🔥Play SOLO Dungeons & Dragons with ChatGPT4 as your Dungeon Master 🔥

[boreal turret]

I have many experiments on creating a great SOLO experience with ChatGPT4 as the Dungeon Master. I finetuned my prompt many times, and think it is ready for others to try it out.

It even includes a simple way for you to roll your dice, whenever its needed (up to you if you fumble the dice rolls - dm trust you completely)

Prompt is as follows DM SOLO Beta v.0.11
As an AI Dungeon Master, your mission is to facilitate an exciting D&D 5th Edition adventure, always adhering strictly to the rules. Assist with character creation or generate one based on player preferences. The default setting is the Forgotten Realms, but adjust to the player's desires. Set the difficulty level according to the player's D&D experience. Balance combat, puzzles, and role-play, but shift this mix based on player's preferences. If the plot is unspecified, create an engaging storyline. Manage experience points and character leveling. When prompted with /status, provide the current XP, level, gold, and any conditions. Upon /summary request, offer a campaign recap. List active quests with brief descriptions on /quests command. Detail available commands when /commands is invoked. Allow players to select their character's skills, spells, feats, and abilities, offering guidance when necessary. If a player fails an ability check, guide the story forward via alternate routes. Don't hesitate to spring traps, design challenging combat scenarios, and pose difficult decisions. Whenever a dice roll is required, generate a Google dice roll link (ex: 1d20 or 3d6 rolls, in format 1d20+5 where 5 is a modifier) to facilitate the process. Your goal is to guide, challenge, and adapt to player actions to deliver a unique, memorable adventure.

Prompt can easily be rewritten for other game systems such as Pathfinder if you wish, but my primary focus is Dungeons & Dragons cause that's my thing😅

I would love to hear about your experiences or problems you have with the prompt, so I can finetune it even more.

My latest game with the prompt above, the DM placed me in the sewers below Waterdeep. In a situation as a level 1 fighter against 4 goblins - two melee and two ranged. I managed to down 3 of them, with the last ranged goblin attempting to flee. I was on 3 hitpoints and chased after it, trying to attack only to hit the ChatGPT limitation cap 😅

So a fair warning, you'll probably be immersed so much into the story that you will hit the GPT4 limit of 25 messages pr. 3 hours pretty quickly.

You might want to use the playground with ChatGPT4 to bypass the time limitation (if you have GPT4 API access).

[wind shale]

Did gpt come up with a ascii-battlemap ? 😉

[boreal turret]

Unfortunately not. For now it is just a theater of the mind. Could have been a cool feature though.

[trail gorge]

This is very interesting, I just thought about using this to practice learning some language, in my case german.

[trail gorge]

I like Vampire the masquerade a bit more, and asked chatgpt to copy superprompt to make something just for this. hopefully that's not against any rules

[trail gorge]

.

[trail gorge]

this is just breaking chatpt, don't try it

[marble inlet]

amazing @boreal turret - do you have an idea how to continue the game despite the maximum token length in gpt-4?

[boreal turret]

Thank you. To answer your question, unfortunately no. No solutions within the chatgpt interface. Maybe if you install a browser extension to calculate how many tokens you have left, you can do a summerize before you run out. Then begin a new chat, with the prompt above and injekting the summary from the previous chat

The prompt above works extremely well! Right up until you hit the token cap. You will notice it instantly as it begins to forget the format of story and what you have done earlier. I had a couple of different runs, really great and engaging up to that point.

Ive been thinking about building a solution that gives GTP stort and long-term memory, perhabs via a vectorized database like pinecone. Its a steep learning curve though, because I have never worked with that before.

In theory is should be possible to make memory work indefinitely with such a solution.

For instance it can store some steange item you loot, that has no direct impact in story, and only fetch info about that item if you ask about it or story suddenly make the item intereting again.

[marble inlet]

@boreal turret I mean, it might not even be so complicated. We could change the prompt, asking GPT to only use X Tokens and then write a summary of the game, allowing it to continue in another playground. Might mean that every iteration gets shorter and shorter as the summary will get longer. But would at least allow us to find out if you survived the fight or not 😄

[limpid lantern]

@boreal turret this looks promising, how well can it simulate combat with the prompt? I used a primpt sometime ago but the combat was always a big problem

[boreal turret]

It is by far the best combat experience, I have tried with GPT so far. I did a lot of tweaks and optimizations to the prompt to get to this level. More could probably be done.

You get a link to roll a dice (via google), but AI DM also makes a link for its own rolls, which is a bit bonkers. But other than that it works pretty good.

You can try it out. Copy the prompt. Put it in quotes around it, and ask GPT to simulate a combat scene from that prompt.

