First, for context, I'm a pretty senior developer, and I've been doing code since 1983, but I'm …

Your use case is one that I have been gently leaning into for small projects at work and personal projects - the LLM is used to give me an initial ideal and code. In a sense, the LLM writes "the one I throw away", and by exploring the limitations of what it produces I both get a better idea of the problem domain and more confidence in actually implementing something useful.

This is exactly how I use it. I’m almost done with a personal project where I didn’t even know how to begin to start, but now I have a decent handle on the codebase.

> I have found that they can get into "death spirals," where their suggestions keep getting worse and worse. I've learned to just walk away, and try something else, when that happens. I shudder to think of junior engineers, implementing the code that comes from these. The death spirals (great term btw) are caused by context pollution. Basically if the LLM gets something wrong in its head (context), it can't move away from it. Just the fact that it exists nudges its decisions towards it. This is where you need to ask it to summarise where you are to a markdown document or similar (the kids call it a "memory file"). Clear the context (start a new chat) and use the markdown file to bootstrap your progress.