Vibe coding is not a 0/1 skill. LLMs generate code as the prompt says, so when you ask what you want, you get it. If you want a specific pattern or architecture, explicitly ask for that. It works really well when you (not the LLM) drive the development.
And how many times do you have to correct that? Because I have not seen that in practice. With Opus / Claude Code it will follow instructions until it doesn't and then suddenly, in the SAME prompt/todo agent blabber, it creates a web worker instead of using Bull which it did first for exactly the same case. Sure, context overload, but this is the same prompt. And the models claiming larger context like Gemini don't fare any better (well, worse actually).
Now you might tell me; make the tasks smaller and more focused, yes, that's true, it performs better, but then i'm just faster myself coding. Most what I use CC for is exploration & frontend; generating slabs of mind numbing crap react and tailwind is fantastic. But for most other stuff, we have patterns and libs in place where i'm much faster and better than AI as it's not much code, just more logic/thinking.