I stopped prompting harder. I started engineering the loop.

I used to ask AI for the whole app in one prompt. Then I would ask again, and again, and get a slightly different mess each time. The problem was not the model. It was the way I was working. One promp