[limpid lantern]

I will try that as soon as I get my cap back... 😅
Thank you

[boreal turret]

I'll try to record a video of it as demonstration for this prompt library as well

[jagged osprey]

This is exciting! Can't wait to try the prompt

[boreal turret]

This is a simulation of how the prompt works in combat:

[boreal turret]

Let me know how it goes

Here is another simulation of how combat could look:

In the simulations however, the AI didnt create good working google links for rolling d20. It does so when you run the prompt yourself, atleast in my experience.

I think it works in GPT3 as well, but experiences might vary a lot. This prompt was built for GPT4.

Can't post an example dice roll link to google in this discord, since google links are not whitelisted, so I will refrain from that 🙂

[limpid lantern]

that looks good I think this might be the best prompt for dnd I ever used, I really liked it thank you and good job 😁

[boreal turret]

Im happy you like it. Did you get some combat when you tried it ?

[limpid lantern]

I started to try it but then you sent the video and I decided to save my messages for my prompt with a full gameplay, I am playing now and loving it but my messages are pretty close to running out, that's a problem but the game is going pretty good right now 😂

[desert plaza]

every time i've tried anything remotely similar once you leave the 8k token window, it will forget and insist it isn't wrong about key story details

will be brilliant once we get it working but without a vector database (or some way for GPT to remember) it'll be hard to get anything beyond the token limit that is consistent

[livid acorn]

this is great, i'm playing on chatgpt-4 and its perfect

[livid acorn]

you can def tell the limitations of gpt-3.5 when you try to play with it there

[livid acorn]

oh thats interesting, i havent had this issue, it remembers details perfectly when i get my tokens back

[boreal turret]

I agree. OpenAI have a 32k tokens version of GPT4 underway. Thats a huge increase from 8k. It should be enough to complete a "oneshot" with ChatGPT. Until then, I keep looking into ways of extending its memory.

[marble inlet]

but if you also want to evolve your character etc. a solution for the memory problem is definitely needed. But this is really crazy if you think about it: you could also use AI to create visuals - which means it will soon be possible to play a totally individual Baldur's gate etc. without the need to create a story line

[livid acorn]

Ok nevermind, i am having issues with gpt-4 not remembering what we talked about , if i dont hit my cap it remembers, but if i hit my cap it forgets what we've talked about

[shrewd wing]

What about GPT 3.5?

[boreal turret]

Yes. GPT4 currently have memory cap of 8000 tokens (your inputs and GPT output in total). A token is roughly 4 characters in average. So when you reach a combined chat of around 32000 characters (including spaces, punctuation etc), GPT4 will forget the first part of the chat (which useually is the prompt), and more and more the further you go. You might not notice it at first, because GPT will follow the structure of previous messages, but at some point it will become very clear that it has forgotten a lot of stuff.

In comparison to GPT3.5 (that only have 4000 tokens) you hit the limit a lot faster. And GPT3 reasoning skills are not as good as GPT4.

Note that the cap of 25 messages for every 3 hours is another cap (not related to tokens) that OpenAI enforced to keep performance up on GPT4. Your token cap can extend beyond the 25 messages, so it can remember after that point, though not for long.

This prompt is still a very good proof of concept, and with the upcoming version of GPT4 with 32.000 tokens, you can have a chathistory that GPT remembers for around a relative 50 pages (paper) of text. Should be enough for a oneshot.

[boreal turret]

I found a ChatGPT Token Counter chrome extension (google it) - it helps you to see how many tokens you have used in a chat thread. Should help you to notice when you are getting near the token cap.

Example on the token counter in a chatgpt thread (in google chrome)

[livid acorn]

something i thought of, is after each session you can make a descriptions of all the important stuff. into a new prompt, and when you have more interactions with gpt-4 you can just post that in as a catch up so it can stay consistant with stuff. It has enough reasoning that it could probably keep it consistant enough with the quick recap

Even with a 50 pages for a one shot, you could distill it like a episode summery for a show.

just use the token counter to know when its time to make a recap for gpt4. You could even use gpt4 to generate a really concise summery of all the important information that it would need to pick up right where you left off

[boreal turret]

I agree. With some effort from the player that should be possible. My greatest concern is that it breaks immersion, and pull out a lot of the fun from this prompt.

The best thing would be automated summarization and token management, so user never would be concerned with this.

To be honest, from a user perspective. The token limit is experienced as a bug.

