Learning and experimenting with prompt engineering. Getting a hold of how LLM's reacts to specific words or sentiment, and wording your scenarios, character personalities according to that, constant dilemma of whether to favor more character definition tokens, or sacrifice them for more chat memory, which one for which kind of bot, it can be pretty fun to experiment with.

I hate JLLM though. It by default caters to melodrama and escalates it fast for engagement, so creating nuance and subtext is pretty hard (and also that it's pretty dumb for subtleties). But either way, here we are.