> you sometimes want to under-specify or phrase things more indirectly to give a color to the implementation or elicit something creative

This has been my experience with most models. If you say "How do I do X? I was thinking maybe Y or Z" then the model will probably try to make Y or Z work. They will very likely not say some third option that is wildly different is better, even if it may be. And actually maybe less so with Claude because sometimes it pushes back.

Actually this seems like it would be an interesting test. Maybe I will come up with some contrived question and ask several models.