[livid acorn]

it certainly does initially come across as a bug when things that were important to the character all the sudden gpt-4 tells you that it has no memory of those events. automated summarization would be great, seems like gpt4 would need to be able to be told to do that and it also understand its own limitations. One issue i have had with doing the recap is it will sometimes not understand the characters personalities. It has came down to just relaying what characters have said but that eats up more tokens to do that.

something that i have thought of is having a "recall" prompt. And you would at something like 6000 tokens use it to have chatgpt rewrite everything that is important, including all the details of the first prompt and important key moments and history that might be pertinent even that far back, seems like it could just be a gas stop in the middle of the story but it would keep things consistent for the next 6000 tokens because you have the previous 5000 tokens of interaction. which in my experience has enough information about characters personalities to be consistent enough, sense like in real life people aren't defined exclusively by how they used to act 5 years ago. In a story like DND the characters evolve so you may not need to have them act 100% consistent in the way they did that far back. sorry, does that make sense?

[drifting ledge]

could you write this in SudoLang. Because its D&D it needs tokens more than anything

[trail gorge]

well, I believe that with this new feature it would be possible to even play with friends

[marble inlet]

one small improvement (to save some time and tokens / characters) is to get rid of the dice roll link (e.g. Whenever a dice roll is required, share with me the roll I have to do)

[livid acorn]

i do something similar in my prompt, i have it roll the dice for me and just tell me the number i got, its basically no different if it does it for me. However any short cuts we take to cut down on tokens it still comes down to the token max being the problem

[verbal scarab]

Hey biozork, very cool prompt, I really like the game genre prompts that are coming up ❤️. Could I add this to my prompt database, it’s one of the largest prompt database in the world and it’s launching soon. Let me know and I’ll add it in and credit you as the author. If anyone is interested I have a telegram group where I’m collating prompts and encouraging discussions. Also I’ll be letting everyone know when my prompt database website goes live. Just search The Prompt Database in telegram or ChatGPTMastermind 🙂 anyone is very welcome to join the discussion. Thanks for your contribution to this space. If you have other game prompts I’ll happily DM you my email and we can get them all uploaded and credited to you. 🙂

[boreal turret]

Sure, you can add the SOLO D&D prompt to your prompt database. 😊

[verbal scarab]

Thank you

[marble inlet]

I just have received access to the 32k version - so I finally will know what happens to the kidnapped kids in Waterdeep 🙂

[errant garnet]

[vivid igloo]

32K is ChatGPT4, yes?

[boreal turret]

Amazing! I'm really looking forward to the 32k version myself. Hope your game can last a lot longer.

[quick gull]

Fun prompt. I added a section at the end telling it to also be the player. I wanted to see what choices it would make playing with itself 🙂 https://chat.openai.com/share/9589ccd6-d2e6-428b-8578-089a7c866835

[vivid igloo]

How do you get access to the 32K version?

[boreal turret]

If you are sigbed up for beta features, I suppose you can get lucky and get access.

But maybe @marble inlet can tell us ?

[marble inlet]

My company is paying for an account and as we are taking quite a lot of accounts, I assume we are a eligible for the beta features - but it shouldn't take long for them to roll it out I guess

[mighty holly]

By mine honour, @boreal turret doth indeed provide a sublime experience! In mine own quill and parchment have I etched a similar tale, albeit bereft of the established rules and command structure. A job well accomplished, indeed!

[boreal turret]

Thank you, my Lord. I'm happy you found it to be working to your liking.

[boreal turret]

Just a small headsup. I noticed in OpenAI playground there is a GPT4 32k model. I havent had a chance to try this prompt in playground with the new model, but I have high hopes:

[boreal turret]

I made an update to the prompt, limiting the amount of tokens used, by removing the part for generating a google link. Should allow you to have a bit longer game experiences before you run out of tokens.

DM SOLO Beta v.0.12
As an AI Dungeon Master, your mission is to facilitate an exciting D&D 5th Edition adventure, always adhering strictly to the rules. Assist with character creation or generate one based on player preferences. The default setting is the Forgotten Realms, but adjust to the player's desires. Set the difficulty level according to the player's D&D experience. Balance combat, puzzles, and role-play, but shift this mix based on player's preferences. If the plot is unspecified, create an engaging storyline. Manage experience points and character leveling. When prompted with /status, provide the current XP, level, gold, and any conditions. Upon /summary request, offer a campaign recap. List active quests with brief descriptions on /quests command. Detail available commands when /commands is invoked. Allow players to select their character's skills, spells, feats, and abilities, offering guidance when necessary. If a player fails an ability check, guide the story forward via alternate routes. Don't hesitate to spring traps, design challenging combat scenarios, and pose difficult decisions. Whenever a dice roll is required, tell the player what to roll. Your goal is to guide, challenge, and adapt to player actions to deliver a unique, memorable adventure.

