Most candidates fail mock interviews the same way bad code fails in production: not during the happy path, but when something unexpected happens. Here's what I've watched senior engineers get wrong, over and over, that has nothing to do with algorithm knowledge.

I watched a staff engineer candidate — 12 years of experience, worked at Stripe — get rejected from a $380k/year offer in 2023 because he said 'I'd use Kubernetes for this' without being asked. The hiring manager told me afterward: 'He kept giving answers to questions I wasn't asking.' That's the whole problem in one sentence.
Most people treat mock interviews like LeetCode. Grind the patterns. Memorize the frameworks. Nail the time complexity. That's table stakes at this point, and it's not what separates offers from rejections at senior levels.
What actually gets people cut: communication under uncertainty. Interviewers aren't watching what you know. They're watching how you behave when you don't know something. That's the real test. That's what you're not practicing.
Here's the specific failure mode I see constantly. Candidate hits an edge case they haven't thought about. Instead of saying 'give me a second to think through this,' they start talking. Fast. Filling silence with half-formed ideas because silence feels like failure. The interviewer logs it as 'unclear thinker under pressure.' Offer dead.
Everyone's been told to ask clarifying questions. Good advice. But candidates over-indexed on it and now it's a liability.
I ran mock interviews for 6 months with a 40-person eng team at a Series B in 2022. We tracked this. Candidates who asked more than 4 clarifying questions before starting a system design problem got rated lower on 'execution speed' by interviewers — even when the questions were good. Interviewers couldn't articulate why. They just felt like the candidate was stalling.
The sweet spot: 2 targeted clarifications, then state your assumptions out loud and start building. 'I'm going to assume write-heavy workload, 10M DAU, and we don't need cross-region replication for v1 — I'll flag where that changes the design.' That one sentence shows more than five questions.
This one's brutal to watch. Candidate is mid-explanation, realizes what they're saying doesn't quite hold up, and instead of stopping they try to fix the argument while still presenting it. The whole thing collapses. Now they're apologizing. Now the interviewer is uncomfortable. Now both of them are trying to recover a conversation that's structurally broken.
What good candidates do: they treat their explanation like a function call. They finish it. Then they evaluate the output. 'Actually, what I just said breaks if the network partitions here — let me revise the consistency model.' Clean. Confident. Shows you can reason about your own work, which is the whole job.

STAR format. Everyone knows it. Nobody does it right.
The failure: candidates spend 80% of the time on Situation and Task. The interviewer already knows you had a hard project. They're waiting for the Action and Result, which you're now rushing through in 20 seconds because you're out of time. I've seen this pattern end offers at Google, Figma, and Vercel-stage companies.
Set a mental timer. Situation/Task: 30 seconds total. Action: 90 seconds. Result: 30 seconds with a specific metric. '$240k in infra savings,' '4x reduction in p99 latency,' 'shipped 2 weeks early.' Numbers aren't bragging. They're evidence.
If I were prepping someone today, I'd record every mock. Not to watch the content — to watch the body language and pacing when things go wrong. That's where the real bugs are. Most people have never seen themselves interview. It's like shipping code you've never run locally and wondering why prod is broken.