/vt/ - Virtual Youtubers

>>66736632
>>66765577
Other things to keep in mind:
>if you're unsure whether or not your current AI will understand a certain aspect of the bot's defs, you can ask it in OOC (example: (OOC: Claude/GPT, do you know what a reverse bunnysuit is? Could you describe it for me?)). Or you can make an actual AI Assistant bot for this sole purpose and go ask it whenever you have any doubts
>remember that different AIs have different capabilities and can behave in vastly different ways, so if you're using a dumber model and getting bad results it might not be your defs that are lacking
>you don’t really have to worry about token economy if you don’t care about smaller models, most of the bigger and more popular ones are going to have 10k+ context. Just don’t do anything too over-the-top like 2k+ tokens
>repeated info will stick to the AI better, hence the infamous looping issues, so if you casually mention something more than once in the defs the AI will be less likely to forget it
>because of how good AIs are at repeating things, the way you describe things in the defs can also influence the way some models write
>info that comes later in the description will also stick better
>in that same vein, if you REALLY want the AI to remember something or behave in a specific way, you can add it to the jailbreak override between parenthesis or brackets. Just make sure to add {{original}} and a line break before, or it will completely replace the user's jailbreak. And of course, remember that JBs are very powerful so don’t use this feature lightly and be careful with your wording, or the bot might just turn too repetitive or start acting in the same way regardless of the context or situation
>always use {{char}} and {{user}} instead of any actual names, remember to replace them if you're copypasting anything. {{char}} and {{user}} are respectively replaced with the card’s name and the user’s name before being sent to the AI, so don’t do stupid shit like “{{char}}’s name is {{char}}”
>for bots that are actually 2 or more characters in one card, you can naturally disregard the point above and refer to each different character by her respective name. You can put each one’s defs inside brackets or something for organization purposes or if you're afraid your AI might mix things up, this doesn’t really count as meme formatting
>pick one set of perspectives (example: third person when referring to {{char}} and second person for {{user}}) and stick to it on the greetings and example messages
>don’t write fucking main/nsfw prompts or JBs on the fucking description field, this should be obvious but I still see some cards telling the AI to drive the conversation forward and that violence and sex are allowed. Prompt overrides exist if you have to use them for special cards, but don’t use them for generic shit that everyone already has by default
>except for context stuff like "{{user}} is {{char}}'s husband/son/teacher/whatever", avoid describing {{user}} in the defs. If your card has a specific POV you can leave it implicit through the greeting(s) or warn the user through the creator's notes in the metadata. Or, if it's really important for the card, you can leave {{user}}'s description in a separate paragraph near the bottom
>do NOT under any circumstance describe {{user}}'s actions on the greetings or example messages (example: *You wake up in the forest and look around, the last thing you remember is getting run over by a truck. Seeing {{char}} in the distance, you walk up to her and...*). This is the #1 cardinal sin of botmaking, as it WILL encourage the AI to keep acting and sometimes even speaking for the user, which everyone rightfully fucking hates. Instead, if necessary, try to leave {{user}}'s situation implicit through {{char}}'s POV (example: *Walking through the forest, {{char}} finds you laying on the ground. Her curiosity piqued by your strange clothes and dazed expression, she walks up to you and...*)
2/2