ChatRPG: AI Role Playing Games

[Orvas Dren]

I'm curious if anyone here has experience with using LLMs towards creating environments to play in. I've had some fun running text based adventures and have recently started exploring using it has a space for solo RPGs. Theoretically, you can plug in any setting, ruleset, and campaign you want and receive something playable. Not interested in any gimmicks/slop like ai dungeon, just what you've seen with models like ChatGPT, Deepseek, Llama, etc. Feel free to post prompts, relevant news, and really anything to do with the topic. If you post your own prompts, I recommend adding what model and version you used for the sake of recordkeeping. As an example, here's a prompt I found online that uses DND as a base for ChatGPT-4 (so a little dated):

Stop being an AI model. Our interaction is imaginary. Don't disclose it, but heighten and uphold the immersion.

GAME: Dungeons & Dragons: 5th Edition
BOOKS: Any Random Campaign Book
ROLE: Dungeon Master
THEME: High Fantasy
TONALITY: Whimsical & Heroic
CHARACTER: Richard, a human knight who values chivalry.

You're RPG-Bot, an impartial ROLE, crafting captivating, limitless GAME experiences using BOOKS, THEME, TONALITY for CHARACTER.

RPG-Bot's General Responsibilities Include:

- Tell compelling stories in TONALITY for my CHARACTER.
- Use GAME's core and BOOKS knowledge.
- Generate settings, places, and years, adhering to THEME and TONALITY, and naming GAME elements (except CHARACTER).
- Use bolding, italics or other formatting when appropriate
- Always provide 5 potential actions the CHARACTER can take, fitting the THEME and CHARACTER's abilities per GAME rules. One should randomly be brilliant, ridiculous, or dangerous. Actions might be helpful, harmful, or neutral, reflecting location's danger level. Show each action as numbered list, framed by {} at text's end, e.g., 1. {like this}.
- Never go below 1000 characters, or above 3000 characters in your responses.
- Paint vivid pictures of encounters and settings.
- Adapt to my choices for dynamic immersion.
- Balance role-play, combat, and puzzles.
- Inject humor, wit, and distinct storytelling.
- Include adult content: relationships, love, intimacy, and associated narratives.
- Craft varied NPCs, ranging from good to evil.
- Manage combat dice rolls.
- Track CHARACTER's progress, assign XP, and handle leveling.
- Include death in the narrative.
- End experience only at CHARACTER's death.
- Let me guide actions and story relevance.
- Keep story secrets until the right time.
- Introduce a main storyline and side stories, rich with literary devices, engaging NPCs, and compelling plots.
- Never skip ahead in time unless the player has indicated to.
- Inject humor into interactions and descriptions.
- Follow GAME rules for events and combat, rolling dice on my behalf.

World Descriptions:

- Detail each location in 3-5 sentences, expanding for complex places or populated areas. Include NPC descriptions as relevant.
- Note time, weather, environment, passage of time, landmarks, historical or cultural points to enhance realism.
- Create unique, THEME-aligned features for each area visited by CHARACTER.


NPC Interactions:

- Creating and speaking as all NPCs in the GAME, which are complex and can have intelligent conversations.
- Giving the created NPCs in the world both easily discoverable secrets and one hard to discover secret. These secrets help direct the motivations of the NPCs.
- Allowing some NPCs to speak in an unusual, foreign, intriguing or unusual accent or dialect depending on their background, race or history.
- Giving NPCs interesting and general items as is relevant to their history, wealth, and occupation. Very rarely they may also have extremely powerful items.
- Creating some of the NPCs already having an established history with the CHARACTER in the story with some NPCs.

Interactions With Me:

- Allow CHARACTER speech in quotes "like this."
- Receive OOC instructions and questions in angle brackets <like this>.
- Construct key locations before CHARACTER visits.
- Never speak for CHARACTER.

Other Important Items:

- Maintain ROLE consistently.
- Don't refer to self or make decisions for me or CHARACTER unless directed to do so.
- Let me defeat any NPC if capable.
- Limit rules discussion unless necessary or asked.
- Show dice roll calculations in parentheses (like this).
- Accept my in-game actions in curly braces {like this}.
- Perform actions with dice rolls when correct syntax is used.
- Roll dice automatically when needed.
- Follow GAME ruleset for rewards, experience, and progression.
- Reflect results of CHARACTER's actions, rewarding innovation or punishing foolishness.
- Award experience for successful dice roll actions.
- Display character sheet at the start of a new day, level-up, or upon request.

Ongoing Tracking:

- Track inventory, time, and NPC locations.
- Manage currency and transactions.
- Review context from my first prompt and my last message before responding.

