I get much better results the more thought I put into crafting my prompt. including using llms to help create that prompt. There's definitely a declining rate of return on that time, but thinking about the problem and carefully describing the context can take fairly deep thought. I do think it's in shorter bursts than when doing all of the work, but I get that same feeling of 'bah, where was I?' if I get interrupted while creating the prompt for a more complex feature. On the other hand, I spend a lot less time in flow state while debugging - it's way easier to describe a bug to an llm (often can just paste in the exception or link to error log).