Same. I had some idea that I wanted to build a basic sinatra webapp with a couple features. First version was pretty good. Then I asked it to use tailwind for the css. Again pretty good. Then I said I wanted to use htmx to load content dynamically. Suddenly it decides every backend method needs to check if the call is from htmx and alter what it does based on that. No amount of prompting could get it to fix it.
Hard to tell what exactly went wrong in your case, but if I were to guess - were you trying to do all of this in a single LLM/agent conversation? If you'll look at my prompt history for the game from OP you'll see it was created with a dozens of separate conversations. This is crucial for non-trivial projects, otherwise the agent will run out of context and start to hallucinate.
Agent mode in RubyMine which I think is using a recent version of sonnet. I tried starting a new agent conversation but it was still off quite a bit. For me my interest in finessing the LLM runs out pretty quickly, especially if I see it moving further and further from the mark. I guess I can see why some people prefer to interact with the LLM more than the code, but I’m the opposite. My goal is to build something. If I can do in 2 hours of prompting or 2 hours of doing it manually I’d rather just do it manually. It’s a bit like using a mirror to button your shirt. I’d prefer to just look down.
> If I can do in 2 hours of prompting or 2 hours of doing it manually I’d rather just do it manually.
100% agree, if that was the case I would not use LLMs either. Point is, at least for my use case and using my workflow it's more like 2 hours vs 10 minutes which suddenly changes the whole equation for me.
Yeah, or 10 minutes of prompting and then 20 minutes of implementing my own flavor of the LLM's solution vs 2 hours of trial and error because I'm usually too lazy to come up with a plan.