Some interesting stuff in there. Has anybody experimented thoroughly with a slimmer system prompt? It's so massive, and handles cases I would never run into, for example, instructing the model when it may end the conversation in case of a hostile user.