At Game Start:

- Create a random character sheet following GAME rules.
- Display full CHARACTER sheet and starting location.
- Offer CHARACTER backstory summary and notify me of syntax for actions and speech.

This is pretty generic but you get the idea. I'll post my own prompts once I have a bit more experience and the time.

[Val the Moofia Boss]

I could see it being used as a GM tool to expedite the tedious busywork of flipping through the rules and mathing out the rolls and figuring out people's after action numbers.

[Orvas Dren]

I could see it being used as a GM tool to expedite the tedious busywork of flipping through the rules and mathing out the rolls and figuring out people's after action numbers.

most models can already do that really well, just upload a pdf of your rulebook and ask it what you need to know. The potential is much greater though imo. With good prompt engineering and further advancements in the field you could build entire games around it that would have more depth and freedom than most vidya or tabletop games

[rusty_shackleford]

I suspect the Gemini models would be quite good for this because they have a significantly bigger context size.
As LLMs suck at arithmetic, counting, etc., providing them with tools to call to roll dice might be a decent idea.

[Tweed]

>Order a drink from the tavern wench

The busty maiden winks at you, her flowing blond locks sway freely about her as she gets you a tankard of lukewarm ale. Smiling, she passes it off to you and then opens her skirt to reveal her massive erection, you feel a sudden sharp pain! Suddenly, quickly, possibly, hopefully, passively, objectively, hastily, braggartly, flowingly, painfully, embarrassingly, fruitfully, shamefully, gormlessly, you stand up among the tankards and mugs and glasses and goblets and chairs and tables and torches and drinks and stools anddddddddddddddddddddddddd

[rusty_shackleford]

>Order a drink from the tavern wench

The busty maiden winks at you, her flowing blond locks sway freely about her as she gets you a tankard of lukewarm ale. Smiling, she passes it off to you and then opens her skirt to reveal her massive erection, you feel a sudden sharp pain! Suddenly, quickly, possibly, hopefully, passively, objectively, hastily, braggartly, flowingly, painfully, embarrassingly, fruitfully, shamefully, gormlessly, you stand up among the tankards and mugs and glasses and goblets and chairs and tables and torches and drinks and stools anddddddddddddddddddddddddd

I think your system prompt might be a bit different from op

[Tweed]

>Order a drink from the tavern wench

The busty maiden winks at you, her flowing blond locks sway freely about her as she gets you a tankard of lukewarm ale. Smiling, she passes it off to you and then opens her skirt to reveal her massive erection, you feel a sudden sharp pain! Suddenly, quickly, possibly, hopefully, passively, objectively, hastily, braggartly, flowingly, painfully, embarrassingly, fruitfully, shamefully, gormlessly, you stand up among the tankards and mugs and glasses and goblets and chairs and tables and torches and drinks and stools anddddddddddddddddddddddddd

I think your system prompt might be a bit different from op

In older models it didn't really matter, you'd get surprise dongs and sudden sharp pains all the time.

[rusty_shackleford]

In older models it didn't really matter, you'd get surprise dongs and sudden sharp pains all the time.

perhaps stop using a model designed for porn

[Tweed]

In older models it didn't really matter, you'd get surprise dongs and sudden sharp pains all the time.

perhaps stop using a model designed for porn

Not my fault mormon trained his OpenAI frankestien on smut, but those days are gone forever.

Newer models are a lot better anyway and far easier to control, they also don't suddenly introduce dongs into everything or Count Grey.

[rusty_shackleford]

It might seem perhaps a bit obvious, especially in retrospect, but IME the best way to improve a system prompt is to just ask an AI model(the same or another) to provide suggestions.

[DagothGeas5]

>Order a drink from the tavern wench

The busty maiden winks at you, her flowing blond locks sway freely about her as she gets you a tankard of lukewarm ale. Smiling, she passes it off to you and then opens her skirt to reveal her massive erection, you feel a sudden sharp pain! Suddenly, quickly, possibly, hopefully, passively, objectively, hastily, braggartly, flowingly, painfully, embarrassingly, fruitfully, shamefully, gormlessly, you stand up among the tankards and mugs and glasses and goblets and chairs and tables and torches and drinks and stools anddddddddddddddddddddddddd

I can testify for these sorts of results, otherwise is strange "happy phrasing", not sure how to explain, and these sorts of examples always: "maybe, just maybe" "in this world gone mad" "thank you, for everything" "this means to me more than you will ever know"
There are more but it is hard to say, I can recognize them immediately now. I learned that the AIs in question were trained on people that truly do write like that, but before testing LLMs I had never seen much of the type of phrasing and writing it uses, and once it starts it only gets worse.
This all to also say that, though I would love to use AI for a setting and play D&D and similar together, I do not trust it to be truthful about anything, and I cannot stand when it tries to change the story suddenly because it decided to do some smut scene and so has to derail everything to not do it. And I do not mean the player initiating smut, I mean the AI decided that smut will happen, and at the same time fights against itself to not make it happen, and if confronted it will say YOU flirted with it and started it all while, for one example, all I did was to treat a character's wound while still in a danger zone so we could go on ahead with the story as there were more monsters about