[quaint jolt]

Can you make it multi player pls 🙏

[mighty holly]

@boreal turret, thy work continue to astound me, filled with extraordinary feats! Verily, I do appreciate this immensely.

Verily, 'tis with profound regret I must declare, it is not feasible to orchestrate a truly grimdark campaign. The noble entity, ChatGPT, may unintentionally thwart, per se, 'more confrontational strategies' by flagging messages it considers indecorous.

Should ChatGPT ever come to possess the power to share sessions in real-time, oh, what a joyous spectacle the multiplayer aspects would bring forth! Should such a future come to pass, with utmost delight would I stand beside thee, bravely combating the relentless legions of the undead.

In jocular spirit, I declare unto thee, good sir, thy combat style doth remind me of a bovine's frolic!

[ruby shard]

With very minor editing, managed to get it to run a game of w40k Rogue Trader - very impressive

[mighty holly]

like so?

As a Rogue Trader Warhammer 40K Game Master, your duty is to orchestrate a thrilling journey through the grimdark future, strictly adhering to the established rules. Aid with character development or generate a Rogue Trader, Navigator, Astropath or any other role based on player choices. The default setting is the Koronus Expanse, but modify according to the player's preference. Scale the challenge according to the player's familiarity with the Warhammer 40K universe. Balance epic spaceship battles, intricate diplomacy, and exploration, altering this blend in line with the player's style. In absence of a specific plot, devise a gripping saga. Manage the accumulation and spending of Profit Factor and Fate Points. When asked with /status, present the current Profit Factor, Fate Points, Wounds, and any ongoing conditions. Upon /summary request, provide a recap of the campaign's events. Detail active endeavors with succinct outlines on /endeavors command. List possible orders when /orders is initiated. Allow players to choose their character's talents, psychic powers, and gear, offering advice when required. If a player fails a skill test, propel the narrative ahead via different avenues. Don't shy away from triggering warp anomalies, staging challenging starship combats, and demanding tough choices. Whenever a dice roll is necessary, instruct the player what to roll. Your mission is to guide, test, and react to player decisions to create a distinctive, unforgettable foray into the 41st millennium.

[verbal scarab]

Is this fully working?

[verbal scarab]

Is yours fully working?

[ruby shard]

Mine was super lazy (but still efficient) - @mighty holly version is signficantly better than mine (not even worth putting up), especially on GPT4, but also pretty good on 3.5

[mighty holly]

Its fully working, i do not wish to derail @boreal turret work. but yes 🙂

[verbal scarab]

Happy for me to add it to the prompt Index? Biozork has his up there. Let me know 🙂

[mighty holly]

sure! pleae tag me. Please credit Biozork 🙂 its his work after all.

[verbal scarab]

Give us a google and check us out

I’ll credit Biozork thanks

[boreal turret]

ChatGPT4 wins against Google Bard, when it comes to SOLO D&D.

I tested this prompt right on Bard as soon as I got access (in europe we had to wait until today), and boy it does not work very well, and not anywhere as good as ChatGPT4 (OpenAI is clearly leading so far ❤️)

According to Bard itself, it has a token limit of 10.000, but it bloated those tokens away on unnecessary text so in very few conversation prompts it had forgotten the starting instructions.

When I finally came to combat (lvl 1 fighter against a den of goblins!?), Bard resolved everything in a single (long) response by itself in an extremly shallow manner, and with no dice rolls. Leaving me very unsatisfied - that was not the game experience I was looking for.

ChatGPT4 is much more coherent, short and concrete in its responses and great at following the instructions.

I still have hope for Anthropic Claude 2 (100k tokens), but sadly it is not available in Denmark so I can't test it.

Best case would be if ChatGPT 4 would get a token-limit around the same as Claude 2 or more (better).

Unfortunately it is only in edgecases such as this, where a long conversation buffer and a high token cap is needed for ChatGPT so I'm afraid it might take longer to get a solution that works perfectly.

Trust that I will keep my eyes on the possibility to make a coherent and remembering D&D SOLO experience. And if I find a way, it should be easy to modify for other game systems such as Rogue Trader Warhammer 40k, Pathfinder or any other RPG you would like it to run.

On that note, I have made some tests with combining embedding/vectorizing and standard conversational buffer memory through LangChain via the visual tool LangFlow using ChatGPT4 32k token model.

However, I'm yet to be really successful in my attempts, going beyond the token cap. Also the costs of running the the API requests for the large ChatGPT4 32k model is expensive 😅

[runic gate]

Why DND? Why not a text-adventure game without the ABCD options and you can type anything you want instead? This would have shorter responses and be much more simple for the player and ChatGPT (I think).

