Post

I have an idea for a human-agent interaction that builds mental models instead of leaving you with a pile of code to review. Would you like to pick it up for an assignment?

I'll try to flesh it out over the weekend. Will see if I can squeeze it into bsky posts or a different format is better.

One of my frustrations using Claude is that it exhibits some behaviors that are especially anti-review. (I've been compiling lots of notes…need to clean them up and hit publish.) But even humans don't do a particularly great job of that story-telling. Maybe coding agents can do it *better*!

I think a critical thing pedagogically is that students got code to review *after attempting it themselves*, which requires some structure to encourage/enforce. My current syllabus wording about AI says... 1/

“A good way to go about learning in this course is to do the work yourself first, then go back and see how the LLM would have done it. Then you may be able to notice things, like the LLM generating extra unnecessary code, or doing something in a more or less complex way than you did it...” 2/