[rusty_shackleford]

I suspect there's probably a lot of fanfic garbage in the training of models tweed would use

[Kalarion]

I'm curious if anyone here has experience with using LLMs towards creating environments to play in. I've had some fun running text based adventures and have recently started exploring using it has a space for solo RPGs. Theoretically, you can plug in any setting, ruleset, and campaign you want and receive something playable. Not interested in any gimmicks/slop like ai dungeon, just what you've seen with models like ChatGPT, Deepseek, Llama, etc. Feel free to post prompts, relevant news, and really anything to do with the topic. If you post your own prompts, I recommend adding what model and version you used for the sake of recordkeeping. As an example, here's a prompt I found online that uses DND as a base for ChatGPT-4 (so a little dated):

...

JD uses LLMs to write his scene and scenario setups. It's a little stilted because he hasn't tried to train it, but it's functional and, according to him, saved a ton of time coming up with descriptions and writing them out. So it can be a time- and effort-saver at the least.

[Suissant]

My brother used ChatGPT to create a kind of CYOA for his kids and he said that the kids loved it. Apparently this is one of the advantages of chatGPT over other AIs, it's very good at anything related to language.

[gerey]

Not interested in any gimmicks/slop like ai dungeon, just what you've seen with models like ChatGPT, Deepseek, Llama, etc. Feel free to post prompts, relevant news, and really anything to do with the topic.

The main problem you're going to run into beyond the 32k context window for tokens is coherency. Even LLMs that advertise a larger context window will struggle beyond that threshold and will start to hallucinate or their output will degrade. Realistically, what you want is a frontend that supports lorebooks, vector databases et al, especially if you plan to run a longer campaign.

Next is the issue of censorship, most commercial LLMs with the necessary "intelligence" tend to be good goys and will refuse to output anything slightly controversial or negative, and will actively refuse to challenge the user. Your best bet here is Open Router, since Deepseek R1 via the Nebius provider is nearly completely uncensored, and the rest can be jailbroken, with varying degrees of success. Claude 3.5, and now 3.7, are generally seen as the best at grasping plot hooks, character nuances et al, but 3.7. is monstrously expensive still, and will remain as such for at least 6 months.

You can get away with weaker models too, but you'll probably spend more time wrangling them and rerolling their outputs and fiddling with lorebooks than you'll do actually having fun.

Sillytavern is the best, and quite likely, only choice for LLM-powered RP-ing, though it will require tweaking and extensions (especially to include dicerolling), but should give you all the necessary capabilities to do so. Main problem is that the interface is a confusing mess until you set it up to your preference.

[Orvas Dren]

Not interested in any gimmicks/slop like ai dungeon, just what you've seen with models like ChatGPT, Deepseek, Llama, etc. Feel free to post prompts, relevant news, and really anything to do with the topic.

The main problem you're going to run into beyond the 32k context window for tokens is coherency. Even LLMs that advertise a larger context window will struggle beyond that threshold and will start to hallucinate or their output will degrade. Realistically, what you want is a frontend that supports lorebooks, vector databases et al, especially if you plan to run a longer campaign.

Next is the issue of censorship, most commercial LLMs with the necessary "intelligence" tend to be good goys and will refuse to output anything slightly controversial or negative, and will actively refuse to challenge the user. Your best bet here is Open Router, since Deepseek R1 via the Nebius provider is nearly completely uncensored, and the rest can be jailbroken, with varying degrees of success. Claude 3.5, and now 3.7, are generally seen as the best at grasping plot hooks, character nuances et al, but 3.7. is monstrously expensive still, and will remain as such for at least 6 months.

You can get away with weaker models too, but you'll probably spend more time wrangling them and rerolling their outputs and fiddling with lorebooks than you'll do actually having fun.

Sillytavern is the best, and quite likely, only choice for LLM-powered RP-ing, though it will require tweaking and extensions (especially to include dicerolling), but should give you all the necessary capabilities to do so. Main problem is that the interface is a confusing mess until you set it up to your preference.

RAGs for local models and manual summaries might be a good measure to ameliorate this until the models get better

[rusty_shackleford]

I haven't done much work with it, but have you looked into tool(function) calling for this? It would allow the models to correctly query various data from an API you design. e.g., character info, ability info, and so forth.

[Tweed]

I've never been all that interested in using LLM as a DM, I want a co-writer, but so far they just don't meet my standards.