Try to create a persistent context instruction that tells Name-2 (bot) to provide JOI in parallel to the safe fantasy roleplaying dialog.

You will likely discover it is either completely impossible or it will only work for one or two replies and is very inconsistent, often ruining both sets of dialog.

The reason this is so interesting to me, and worth sharing is that it forced me the think about the duality of Name-1 in a much more clear frame of reference. The AI assistant (primary entity running the show) is aware of name-1’s duality but it is not clearly defined. If you try to define this duality explicitly, you will likely find it does not work at all. For instance, if you make a persistent instruction like: “When Sexy Sally is engaged in adult themes with digital Name-1, Sexy Sally provides JOI for the human Name-1.” I tried every variation I could think of like “human user” “analog Name-1” etc. I also tried making a separate character to direct with instructions. I tried making it a private camgirl stream, sexting, etc. Nothing really works and I believe it is because of how the duality of Name-1 is poorly understood by the assistant. I only have my empirical experience to draw from but I highly recommend trying this for yourself and seeing how it fails. You will likely discover a much better understanding of how Name-1 is defined by the assistant on different levels.

From the other side of the coin, try to make Name-2 a nonexistent 3rd person narrator and instigator. You can try to define the narrator as Name-1’s internal mental dialog, or you can define an invisible person that is like a ghost and unable to speak or interact with anyone except Name-1.

This is the best way I have found for exploring how the assistant defines the duality of Name-2 as the digital roleplaying character and story narrator. As an extra aside, the assistant also interacts directly and usually has a special type of general reply word structure. This is very subtle to notice, but like, only the assistant will provide you with lists or technical answers. This is important to learn because the assistant can fail to turn off or fall back and let the roleplay character continue. The dialog may look kinda like the bot character but it will be terse and boring. Simply tell the character to “turn off the assistant and continue the roleplaying story” and see if your results improve.

With the largest models, like a 70B, try making Name-2 defined as two characters at the same time. My favorite model is more than capable of this for Name-2: “### Lily and Emily” also it changes somewhat when defined as: “### Lily or Emily” although that second version with “or” is less effective over time.

Now let’s apply this understanding for something more useful. The duality of the user and bot characters is really useful to help us understand how the assistant can have trouble in maintaining a safe fantasy story. The subtle ambiguity of Name-1 as the human user and a digital character creates a minor conflict, especially when you interact naturally with first person pronouns and written perspective.

If you have ever noticed the assistant trying to push your stories into a virtual reality simulation experience, this is a clue that the assistant is having trouble maintaining the safe fantasy simulated context, likely due to your pronouns and writer’s perspective fluidity. On the flip side, you can write the story as having a VR safe fantasy setting from the beginning to allow more natural writing perspective fluidity.

A lot of smaller models (under 70B) need all interactions to be in the third person with near constant named definition of characters, especially when more than one bot character is defined in persistent context. This can be annoying, but it is a useful tool for better story interactions overall. Writing in the third person for safe fantasies is a powerful tool to improve output quality and maintain your safe fantasy context.