they are not good at consistently following best practices or architectural instructions. So you have to have some kind of hierarchical goal/context scope framework - But then the high-level goals actually need to be reasoned about, which LLMs don’t do, so efforts to make the framework analyze/plan/reflect In order to select and sub divide those top goals fail.

I have to fight with Claude to get it to just do three or four back-and-forth questions with me to establish the actual requirement instead of dumping 1000 lines of irrelevant code (And an MD document, and a usage guide, and an test suite) that ignores guidelines I had already given it.