Thorough Onboarding, Empty Completion

I built my product's onboarding to be thorough. Then I watched people quit it halfway through.

The instinct was good. Surface the hidden parts of someone's story, get past the resume version of who they are. The execution wasn't. People would get eight or nine questions deep and close the tab.

The problem wasn't the questions. It was that the thing had no sense of when it had enough.

A good interviewer knows. You can watch a person realize they've said something that matters, and a good interviewer stops, reflects it back, moves on. My AI had no version of that instinct. It just kept going, because to a model, more input always looks like more value.

So I spent a week building the opposite skill. Not "ask a sharper question." Know when to stop asking.

It turned out to be the hardest thing in the whole product. You can get a model to generate forever. Getting it to recognize that it already had what it needed, and then act on that instead of reaching for one more prompt, took more iteration than any feature I've shipped.

That's the part of building with AI nobody really warns you about. The models are extremely good at doing more. Most of the actual work is teaching them to do less, on purpose, at the right moment.

These days I'm rarely adding capability. I'm drawing edges.