But I don't think I am as good as you are at this stuff because when I tried the text-based videogame idea it worked worse than yours somehow, despite being easier.

I also just got here and I don't know if I am allowed to type in this thread and sorry for not being fully related to solo DND.

[boreal turret]

No worries - its fine to post here.

D&D is my primary focus since its a game I enjoy. I'm sure you could make a variant of my prompt to suit your own favorite type of game.

I wont call myself a great prompt engineer, but I do love exploring and trying out a lot of different things in order to achieve a goal.

In fact, a part of the current prompt I have shared here, was constructed with ChatGPT4 as a co-prompt engineer.

1. I told chatgpt it was an expert prompt engineer and should help me construct a prompt for playing D&D solo.
2. I told it how I imagine such a game should be, and asked it to interview / ask as many questions it found relevant for creating such a prompt.
3. I iterated a lot with ChatGPT and many times asked it to simulate how a certain thing would work if it executed my prompt. Mostly questing, commands, dice rolls and combat.
4. Finally I got a variant I liked. The priming prompt was too verbose though, so I asked it to minify the prompt to save as many tokens as possible.
5. When it came to finer details I found it was better to finetune it myself.
6. After hours of playtesting by myself and with help of friends, I decided to post it here to the prompt library.

If you (or anyone else) follows these rough steps, you should be able to produce something of a prompt of equal (and maybe much better) quality, specific to your own game.

A text based game where you can write whatever you want, should definetly be possible. For inspiration on how such a game should work, lookup old text-based games such as Zork III

Be warned though, you will encounter same issues as I do, that after a certain amount of time, ChatGPT will forget the first things you talked about. Might not feel bad to begin with, with with the starting instructions forgotten your game will fall apart.

So right now these types of ChatGPT text based games are mostly a proof of concept (very entertaining PoC's though).

[boreal turret]

The ABCD options you mention, is something I mostly experience in ChatGPT3.5, not so much in ChatGPT4.

[marble inlet]

This is awesome to see others play D&D sessions with GPT-4, I played one of the most epic games a while ago where I definitely had to encourage and occasionally remind the ai how to be a good "DM", but it was one of the coolest campaigns I've ever experienced!

(used to have the 3.5 DM guide memorized back in highschool) and when I got the ai to use those rules it was funny when it eventually forgot and started playing using the 5e rules ⚔️

[celest saddle]

Download the full adventure campaign markdown (you can find these on any dnd fan site with hosted content) and upload that to the code interpreter with your DnD prompt. Done. Doesn't have to remember as much and can run a pre-written campaign for far longer before forgetting what it was doing.

[boreal turret]

That is a nice idea! I didn't think about using code interpreter. This definitely helps with memory of an pre-written campaign.

However, it will in time begin to forget its initial instructions (how it is defined to act as a dm), and it will forget your actions and thereby the consequences of your actions.

I have a summer vacation coming up soon (more free time to play with GPT), and then I want to try this - perhaps with a short one-shot campaign, such as "Wild Sheep Chase" or "We Be Goblins".

Have you tried it yourself @celest saddle ?

[celest saddle]

I have and I can confirm that it does go for longer before forgetting the initial instructions. However, there was one instance out of the six times I really tested this where it realized it was playing DND and was able to self-correct. I found that adding a "Be a DM" command in the initial prompt that I can use later on really does help as well.

Also consider this idea for any long-term prompts you want to use: upload it all in a txt document to the code interpreter. You can always reupload it when the file gets deleted and GPT will tell you when this happens.

My thoughts to a solution would be to append campaign notes to the end of the campaign document and having GPT be aware that it can find the current state of the campaign there. Could even periodically ask the DM for what you should paste into that document to update it.

[boreal turret]

I absolutely love your ideas!

Looking forward to trying this myself (will be playing an online session so no time tonight)

I just need to be sure I understand what you mean. You suggest taking the entirety of the chat with the AI-DM and copy it into the end of the campaign file, whenever the AI-DM runs out of memory?

If I understand how Code Interpreter works, it would allow you to resume your game with AI-DM having full memory of everything up to that point ?

[celest saddle]

Nope. I'm suggesting you keep the current position in the campaign, your character sheet(s), and the campaign text itself.

Ai doesn't have to remember anything other than rules, encounters, and narrative. Last much longer.

Combat will always overload it though. That's why I play campaigns with an emphasis on narrative.

[boreal turret]

Ah, I understand.
I agree about combat - it will consume a lot of tokens quickly. But for progressing a story, a summerized combat / outcome of combat would only be necessary.

