I Built an Interview Tool for the Head. The Real Problem Lived in the Mouth.

A strange thing happens when you build a product about a problem you thought you understood.

You start with a clean theory. The theory feels obvious — so obvious you barely notice it's a theory at all. You build the whole first version on top of it. And then real people show up, use the thing, and quietly demonstrate that your obvious truth was a confident guess wearing the costume of a fact.

This is the story of the assumption I got wrong, how I found out, and the lesson that's outlasted the mistake.

The assumption

I started building Intervues believing interview preparation was a knowledge problem.

The logic seemed airtight. People fail interviews because they have gaps — a data structure they never fully grasped, a system-design pattern they can't quite recall, a concept they half-know. So the job of an interview-prep tool is simple: find the gaps, fill the gaps. Quiz them, surface what they're weak on, help them learn more. More knowledge in, better interviews out.

I had nine years of engineering behind me and three years of conducting interviews on the hiring side. If anyone was qualified to define the problem, surely it was me. So I defined it, and I built toward it, and I felt good about how clear the whole thing was.

That clarity was the problem. I just couldn't see it yet.

The correction

The correction didn't arrive as a single dramatic moment. It arrived slowly, in the form of transcripts.

I started reading real interview transcripts — actual sessions, real candidates, the unedited record of what people said when someone asked them a question and waited. I read a lot of them. And somewhere around the fortieth or fiftieth, the pattern I expected to see simply wasn't there.

I had expected to see knowledge gaps. People stumbling because they didn't know the thing.

That's not mostly what I saw.

What I saw, again and again, were people who clearly knew the thing — you could tell from the fragments, the half-sentences, the eventual recovery on the drive-home version they'd type out later — but who could not get the knowledge out of their head and into the room in real time. The answer existed. The retrieval worked. The translation broke.

I watched candidates with obvious competence freeze on questions they could have answered in writing without breaking a sweat. I watched people give a tangled, apologetic non-answer and then, when the interviewer rephrased it more gently, deliver the exact thing they'd failed to say thirty seconds earlier. The knowledge had been there the whole time. It was never the bottleneck.

I had built a tool aimed at the head. The real problem lived in the mouth.

Why I couldn't see it from the inside

Here's the part that's humbling, and worth sitting with if you build anything.

My assumption wasn't stupid. It was reasonable. It was even partly true — knowledge gaps are real and they do cost people interviews. But I had mistaken a contributing factor for the main one, and I'd done it because my own vantage point hid the truth from me.

As an interviewer, I had spent three years watching people fail and unconsciously attributing the failure to what was easiest to measure: whether the answer was right. As an engineer, I'd spent nine years in a world where knowing the answer basically is the job. From inside both of those roles, "interview prep = knowledge" looks like common sense.

What I didn't have was the candidate's interior experience — the specific, private horror of knowing something and watching it refuse to come out of your mouth while a stranger waits and a clock runs. I'd never systematically looked at that, because my theory told me it wasn't the important variable. The theory was protecting itself. Assumptions do that. They quietly filter the evidence so the thing you already believe keeps looking correct.

It took a hundred transcripts to break through, because data at that volume stops being something you interpret and starts being something that interprets you.

What changed because of it

Once I saw it, I couldn't unsee it, and the product had to change.

A tool built for the head and a tool built for the mouth are not the same tool. So we rebuilt around the real problem. More room to pause, because thinking takes time and a pause is not a failure. Patience with a thinking silence instead of cutting it off after a second the way most voice AI does. Follow-up questions that test whether you can defend an answer under a little pressure, not just whether you can produce one. An honest report at the end — what you actually sounded like — instead of a cheerful, useless "great job!"

And the deeper repositioning: Intervues stopped being a place to learn more and became a place to practice saying what you already know. To get the reps out loud. To do the clumsy first attempt in private, on purpose, so the clumsy first attempt doesn't happen in the room that counts.

None of that would exist if my first theory had been right. It exists because it was wrong, and because I found out.

The lesson that outlasted the mistake

I've thought about why this one stuck with me more than most build lessons, and I think it's because it's not really about interviews at all.

It's this: you do not understand a problem until you have watched a lot of real people run into it. Not two. Not the five who agree with you. A lot of them, enough that the pattern stops being a story you can spin and becomes a fact you have to accept.

Everything you believe before that point is an educated guess. Some of your guesses will be good — mine wasn't entirely wrong, just wrong about what mattered most. But all of them are guesses, and the dangerous ones are the guesses that feel like certainties, because those are the ones you never think to check.

So the job, if you're building anything, is humbler than it sounds. Build the thing. Put it in front of real people. Watch them use it. And then — this is the hard part — let them correct you, even when the correction costs you the clean theory you were proud of. Especially then.

I was confident interview prep was a knowledge problem. The transcripts disagreed. I'm grateful they did, because the product is honest now in a way it never would have been if I'd been right.

The answer was always in their heads. My job was just to stop assuming I already knew why it wouldn't come out.

I'm building Intervues — voice-led AI interview practice, rebuilt around the way real candidates actually speak. Founding cohort opens July 2026. I write about interviews, building in public, and being usefully wrong.