This is a special post for quick takes by Dalcy. Only they can create top-level comments. Comments here also appear on the Quick Takes page and All Posts page.
I thought it was pretty hard to come by relatively time-bounded (on the order of few hours max) exercises for the skill of coming up with good concepts,[1] but it seems like stuff from recreational mathematics is potentially a good source for them, eg "find a simple rule for deciding the winner in a game of Green Hackenbush," which will involve inventing (as long as you're not already spoiled!) a bunch of new concepts like how to chunk a given game graph. GPT[2] also recommends Domineering, Chomp, Sprouts.
Though I haven't tried them yet[3], this is possibly better than 1) [learning new boardgame] since there is a pretty object natural / objective answer / concept to be discovered for the task (and also easier to evaluate / credit assign success, compared to winning a boardgame), 2) other suggestions I heard like Baba is You which seemed less about concept formation, and 3) much more time-bounded than various versions of "reinvent a field X from scratch".
For purposes like, 1) training your object-level concept formation skill / taste by drilling such exercises, 2) doing it more deliberately by explicitly recording thought traces via introspecting while doing such exercises and reflecting on the cognitive strategies that come up, and 3) the same as 2 but for getting data on agent foundations-ish questions, and 4) doing all of these myself, or as exercises in workshops or research intensives for others.
Be careful with getting spoiled though, eg GPT is sooooo eager to explain & give hints to the exercise at every opportunity despite strongly worded instructions against it. I recommend just asking for problem description and nothing else.
Playing around with pergraphs has made me better at coming up with good concepts. (Even though it is redundant with automata theory, but I don't know automata theory that deeply.)