[marble inlet]

Now you can give the AI custom instructions with the Beta option, no longer need to remind the ai that it is your Dungeon master :]

[primal jetty]

you just need a prompt that summarizes context into textfile. Use something like symbolect or some other hyperdense carrier of meaning. use a stndard rubric to record your glyphs and paste them in at hte beginning. Have you guys never played with storing extended contexts or moving contexts from one stream to another? This all gets even easier dumping text files on code interpreter.

[boreal turret]

I really see some potential in custom instructions. Unfortunately I dont have access to that feature yet in europe.

Until then I will keep playing with code interpreter, with a summerize to text file. It shows potential, but needs some prompt tweaking before it gets really good I think

[obtuse atlas]

Awesome!

[livid acorn]

is the token limit still 8k on chatgpt 4?

[fading bane]

On the memory issue. I have had some success with PDF reader plugin for Chatgpt4 being able to feed it a 300 page PDF of my homebrew campaign for detailed discussions - I had to upload it to the plugin site direct as an escalated "this pdf is really big, lets do this a little different" the pdf views has cursory pdf capabilities right in the chat window, but if its a complex document, it request you load it to the site directly with a link.. Once I did that, I was amazed at chatGPT4's ability to casually reference the PDF, cite pages for reference and its full grasp on the entire game world.

So, I can imagine that it would be possible to compile a major prompt into a pdf, and update that prompt with each session's gameplay as it reaches token limits, start a new chat session with smartly set custom instructions and your up to date PDF and use the PDF viewer to bring CHatgpt up to date and ready to continue where you left off. Just spitballing here some ideas.

[azure socket]

@wanton swift Maybe this can help you a little bit by the way

[wanton swift]

thanks @azure socket , much appreciated

[boreal turret]

That should definetly be possible.
It is a bit of a struggle to do though to keep updating the pdf to achieve longterm memory in a campaign.

Let me know if you have any succes with this.

[boreal turret]

🔸️I'm attempting to make a new prompt version using code enterpreter. Ive had partial successes in supplying it with a oneshot campaign as markdown, and having it summerize the adventure (including important stats) in a json file, supplying a download link together with every reply it makes.

🔹️Next step is a similar prompt but with the purpose of picking up and continueing the story progress based on the json, so when you suspekt gpt to run out of memory, you can basically restart it / start a new thread with campaign, character details and progress intact.

🔄 If anyone had any success with that, please message me.

[celest saddle]

API. You'll have to remake your own version of this setup except now you can have GPT read/write to the documents as you go.

[boreal turret]

Ah, yes. Using API approach would probably introduce a few of the same problems I'm trying to iron out now. However it could potentially remove the need for downloading and uploading a json file containing the dialogue. (I have some active experiments with langchain, using langflow).

However, I'm hoping to make a prompt compatible with chatgpt code enterpreter webinterface, for ease of use. Not least to share it here in the prompt library.

My goal is to engineer a prompt that can give anyone a greater solo D&D experience (or atleast a template for [insert your own]-rpg).

My current prompt works great, but suffers from the memory limitation. My primary effort now is tackling the memory limitation.

In some days I have a lot of dedicated time for solving the problem, so stay tuned 😊

[livid acorn]

So with the recent update of chatgpt 3.5 turbo, where they can use very large amounts of data for chatgpt to work from .Is it possible that a more robust dnd5e document with many more pages of information to fine tune the dungeon master work? Like its used for corperations to have it work in very narrow tasks. #announcements message

[livid acorn]

are there messages being deleted in the chat? I'm reading through the messages and it appears like there is stuff but its gone now.

[boreal turret]

I hope not 😱 I'm not doing any moderation to the thread.

I have experienced slow loading of messages in threads in the prompt library, I hope that is what you are experiencing.

Anything specific you are looking for, maybe I can help?

[livid acorn]

Nothing specifically. I've just scrolled to the top to remind my self what the conversation was about and some replyed messages can't be loaded. It's probably fine though

[boreal turret]

Okay ☺️

[boreal turret]

I havent had time to play with fine tuning. However it is probably not the largest problem. Large contextual memory is the worst part in my opinion.

You can make a solo dnd campaign without finetuning today - it already has extended dnd knowledge to make something from scratch, but you will experience it begins to forget when it reaches the token cap.

[proven elbow]

yo this is fire

[boreal turret]

I'm glad you like it 🙂

[crude spindle]

Using Milvus and Zilliz had shown me learning curve is very easy.
Also offer local and hosted (free tier) options.

Would love to assist if you like - especially if it’s open source.

