There was an article on HN a few days back on how it’s very hard to convey all the context in your head about a codebase to solve a problem, and that’s partly why it’s actually hard to use AI for non-trivial implementations. That’s not just limited to code.

I don’t use AI for most of my product work because it doesn’t know any of the nuances of our product, and just like doing code review for AI is boring and tedious, it’s also boring and tedious to exhaustively explain that stuff in a doc, if it can even be fully conveyed, because it’s a combination of strategy, hearsay from customers, long-standing convos with coworkers…

I’d rather just do the product work. Also, I’ve self-selected by survivorship bias to be someone who likes doing the product work too, which means I have even less desire to give it up.

Smarter LLMs could solve this maybe. But the difficulty of conveying information seems like a hard thing to solve.