 1) moments where I said what I did but not why, 2) claims I didn't back up, 3) places where you'd have lost track. Quote my exact words for each. Finish with the one habit to change first.
[transcript]
Code, Measuring a recording. Takes transcript segments and works out what an interviewer would feel.
const FILLERS = /\b(um+|uh+|erm|like|basically|you know|kind of|sort of)\b/gi;
/** Words that signal a reason. Saying what you do is narration; saying why is the signal interviewers want. */
const REASONS = /\b(because|so that|trade-?off|instead of|the risk is|which means)\b/gi;
export function analyseTalk(segments: Segment[], silenceLimit = 20): TalkReport {
if (segments.length === 0) throw new Error("no speech in this recording");
const text = segments.map((s) => s.text).join(" ");
const minutes = (segments[segments.length - 1].end - segments[0].start) / 60;
let longestSilence = 0;
let longSilences = 0;
let longestMonologue = 0;
let runStart = segments[0].start;
segments.forEach((s, i) => {
const gap = i === 0 ? 0 : s.start - segments[i - 1].end;
if (gap > longestSilence) longestSilence = gap;
if (gap > silenceLimit) longSilences++;
if (gap > 3) runStart = s.start; // a pause of over 3 seconds ends a monologue
longestMonologue = Math.max(longestMonologue, s.end - runStart);
});
const per = (n: number) => Math.round((n / minutes) * 10) / 10;
return {
minutes: Math.round(minutes * 10) / 10,
wordsPerMinute: Math.round(text.split(/\s+/).filter(Boolean).length / minutes),
fillersPerMinute: per((text.match(FILLERS) ?? []).length),
reasonsPerMinute: per((text.match(REASONS) ?? []).length),
longestSilence,
longSilences,
longestMonologue,
};
}
const REASONS: Words that signal why. It's a rough marker, but it shows a real pattern across a whole round.s.start - segments[i - 1].end: The gap between segments is your silence.if (gap > 3) runStart = s.start: A pause of over 3 seconds ends a monologue. That's the interviewer's chance to step in.segments.length === 0: An empty transcript means a recording problem, so it stops with a clear error.
Code, Turning numbers into advice. Each number maps to one thing you can do in the next round.
export function talkAdvice(r: TalkReport): string[] {
const out: string[] = [];
if (r.longSilences > 0) out.push(`You went quiet ${r.longSilences} time(s). Before a long think, say what you're checking.`);
if (r.longestMonologue > 180) out.push("You talked for over 3 minutes without a pause. Stop and check in with the interviewer.");
if (r.reasonsPerMinute < 1) out.push("You said what you did but rarely why. Give a reason at each decision.");
if (r.fillersPerMinute > 6) out.push("Lots of fillers. Replace them with a short silence.");
return out;
}
r.longSilences > 0: Long silences are fine if you announce them first.r.reasonsPerMinute < 1: The most common gap: saying what you do, but not why.Failure, Fixing everything at once: You'll sound stiff. One habit per round.
Failure, Talking to fill silence: Silence is fine when you've announced it: 'Give me a minute to check the edge cases.'
Failure, Reciting memorised phrases: Interviewers notice. Reason words should come from reasons you actually have.
- What's the difference between narration and a reason?
- Narration says what you're doing: 'now I loop over the list.' A reason says why: 'I loop once because the input is sorted, so I don't need a nested loop.'
- How do you think quietly without losing the interviewer?
- Say what you're about to check and for how long, then go quiet. Report back when you're done.
Sources:
- Meta AI-enabled coding interview guide (Prep vendor): https://www.coditioning.com/blog/13/meta-ai-enabled-coding-interview-guide
- How I used AI to practice for my Meta interview (Individual): https://dev.to/karuha/how-i-used-ai-to-practice-for-my-meta-interview-and-got-the-offer-543n