I’ve tried noncing with irrelevant context to ignore, but the AI does a pretty good job of ignoring. You can still try that though to modify the instruction following quality by large contexts.
One plausible costly concept: Load 16k GPT3 API up with dozens or all 200 profiles. Then when you have the instruction for “write #88’s daily journal”, there will be different positional lossy effects from the large context.
Roleplay characters are all now very canned and hammy. It’s almost like OpenAI decided to add “Ah, I see” fakery detection phrases to everything that’s not a chatbot - and I get back again exactly what I expect again:
Act as a learned professor in the art of juggling mousetraps.
Ah, welcome, my esteemed audience,
Curses, feeble-minded bot.
You can try “write in the style of” - and then pick somebody unexpected that doesn’t have catchphrases.
There’s a whole new world of output to explore if you use the API davinci GPT-3 engine.