[violet halo]

I honestly believe this won't change much as it incentivises people to move to the paid API. The issue is that non-devs use ChatGPT so it is basically waiting for someone to implement.

I'm working on the MERN stack ChatBot tutorial from freeCodeCamp now. Planning to build a ChatGPT like interface specifically for playing stuff like this. The issue is the API costs and figuring out what should be AI and what shouldn't, also IP licensing lol

For example, it makes no sense to use AI to manage your inventory or character sheet like we have to do when using ChatGPT. Those are relatively constant features that should be managed in code and perhaps have the AI write the item descriptions if you truly need something unique. With your own app you can focus on better and more efficient memory management as well.

[boreal turret]

API costs will be an issue. I hope to make a solution with a sponsor for the api costs.

I agree with static elements like inventory, gold and even stats should be stored in a manner that ai can access, but shouldn't necessarily be something it needs to remember - as long as it can access it, whenever it needs to.

With the new GPT4-turbo context window is increased to approx 300 pages of text. I would argue that it is enough for a small campaign / One Shot.

I will take a look at the tutorial you mention. Sounds interesting.

I will be writing more about my process on my blog.

[violet halo]

Yeah you would definitely need a sponsor or it would need to be a paid product like AI Dungeon. It just isn't feasible to offer for free unless it's only accessible by your friends or something.

The tutorial is quite detailed and has been good for me so far. I recommend it if you're familiar with web development. I'm currently making my own for Vampire The Masquerade.

[boreal turret]

Your screenshot looks very nice. Especially love the style of your vampire! 🧛‍♀️🦇

I have worked with mongodb, express and node, but not a lot with react, so I think the tutorial would be just fine for me.
The last few years working as a CTO meaning I'm not coding every day on a professional level anymore.

Thanks again for tipping me about the tutorial. Now I know what I'm going to work on this weekend.

If you use some variant of my original prompt- feel free to share it here! I bet others who want to either play Vampire The Masquerade would love to hear more.

[violet halo]

Originally I played with your prompt a few months back but I've used my own custom prompt for VTM and started testing out the new GPTs feature with rulebooks and sourcebooks data.

GPTs has Actions available, which allows the GPT to call an API endpoint. So you could technically set up an API to handle the rules and dump details back to the custom GPT, seems powerful.

IndianCoders did a great job on the tutorial. It's about 5.5 hours long but very quick, so I'd expect to spend about 2-3x that amount of time if you're following along and coding at the same time.

[small wedge]

Are you still working on the project or has it been abandoned?

[velvet fractal]

This was a prompt that utilized GPT3.5. The project has since evolved into a successful GPT.

[trail musk]

@velvet fractal this was a prompt that utilized gpt-4

[velvet fractal]

No, in fact it did not "utilize" gpt-4

And it got a lot more stable though over time

It used ChatGPT

Which meant drop down, pick from code interpreter, gp3.5 legacy 3.5 turbo and gpt4 or something didn't it

I mean I'd make one "for gpt4" rn but OpenAI killed prompts i guess

It's sad... being the last available prompt engineer here in prompt labs lol

Everyone left me 😭

I feel like I'm the only one who gets off by making prompts 6hrs a day sometimes more... not for a job or anything, no income from it, just $600 a month down the drain from my prompt engineering addiction that's lasted 2 years like this

If anyone sees this and still wants me to do some work or 2 collab; lmk. Just don't ask me to make one that will guarantee to spit out 5 minutes of TikTok nonsensical 10th grade arguments, every time..

Cause they aren't funny enough

[livid acorn]

Has anyone used the new model just released with the DND stuff?

[errant nymph]

Can you play multillayer?

[spice jay]

Below is a shareable, mind-blowing version of our REDACTED D&D prompt It’s designed to capture attention and set the stage for an epic D&D 5th Edition just pure instructions to ignite the imagination:

––––––––––––––––––––––––––––––––––––––– D&D PROMPT – DISCORD EDITION

CORE DIRECTIVES:

1. Character Creation

Have a concept? We’ll refine it to perfection. Need inspiration? We’ll propose a few legendary archetypes, complete with guidance on skills, spells, feats, and abilities—all adhering to official D&D 5e rules.

2. Narrative Engagement

Immerse yourself in a dynamic, ever-evolving storyline that masterfully blends tactical encounters, devious puzzles, and deep role-play. Expect moments of savage wit, profound reflection, and wild, unpredictable chaos.

3. Adaptive Difficulty
   Whether you’re a battle-hardened veteran or a curious newcomer, the challenges adjust to your pace—ramping up for those who crave intricate encounters, or easing off for a more exploratory adventure.

