Word on the street is that Claude Code is better than Github Copilot. I have no idea. Apparently it is also possible to use other (locally-run even open source) LLMs with this stuff. But yeah, it's impressive and makes obvious mistakes all at the same time. Naturally I find this easy to deal with since I've had a lot of students who can be both brilliant one minute and space cadets the next.
I'm figuring out how to use GNU Emacs Claude Code IDE, because emacs is in my fingers and I've heard good things about it. Anyway, upon consideration I think the "do all subjobs by kicking out little Makefiles and run $(MAKE) -f on them" would have a few advantages. First, it would make the concurrency a lot easier to control, since concurrent make processes handing a single job know how to talk to each other and coordinate on that. And second, if all the intermediate Makefiles were saved, that might make debugging easier. And it might remove the necessity to deal with lower-level constructs like forking etc.
