> You are discussing with a llm? Never happened to me and I use llms all the time. Why would you need to discuss if you know best? Just tell it what to do and course correct it. It's not rocket science.

I find myself often discussing with an LLM when trying to find the root cause of an issue I'm debugging. For example, when trying to track down a race condition I'll give it a bunch of relevant logs and source code, and the investigation tends to be pretty interactive. For example, it'll pose a number of possible explanations/causes, and I'll tell it which one to investigate further, or recommendations for what new logging would help.

I find it exhausting in a yes-man kind of way where it does whatever you told but just somehow wrong. I think your human case is the reverse.

Just like a person would. If you want it done right you have to do it yourself. Or you have to tell the LLM exactly how to do it.

Often I find it easier to just do it myself rather than list out a bunch of changes. I'll give the LLM a vague task, it does it and then I go through it. If it's completely off I give it new instructions, if it's almost right I just fix the details myself.

In so many ways, LLMs are like that very energetic and confident Junior developer you hired straight out of college who believes he knows everything but needs to be course corrected constantly. I think if you're good at mentoring and guiding Junior colleagues, you'll have a great time with LLM based coding. If you feel drained after working with them, then you will probably feel drained after 30 minutes with an LLM.