Hey, folks. Next turn of my experiments and ideas. Buckle up, it's gonna be wild this time.

We're used to defining alternative scenarios by swiping the starting messages. A problem is, that this way, we cannot place detailed instructions inside a scenario part of the character card - we often just write the general guidelines or leave the scenario field empty since nuances would have to change according to a starting message. Additionally, a starting message is always the same when we make it a scenario trigger - providing consistent formatting at the expense of repetitive re-plays of the same scenarios. What personally makes me roll my eyes is that we need to write the new card for another character if we want to roleplay literally the same scenario but with a different character in our deck - or we need to create a card with just the scenario for group-chats.

Thus, I started thinking - what if we could make "blank" characters - appearance, personality, quirks, goals, speech patterns and behaviors - only that aka the raw "avatars" without a scenario, without a starting message?

When you check on it, the prompts are just the formatted walls of text in the end - so where the actual "scenario" parts come from does not matter. Prompts always start with our system prompt, then whatever is in a {{char}} card (in the order specified by those UI boxes "depth"), then {{user}} persona and that's all - as simple as that.

Thus, I started experimenting. I cleaned up one of my character cards from all the scenarios and starting messages, then I described the scenarios I wanted to have - but not through the starting messages, not within a card itself - but in a lorebook, with detailed, specific instructions in a JSON object (string) format. I switched to the insertion order "after char", then I made the entries sticky for 10 000 messages (enough for any roleplay), I stopped the recursion and... checked how it works.

In short - it works perfectly. Not only I am able to keep my formatting from a system prompt intacts without giving it examples in a starting message, but I also get the scenario I want and it became even better. Really - better, not worse. Why? First, {{char}} is roleplayed better because its definitions become shorter - even a very detailed character definition ends up around 500-600 tokens so LLMs pick up the personality traits etc. better and roleplay it better. Less is more with LLMs. Secondly, scenario consists of very specific, detailed instructions to set it up, which is more direct instruction than a starting message. It is a set of instruction guidelines as scenario, not the ambiguous, prose written starting message. LLMs seem to understand it better. Third, starting message becomes different with each roleplay - sometimes it's peaceful, just as written in a scenario set-up, sometimes it goes wild. For example, once, a {{char}} threw a thug into my office and started the shootout, even though a scenario simply defined {{char}} entering my office. Experimenting with temperature works great here, keeps roleplay more grounded due to the detailed scenario guidelines in a prompt while allowing such things to happen through scenario interpretation by the LLM due to a higher temperature value than normally - when we want it, of course. We can keep the temperatures unchanged but the structured instructions for scenario make rising it possible without a consistency loss.

Last but not least, you can hot-swap both the scenarios through a lorebook and swap those "empty" characters while having access to your scenarios at all times. You can mix them together or change one scenario into another by triggering the keywords during roleplay. You could theoretically write different sub-scenarios, different sub-stories and trigger them half-roleplay or make it all RNG through the group entries rolling mechanism described in my other posts - to create the random "routes" that roleplay divides into - or to pick up the route at a point of your choice.

Take a look at how I'm doing it and at the results.

Lorebook template. It's important.

Character card - just to show it's empty with a generic starting message - only a detailed description of appearance, personality, outfits, goals, quirks, mannerism etc.

Example 1 - Scenario 1 - triggered through a defined keyword: SC1

Example 2 - Scenario 2 - triggered through a defined keyword: SC2.

How prompt looks like. It's literally the same as putting it all into a card. It's always a good idea to separate \"scenario\" as an object (with a label and brackets), even if we write it in natural language, not JSON. It still works, I've tried. So - in conclusion - it really works.

At the end, catch a great song about the AI and cyberpunk-like scenario we're already living in, from one of my favorite band :-P That's what I am listening to while coming up with those crazy ideas :-P

FIREWIND - Welcome To The Empire (2020) // Official Music Video // AFM Records