The original version of my site aidev.codes was a bit like that. The concept was to keep refining the spec and each time it would regenerate starting with whatever section had been modified. There was even a way to clone an "app spec" like with Codepens.

People generally didn't seem very interested. Also the bigger problem was that the gpt-3.5 models really wouldn't return the same thing consistently so a minor change could throw other things off.

Also a conversation feels like a much more intuitive way of doing it. So I switched to that.