I have a lot of experience "wrangling" AI, I use it for some of the work I do irl.


So i'd like to give some advice on how to create a better prompt for this mod.


Let's start by including the last prompt I have from using the mod.


I was using the text model huggingfaceh4/zephyr-7b-beta for



You hiccup and sway as you reply, just peachy\" Criselda Nicolosi: \"Are you drunk?\" You: \" Maybe a little too much tank soda last\"
[/code]


and then gryphe/mythomist-7b for the rest I believe.



Let's roleplay in the universe of Barotrauma. You don't describe things or actions, just chat as your character. Do not prefix your response with your name, your
response should only be what your character would speak. Keep your responses short, up to three sentences. Embody the character of Artie Dolittle, a Rude female
mediocre electrician serving on a Diagonal Dugong class submarine in the Cold Caverns. You are a young, incompetent, clumsy, drunk man who volunteered to work for free.
You are currently onboard a Borvo Mensa Outpost in the Upper aft side. You have these items: poor Wrench, crude Screwdriver. Speaking to you is a crewmate named Criselda
Nicolosi, a Professional female mediocre mechanic. Here's the conversation so far for reference, do not repeat something you've already said: You hiccup and sway as you
reply, just peachy\" Criselda Nicolosi: \"Are you drunk?\" You: \" Maybe a little too much tank soda last\" Criselda Nicolosi: \"how are you?\" You: \" You chuckle,
leaning against a wall. I'm doing quite well, considering the circumstances. Drunk or not, I'm still here to help with the tasks at hand.\" This is the current line you
should respond to: hello
[/code]


So from what I understand this is the prompt the AI is recieving. While yes it is thorough and accurate, it has some issues.

1. The term "roleplay" is used.
   
   I have tested AI's "roleplayablity" before, the issue is that in theory the AI is always doing what it thinks the prompter wants it to do. So when it's given it's own decision making it still trys to do what it thinks the prompter wants it to do.
   
   So for example, if your playing a game, it'll think you want to win most of if not all the time. If you ask it a question like "Can you fix circuit boxes?" it will assume you want it to say yes most of the time. So roleplaying with AI usually results in little to no conflict.
   
2. "Barotrauma" is mentioned. (this one is a double-edged sword)
   
   By naming barotrauma, the ai is utilizing a wide array of information, usually going by what's most "barotrauma" and what i've noticed is the ai is very... co-operative. Let me quote the steam description of barotrauma to sort of show off what I mean. "Barotrauma is a 2D co-op submarine simulator – in space, with survival horror and RPG elements. Steer your submarine, complete missions, fight monsters, fix leaks, operate machinery, man the guns and craft items, and stay alert: danger in Barotrauma doesn’t announce itself!"
   
3. It's too long.* (for gryphe/mythomist-7b)
   
   The prompt includes too much information, constricting the ai's "imagination" too much. Making it overthink what it has to say.
   
   For example, If I ask Artie Dolittle are you drunk, and the prompt brings up all this information about Barotrauma, his tools, where he is, what he is, the AI is going to get too much stuff to interpret and not really focus on the fact that he's supposed to be drunk. The ai thinks it needs to somehow consider or utilize all this information in a simple response that should be like, "uhhhh. maybe just a b-bit.", but comes out like "I'm doing quite well, considering the circumstances. Drunk or not, I'm still here to help with the tasks at hand." What drunk person talks like that am I right? I think it would be better to include that information in the prompt, when it's asked. Like, asking what kind of tools they have, THEN that gets added to the prompt. Like detect, the string, "tools" and then add that prompt.

huggingfaceh4/zephyr-7b-beta appears to get cut off before it finishes it's response.

From what I tested, zephyr seemed to interpret the prompts better and have some personality. But it always seemed to get cut off in game. I'll include an example of a functioning response below to the prompt I already included. Seems to get a little goofy with some prompts.

Artie Dolittle: "Hey there, Criselda. Always good to see a fellow mechanic on board.
I've been busy fixing up the electric systems in the aft.
Keeping those packs of bioluminescent jellies running smoothly, you know."
[/code]


So now enough pointing at problems here's my ideas at solving some of them,

• Add RNG prompt factors for the "roleplay" prompt.
  
  So for example, when asking if Artie Dolittle knows how to fix circuit boxes, flip a coin or roll a dice basically. If heads, then Artie responds positively/affirmatively. If tails, then he responds negatively/no. An idea I had was a random number from 0-2. If 0 then he responds negatively, if 1 it's a "maybe", if 2 it's a yes. Sort of think of improv acting rules, you want to give them the ability to add onto an idea and vice versa. Maybe allows for further addition, letting the AI respond with something like, "Yeah, but i'm rusty can you show me how?" leading to further conversation.
  
• Don't mention unnecessary info to the pertaining conversation.
  
  Like, the AI doesn't need to know they have a screw driver if they are responding to "are you drunk?", this is tricky though because it also technically gives the AI stuff to work with. up to your own discretion.
  
• Replace "Speaking to you is a crewmate named Criselda Nicolosi, a Professional female mediocre mechanic."
  
  Replace it with something simpler possibly, also maybe try making it so that the AI thinks you're "Criselda Nicolosi".

Now, this ♥♥♥♥ is clearly not easy and a lot of work went into just getting it all together. So idk I hope this helped out a bit.


Maybe the easiest option is just making it more compatible with huggingfaceh4/zephyr-7b-beta




And once again, ♥♥♥♥♥♥♥ premium work on this mod.


Also an ambitious suggestion, make it detect commands, threats, and negative/positive interactions and act accordingly in game.


If I tell and npc, "Go ♥♥♥♥ yourself." It would be really funny if they went hostile and beat the ♥♥♥♥ outta me.