I found that it normally takes one prompt early-on to go from 'vibe-coded spaghetti' to something having a decent architecture.

I cap my effort at 2-3 prompts. One to investigate obvious mistakes with a top model, 1-2 to try to fix them.