Comment by vidarh

Claude has outright told me "this is getting tedious" before proceeding to - directly against instructions - write a script to do the task instead of doing it "manually" (I'd told it not to because I needed more complex assessment than it could do with a script).

There are fairly straightforward fixes, such as either using subagents or script a loop and feed the model each item instead of a list of items, as prompt compliance tends to drop the more stuff is in the context, but, yes, they will "get bored" and look for shortcuts.

Another frequent one is deciding to sample instead of working through every item.