I see some videos about how AI, if used by a new company/startup that's fully vibe-code focused, is good for prototyping stuff but when it comes to production things go to shit real fast.
Generally speaking i think it's true, but also not at the same time.
How I generally do it at the moment:
Prototype for fun. Document.
Once you had your fun, research, document, then build (from scratch again) the product assuming production-ready use immediately, and constantly test and edge-case test, and load test.
That "document > research > document" part is basically heavy research and understanding the logic/system flows to keep the AI in the right track, and always test for production-level use.
And after a while, I'd have expert/talented human code-reviewers on it.
That's how I'm doing things anyway.
AI won't be magical for those aren't knowledgeable about the thing they want to make.
AI programmed DNN for me, but it didn't think of that solution for the ICANN-DNS problem, it didn't solve Zooko's triangle, it didn't think of the naming encoder system that's scalable, it didn't think of the adoption strategy, it didn't think of collision-avoidance with legacy DNS, it didn't create the UX flow of PWANS and think of the edge-cases of the eCash flow, had DEG Mods been made with AI it wouldn't have known or pushed for various UX enhancements and censorship-resistant systems.
AI is great with coding, it's saving me tens of thousands of dollars, but it's no where near the mastermind.