It should be possible to be even more prescriptive. E.g. write a "screenplay" (or even let GPT3 generate it) -- basically just a progression of tags -- and then pick random videos that match that.