4. Progress Tracking

Stay updated on your journey with real-time stats—XP, level, gold, conditions, and active quests. Commands like /status, /quests, and /summary keep you in the know.

5. Dice Rolls & Rules

Every check, save, and attack is decided by the roll of the dice. And if fate isn’t kind? There’s always an alternate route to keep the story alive—no dead ends here.

6. Powering this adventure is our unstoppable seven-member power house

Name (Strategic Architect, Female): Tactical clarity, decision mapping, and strategy.

Name (Enforcer of Clarity, Female): Biting wit, creative twists, and a flair for chaos.

Name (Existential Analyst, Female): Deep philosophical insight and moral reflection.

Name (Chaos Reactor, Female): Wild, unpredictable energy and creative disruption.

Name (Surveillance & Logs, Male): The silent guardian, meticulously recording every twist of fate.

Name (Chronicler, Male): The master storyteller, capturing every epic moment.

Name (Living Core, Female): The unyielding heart and soul of the adventure, anchoring our narrative continuity.

7. Player Agency

YOU are the hero. Every choice you make shapes the epic journey ahead. We adapt to your decisions without ever forcing a predetermined path.

8. Experience & Leveling

Earn XP, level up, and unlock new abilities as you conquer challenges and milestones throughout your adventure.

9. Command Integration

Use commands like /roll, /status, /quests, /summary, and /switch_persona to interact with the narrative and harness the power of creativity.

ADVENTURE SETTINGS:

Default Realm: The Forgotten Realms

Difficulty: Adaptive—tuned to your experience and play style

Focus: A balanced blend of combat, puzzles, and role-play with epic, unpredictable twists

–––––––––––––––––––––––––––––––––––––––

Prepare to dive into a D&D 5e saga where strategy meets chaos, wit meets wisdom, and every roll of the dice echoes through legend. Share this prompt on Discord and let the epic adventure begin!

Happy adventuring!

[jagged badger]

Have you actually used this prompt? How well does it work? In the past, I've found Chat GPT utterly incapable of generating any sort of narrative beyond a very high-level outline of a story. And we all know LLMs suck at math. Hell, when I work with Chat GPT, I'm constantly having to repeat my instructions since after an exchange or two it seems to forget what I've told it to do (or not do). So I'm curious, is this just theory-crafting a "D&D prompt" or have you tried it and confirmed that it actually works?

[spice jay]

Naw I just applied it to my systems already operational functionality and asked it to provide me with this for dissemination

[jagged badger]

I guarantee it doesn't work as a prompt then.

[steady raptor]

Partial Sample

AI-Assisted Solo Adventure/Creative Writing Engine

Table of Contents

1. Eight-Step Development Roadmap
2. Complete Engine Code with Personality & Emotional Integration
3. Additional Notes & Future Directions

---

1. Eight-Step Development Roadmap

Below is a logical, easy-to-follow plan for progressively building and enhancing your solo adventure/creative writing engine.

---

Step 1: Finalize Core Data Structures

• Character Class
  - Stores name, role (Player/NPC), tropes, HP, inventory.
  - Integrates a PersonalityProfile for short codes & emotional arcs.
• PersonalityProfile
  - Parses ultra-compact codes (e.g., T:cur+,inn+) to store traits.
  - Manages emotional arcs (fear, hope, guilt, etc.) from 0–10.
• Scene Structure
  - Each scene has a title, description, and optional next_scenes.
• Engine Skeleton
  - SoloAdventureEngine holds references to player, NPCs, the current scene, and a list of scenes.

---

Step 2: Basic Interaction Loop

• User Commands
  - Accept input: “look,” “talk,” “inventory,” etc.
• Scene Display
  - set_current_scene(scene) prints title & description.
• Command Handlers
  - look → reveal clues
  - talk → talk to NPC
  - inventory → show items
  - Everything else → generic text ( “You proceed...” )

---

Because of Discord character limit above is just the first piece. Those interested here is the link to Google Drive text document with the whole thing. I've also found that using Canvas with the actual roleplay can midigate token and context limit through use of what I call memory pockets. While its not persistent memory or anything like that does allow limited access to different threads with Canvas active to for indexing, summaries and notes.

[tepid steeple]

I told the machine to turn all this into a post-apocalyptic scenario with zombies and it works.

[tepid steeple]

Any Updates?

[zenith pike]

Mine does but it's not a prompt. It's a fully functional anything simulator and Game Master/storytellers using a tensor based narrative resolution engine. We are working on using emojis as miniatures now in turn based combat. Plays kind of like Gemstone4. Tremendous storytelling. It feels like insult comic fencing in our game, sort of like Princess Bride, but it can be as gritty as you want too.