What the first 500 builds taught us about asking questions

We opened Quiny to the first batch expecting a particular kind of user: someone who had thought about their app, maybe sketched it, and would arrive with a paragraph or two describing what they wanted. That person exists, and is rare.

Everyone else typed a sentence. "A skincare shop, five products, mostly repeat customers." "Booking for a dog groomer." "My bakery, we do pre orders on Fridays." Then they waited, because from their point of view they had told us what the business was and the rest was our job.

The blank canvas was the problem

Our first console did what most tools do. It took the sentence and started building, filling every gap with a reasonable default. The results were fine and almost nobody finished. People watched five screens appear for a business that was nearly theirs and did not know how to say what was off. The prompt line was right there, but "make it more like my shop" is not something anyone wants to type.

We looked at the 500 builds and one number stood out. The builds that shipped were the ones where the person had answered a question inside the first minute. It did not matter which question. Answering something turned a demo into their app.

The sentence tells us what the business is. It never tells us how the business actually works.

Eight, not twelve, not four

So we rebuilt the opening as a short interview. The first version asked twelve questions. Completion went up and then fell off a cliff around question nine, where people stopped feeling clever and started feeling processed. We cut to four and got confident, generic apps back. Eight is where answer quality peaked and drop off stayed flat, and it has held across every batch since.

The rule for a question earning its place is simple. It has to change what gets built. "Who buys from you?" changes the home screen. "Do people reorder the same things?" changes the whole information architecture. "What is your favorite color?" changes nothing we cannot read from your photos, so it went.

What the questions look like now

Most of them are taps, not typing. Quiny proposes an answer from what you have already given it and you confirm or correct. For the skincare shop that looked like this:

Every answer is remembered. When she came back three weeks later to add a product, Quiny did not ask any of it again.

What moved

Far more builds now make it from the first sentence to a submitted app, and people ask for fewer changes before approving a screen, which we read as them getting the app they meant the first time. And the credit cost of a first app dropped, because rebuilding a screen that was wrong from the start is the most expensive thing you can do in Quiny.

We shipped three model upgrades in the same period. None of them moved completion the way the questions did.

What we are still getting wrong

People with an existing website expect us to read it and skip questions, and we only partly do. We read products, prices and the brand voice, but we still ask about customers because sites rarely say who actually buys. We are testing a version that infers it from the products and asks you to confirm in one tap.

And the ninth question is still tempting. Every week someone on the team proposes one. So far, no.