Honest take: we usually rebuild it
AI prototypes are brilliant for discovery. They're a poor foundation for production — and 9 times out of 10 it's faster to rebuild cleanly than to renovate.
Patching costs more than rebuilding
Untangling AI-generated spaghetti — chasing inconsistent patterns, missing types, half-finished abstractions — usually takes longer than starting clean. The economics flipped.
Production demands different bones
Auth, RLS, migrations, background jobs, rate limits, observability — none of that comes out of a prompt. A real architecture is a different shape than a demo.
But the prototype wasn't wasted
It answered the questions a doc never could: what users actually click, what flows feel right, what features you don't need. That's the real output. That's what we build from.