Comment by diggan

> The creative part for me includes both the implementation and the design

The implementations LLMs end up writing are predicable, because my design locks down what it needs to do. I basically know exactly what they'll end up doing, and how, but it types faster than I do, that's why I hand it off while I go on to think about the next design iteration.

I currently send every single prompt to Claude, Codex, Qwen and Gemini (looks something like this: https://i.imgur.com/YewIjGu.png), and while the all most of the time succeed, doing it like this makes it clear that they're following what I imagined they'd do during the design phase, as they all end up with more or less the same solutions.

> If you like your expensive AI autocomplete

I don't know if you mean that in jest, but what I'm doing isn't "expensive AI autocomplete". I come up with what has to be done, the design for achieving so, then hand off the work. I don't actually write much code at all, just small adjustments when needed.

> and I find understanding my whole implementation faster

Yeah, I guess that's the difference between "vibe-coding" and what I (and others) are doing, as we're not giving up any understanding or control of the architecture and design, but instead focus mostly on those two things while handing off other work.