Knowing how to build it was the problem

One of the stranger problems I had building an AI product was that I was pretty good at finding the next thing to build.

There was always something sensible waiting for me: improve the flow, tune the scoring, fix the language, clean up the review notes, make one awkward step feel finished. None of that was fake work. Usually the product was better when I was done.

It was also very convenient work.

Product work comes with the comforting little pleasure of completion. A screen changes. A test passes. A problem that was red turns green.

Demand is ruder because another person has to do something, and I don't get to do it for them.

I built that product end to end, and Stripe still showed $0 in MRR.

I don't know exactly why, and the record doesn't support a tidy diagnosis after the fact.

What I can say is that I kept giving myself questions I knew how to answer while the important one stayed with the customer: would a specific person choose this strongly enough to give up money, time, effort, an old habit, or even the energy to come back?

That is an annoying division of labor. Cleaning up the product can happen on my schedule; the choice doesn't belong to me.

AI makes the temptation worse because the next missing capability can be close enough to build before lunch. That work is productive. It just may be producing a better answer to a question nobody has chosen to ask.

So the next action I want is one that requires something from outside the product. It does not have to be a charge on day one, but it does have to ask for a real choice: a return, a changed habit, time spent, effort given, or money paid.

Praise and a nicer demo can help. They cannot make that choice for the customer either.

I'm still going to build things. Next time, the build needs to make demand easier to observe, not give me another week of very respectable avoidance.