But there is a spectrum here. AI is a cruder, less fine-grained method of producing output. But it is a very powerful tool. Instead of "chiseling" the code line by line, it transforms relatively short prompts along with "context" into an imperfect, but much larger/fully formed product. The more you ask it to do in one go, usually the more imperfect it is. But the more precise your prompts, and the "better" your context, the more you can ask it to do while still hanging on to its "form" (always battling against the entropy of AI slop).
Incidentally, those "prompts" are the thinking. The point is to operate at the edge of LLM/machine competence. And as the LLMs become more capable, your vision can grow bigger.