This is where I’m aiming for now. We’re not going to put away our power tools, but we also should not be using a circular saw to carve inlay panels.

LLMs are absolutely capable of writing beautiful code, the important step here is telling it what beautiful is. You need to have an opinion and describe it articulately to an LLM. You need to be a Rick Rubin of code; he plays no instruments, can’t read sheet music, but his opinions are what matter.

I have been trying to perfect a workflow where the human is engaged and doing what they love and the LLM is kept away from what it’s bad at. This is essentially a pair programming setup where the human is in control but will pass the conch to the LLM from time to time when they don’t wish to type something, the rest of the time you should be asking it questions.

We will need to start thinking about our human engagement as a resource that is expended on a dial. I have a fixed amount per day and it doesn’t need to be evenly distributed, but it needs to be dense in areas that “matter”. What matters is subjective and where your Rick Rubin skills come in. The metric of quality is the difference in the distribution you have decided and a “true” distribution of what matters. So yes, maybe I don’t need to fully understand all the code for some control panel I’ve built or a blob of test code the LLM spit out, but I absolutely need to know and have an opinion on how the core machinery of the codebase works.