Garry Yeah I find it kind of ironic that in some ways LLMs are forcing a return to waterfall development after all these years. In theory, if an LLM can shorten the coding time, you can iterate fast enough to overcome the criticisms of waterfall, mostly that it isn't responsive enough to changes in requirements, or new discovered requirements. In theory.
The problem is that I have a lot of professional experience and intuition in between my ears that is "just there" when I need it without necessarily having to think much about it, and it's bothersome (though not without some benefits) to translate that back into hard specs and guardrails for an LLM. So when adopting any aspect of assisted development, on this side of the process it just looks like another blasted thing I have to do in my copious spare time.
All I know for sure is that a naked prompt request to an LLM is wrong about half the time, sometimes hilariously so. The trick is providing the context while not adding so much to my own plate that it's not worth it. Also, I have ethical and privacy concerns using cloud models, so I have to get by with something local like Qwen 27B -- though I understand it's no slouch in it own right and should run very well in Apple's unified memory design.
This kind of thing has been the pattern several other times in my career: I am generally not a first adopter but I can also see when something is going to become a permanent and expected fixture in the stack.
