I didn't see the presentation but judging by your description, this is achievable using in-context learning.