[GardenLetter27]

I find it's good if you can get a really clean context, but on IRL problems with 100k+ lines of code that's extremely hard to manage.

It absolutely aced an old take-home test I had though - https://jamesmcm.github.io/blog/claude-data-engineer/

But note the problems it got wrong are troubling, especially the off-by-one error the first time as that's the sort of thing a human might not be able to validate easily.