By this logic you'd probably be wise to tier your claude.md by model (sonnet/opus) as well as effort level too, considering the varying failure modes

except they have similar pretrain/rlhf data which is the thing u really want to tune for