> Model capabilities are evolving so fast that what was true last month may not be true today I dunno. I think the harnesses are getting a lot better but older

> Model capabilities are evolving so fast that what was true last month may not be true today I dunno. I think the harnesses are getting a lot better but older models still do the bread and butter just as good. > Recently the head of Claude Code, Boris Cherny, said to delete all your AGENTS.md Claude.md files and just see what the models can do nowadays It was pretty obvious early on (which is only like a year ago) that you shouldn’t have much outside of things the model/person has no way of knowing inside your agents.md. Basically just a pointer to your docs. That OP didn’t know this as well as they listen to number one shovel salesman Boris makes them seem unserious. > My process for learning a new domain or technology used to be modeling it in code: build to learn This is still the process. My guess is OP is actually just an average / below average developer, so AI causes this reaction. I also have this reaction anytime I use AI in area I’m a novice in. You still need to learn domains deeply to unlock the ability to use AI effectively. Domain dependent, but if you’re not seeing AI make mistakes consistently you need to learn more about how the tech you’re using works. You’re missing things. The world is still full of ambiguity and tradeoffs that have no objective correct answer. I don’t think AI or humans have solved all these cases yet.