TalkMeThrough an interviewer that asks the follow-up a form never could.
- Started
- Nov 2026
- Runs on
- Any browser
- Kind
- Research and customer stories
- Status
- Closed alpha, one person
Send a link instead of a form. It asks, listens, and asks again.
People answer by talking, the way they would to a person, and every next question is built on what they just said. You get the transcript and a summary in their words, and nothing is used until they have said it may be.
- In numbers
- 3kinds of interview3languages: English, German, Turkish0accounts needed to answer
A form gets a sentence. A conversation gets the story.
Why it exists.
I wanted the people who use my apps to tell me how they came to them, in enough detail to write a customer story. A form gets one careful sentence. A call needs two calendars.
So I ran the interview as a Claude chat first. It worked until I sent it to someone: they needed a Claude account to answer. TalkMeThrough is that interview on a link. Nobody installs anything or signs up. They open it, tap the mic and talk.
Six screens from one interview.
What it looks like.
Seven questions about a cargo bike.
One interview.
A try-out on production, 13 November. Every question is the interviewer’s, word for word. I typed the answers one at a time, after reading each question; the bike and the person are made up.
- Question 1
What’s something you bought in the last year that you’re really glad you did?
A cargo bike. We bought it in March.
- Question 2
Where were you when you actually decided to buy the cargo bike?
At a bike shop in Kreuzberg, on a Saturday morning. We had booked a test ride, and I think we both knew on the way there.
- Question 3Built on their words
You said you both knew on the way there. What had you been doing before that, for getting around and carrying everything?
The car, for everything. Daycare, shopping, all of it. One rainy Tuesday I spent twenty minutes looking for parking outside daycare and was late for a meeting, again.
- Question 4The trigger
After that rainy Tuesday, what did you do?
Nothing, at first. Then our neighbour went on holiday and lent us their cargo bike for a week. After that week, getting back in the car felt silly.
- Question 5
What made you hesitate about buying one of your own, even after that week?
The price. Almost five thousand euros. And we have no garage, so I was sure it would get stolen in the first month.
- Question 6Follows the worry
You were sure it would get stolen in the first month. What did you do about that before you bought it?
We found a bike box in the courtyard you can rent for twelve euros a month. Once that was sorted there was nothing left to argue about.
- Question 7
What does the bike let you do now that you couldn’t before?
I get to daycare in ten minutes, we chat on the way, and I haven’t been late since.
Three kinds of interview, one set of manners.
How it interviews.
One question per message. It starts from a real moment rather than a topic, it never asks what the brief already says, and it never compliments an answer. Whoever sends the link writes a short brief and a private goal, which the person answering never sees, and the interviewer uses the goal to decide which threads are worth following and when it has enough.
Customer story
For a case study or a blog post. The moment it went wrong, the turn, what changed, and lines worth quoting. It asks for their role once, so the story can credit them.
Jobs to Be Done
A switch interview after Bob Moesta: the day they bought, back to the first thought, the event that made it that day, and what worried them.
360 peer review
What a colleague actually saw. A real moment first, and only then a rating, in plain words. “I haven’t seen that” is a good answer.
The first people to use it rewrote it.
What testing changed.
The first transcription model dropped whole sentences from real answers and wrote “Claude” as “cloud”. ElevenLabs’ Scribe kept every word, so it does the listening now.
A friend who builds voice debriefs into their own app tried it and told me to land on the mic, not on a text box. They also caught the pull quote keeping every “um”. Taking fillers out is not one rule: in German “um” means “at” and “er” means “he”, so only English loses those two.
A quote has to be words the person said, in the order they said them. It may leave out up to three, a false start or a stray “and”, but never a “not”.
My own try-out lost the first words of two answers. The mic takes a moment to open on a phone, and I was already talking. The button now pulses until it is really listening, then clicks.
A model that doesn't stop to think.
Under it.
The interviewer is Claude with its thinking switched off. A question comes back in about a second and a half; with a little thinking it took seven, which is a long silence in a conversation.
- App
- Next.js on Vercel
- Data
- Postgres on Neon
- Interviewer
- Claude Sonnet 5.5, via Vercel AI Gateway
- Listening
- ElevenLabs Scribe v2
- Voice
- ElevenLabs Multilingual v2
- Languages
- English, German and Turkish
Whether people want to talk to an AI about something that matters.
It is a closed alpha. I create every interview, and everyone who has answered one got the link from me. A stranger getting a link from a company they buy from is a different conversation, and I don’t know yet how that one goes.
- What it is not
- Not whether the questions are good. What testers have asked to change so far was the screen and the transcript, not the conversation.
- Being tried
- Saying plainly, before the first question, that an AI is asking and who will read the answers. And letting anyone type instead of talk.
Everything on this page was designed, built and shipped end to end.
The design, the code, the launch, the support replies and the marketing that got anyone to look. All of it in my own time, alongside leading design at NewStore.
- What that covered here
- 01Why a conversation instead of a form02The interview, from mic to the four forces03Three kinds of interview behind one link04What the first testers changed05A fast model, English, German and Turkish