What I find helpful in a large project is whenever Claude goes way off the rails, I correct it, and then tell it to update CLAUDE.md with instructions in its own words how to not do it again in the future. It doesn't stop the initial hallucinations and brainfarts, but it seems to be making the tool slowly better as it adds context for itself.

Has this got anything to do with using a stronger typed language? I've heard that reported, not sure whether it's true since my python scripts tend to be short.

Does it end in a forever loop for you? I used to have this problem with other models.

I can second that. Even on plain CRUD with little to no domain logic.