... This worked very well for a while, until I got a cohort who were already *deep* into LLMs who would encounter the errors, not wait for the rest of the class to catch up, paste the message into ChatGPT, and set out on confusing, frequently completely wrong paths that it was difficult to bring them back from (2/2)

Argh, this sounds very familiar, I've been trying to teach *from the very start* that encountering errors is useful - they're opportunities to learn, and to think / design in successively more detailed levels of abstraction - working "outside in" - calling methods which don't exist yet to sketch out the general shape of a procedure, watching it fail, then defining them... (1/2)