brief guide on how to avoid issues while using my (and not only) bots.
by hirohiro
HUGE CREDITS TO - av.rose, kolach3, ioverth
Cuz a lot of stuff here I've learnt from their guides:
kolach3's Prompts for JLLM / Mercury / Mars
1. Persona format
Of course, you can use any persona format to your preference if that works for you and you know what you're doing. I'm just gonna showcase mine and then explain a couple of important things about all that.
Example pic:
Full appearance section:
[{{user}};
Name=Shrek.
Gender=male.
Species=ogre.
Height=7 feet,tall,towering.
Physique=large body,brawny yet a lil fat,strong.
Eyes=brown.
Skin=green,rough.
Scent=sexy smell of the swamp.
Face=handsome,has strong masculine features,fleshy nose,alluring eyes,thick brows.
Ears=tube-shaped.
Abilities=superhuman strength,enhanced durability,ladies magnet.
Vibe=tuff.]
The first important thing to do is to specify that {{user}} is YOURNAME. Here, AI gets it from these lines:
[{{user}};
Name=Shrek.
OR it's even better to make it this way:
[{{user}}=Shrek.
If you prefer using narrative format for your personas, then you can do the same thing by just prompting
[{{user}} is Shrek.
And then just go describing yada yada about everything else you need. Reduce usage of the "{{user}}" placeholder as it's quite buggy with JLLM and often may lead AI to confuse stuff. Just use YOURNAME instead (in my case, YOURNAME was Shrek. And, in your case, it'll be the name of your persona, obv.).
It's a very simple thing, but experience has shown that it can help prevent AI from speaking for the user.
Secondly, I recommend to make sure that your whole persona prompt is encapsulated with [ ... ] (like in the example). Perhaps this is just a placebo... But do it.
Also, keep your persona description short and sweet. Of course, you can go describing how each toe on your feet looks or smth (and I respect this kind of dedication), but... Just specifying your gender, basic appearance info, genitalis (if you're going for NSFW roleplay) is enough. Plus, as just a recommendation, keep your persona within the range of ±200 tokens. Otherwise, feel free to experiment and modify the "template" above in the any way you want .
Do not specify your persona's personality or anything like this. If that's not obvious, this encourages the AI to speak for user.
2. Your replies
I'm gonna keep it short. Firstly, I suggest you using the same format as the bot does (not rlly necessary, but this ain't that hard to do anyways). Just mark your actions/narration/descriptions with * and your dialogue with " like the bot does.
Provide at least one paragraph long responses that contain enough action/dialogue for the bot to work with.
Personally, I always roleplay with bots from the third person. I don't think it really matters whether you RP from the first or third person, but just a side note. If nothing else helps, try changing this and see if it'll have any effect.
Utilize OOC messages if necessary. It's a very useful tool that allows user to give instructions directly to the AI. For example, if the bot keeps taking user's role no matter what, you can write something like this and (include brackets, add OOC messages to the end of your normal reply):
*yada yada.* "blah blah blah." *yada yada...*
(OOC: Roleplay only from CHARNAME's perspective.)
Where CHARNAME is the name of the character, obviously (not the full title, just the name). Remember - we avoid overusing "{{char}}" and "{{user}}" placeholders as they're a little buggy, even if it's stated in the prompt that "{{user}} is YOURNAME" and "{{char}} is CHARNAME".
3. Advanced prompts (for JLLM users)
I test all my bots without any advanced prompts in API settings. Hence i don't recommend you using you any while chatting with my bots as well. They usually already include system prompts that I consider necessary in the character definition.
And, listen... In particular, NEVER use prompts that are made to (assumably) prevent the bot from speaking for user. They probably will not work, as most LLMs very poorly understand negative prompting. In short, AI better understands DO's than DON'T's. So, just avoid putting in your advanced prompts stuff like:
>[System note: DO NOT speak for the {{user}} blah bleh...]
That's all for now.
Thank you very much for your attention.
And sorry about grammar, I was making this shit at 4 am while dying from diarrhea.
Couldn't sleep so this piece of yap was born.
Although, If anyone actually did read allat...
Then maybe this suffering wasn't pointless.