generally agree, though i do use llm for code reviews, reference work (finding documentation or other knowledge, determining if some language has some feature that i know how to describe from implementation experience in another language), and ideation, and reasoning. ideation and reasoning are particularly rewarding if you're not a total dumbfuck, because llms like claude, chatgpt, and gemini will mold themselves to your capacity for reason, reflection, inrospection, etc. to varying degrees.

i've used claude to generate code a couple times. one was to perform some very technical debugging that is far outside the common python web development patterns, and one was to find a function in an attribute of an instance of a class. the function could be wrapped with n decorators, so particularly fragile runtime introspection had to be performed to dig into closures and cells. that latter isn't even documented, and if it weren't for me being open to being wrong in a code review i probably would have merged some tech debt time bomb into the code base.

i think of using these llms as hypothetical partners. if i want a hint of how i might implement something, i could ask, but i definitely don't want some llm changing code in the code base, or solving problems for me. solving problems is part of the fun of programming. the people who don't want to solve problems don't enjoy programming and aren't problem solves. they're just warm bodies that want to make programmer money without any of the wisdom or desire to improve at something they care about.