I've found claude-code good in a multi-million line project because it can navigate the filesystem like a human would.

You have to give it the right context and direction — like you would to a new junior dev — but then it can be very good. Eg.

> Implement a new API in `example/apis/new_api.rs` to do XYZ which interfaces with the system at `foo/bar/baz.proto` and use the similar APIs in `example/apis/*` as reference. Once you're done, build it by running `build new_api` and fix any type errors.

Without that context (eg. the example APIs) it would flail, but so would most human engineers.

Well I have also worked on systems of multiple millions of lines, well pre-llm, and I sure as he'll didn't actively understand every aspect of it. I understood deeply the area I work on and the contracts with my dependencies as well the contracts I provide. I also understand the overall architecture. We'll see how it goes if my project grows to that point, but I believe by clearing documenting those things, and overall focusing on low coupling, I can keep the workflow I have now, but with context loading for every session. Time will tell.

In general though, its been a lot of learning on how to make llms work for me, and I do wonder if people simply dismiss too quickly because they subconsciously don't want them to work. Also "llm" is too generic. Copilot with 4o sucks but claude in cursor and windsurf does not suck.