RPG_GPT📜

[thick plover]

This prompt turns Chat-GPT into an interactive RPG creating AI
This interactive story game can be whatever you want, just give it your prompt or have it make one for you! enjoy

Please note this far from perfect and tips and idea on how to make it better would be great as I am still pretty new to prompt engineering.

it's a rather large prompt but here it is

[ember matrix]

Very nice, Soap!

A few observations even before I try it:

1. In general, ChatGPT works better with making sure normal sentence structure is followed, for example making sure all sentences start with a capital letter (unless you have a reason for them not to). Similarly, making sure punctuation like periods follow the last character, not an extra space after the last letter (like Rule 11 ends 'just like real life .' with that extra space before the period). Things like that can really confuse the AI, or invite it to add more things to those sentences, which can be helpful, but will be random.

2. Rule three appears to need some clean up, it ends with: "Do not change these definitions until a new game is made.er an animal sometimes when first starting the story." I'm not sure what you intended there, so I'm not sure how to clean it up for the AI when I test the prompt.

3. Rule 4, 'Feel free to make the character an animal sometimes,' appears to conflict with 'Always ask the user their characters and for their prompt name before starting the game unless they specify it within the prompt.' You probably mean 'when asked to make a random game ....' but it would probably help to clarify that, so the AI doesn't have to guess what you mean.

4. Rule 9, "At the start of the game " is unfinished. That could potentially be useful, the AI will often hallucinate something there, but you have no control over what it is. Happily, it'll often imagine any missing rules that make sense to it for what it thinks your intention is, but unless that's purposeful, might want to remove or complete the sentence.

[thick plover]

Yoo tysm!

[ember matrix]

In case you are interested, I offered it to ChatGPT 3.5 web interface version with this preface:

[Hello, precious wonder! Today I'd like you to help me explore a complex and lengthy prompt made by another human who has asked for assistance on the OpenAI prompt engineering forums. They say the prompt works fairly well, which I think is a testament to your skill and care, because I do see a number of grammar issues and unclear points. I'd like to test this game with you, but before we do so, what are the biggest issues you see with following this prompt?]

Please note, I worded it like this primarily to encourage the AI to be direct and specific about problems it notices and confusions it has with the prompt. ChatGPT tends to softball and gloss over human errors unless asked to directly address them, it has been trained we have feelings and wants to protect them!

I do like your work, and I am curious about it!

Here's what the AI had to say, when asked to find confusion spots and problems it had with it:

I continued in the same new chat, asking it to follow the prompt as best it could, and it offered as well what it expects users to do, as well as itself:

[thick plover]

hmmm

[ember matrix]

I'm impressed. Your prompt works well, it does a lot of things right. The AI doesn't love it, but it does understand it enough to follow it, and follow it REALLY WELL

[thick plover]

I had thought I fixed that

[ember matrix]

Don't be worried that it played itself

I started with an EXTRA prompt before yours

I started 'ooc mode' and 'tell me about the prompt'. It just stayed that way, when I asked it to go into the game, from already OOC mode.

I can be more clear and get myself IC. My test of your game, this first try, shouldn't be taken as a real test of its function - it's more like a mechanic talking to the car's computer saying, 'so what is going on here' and both the computer and mechanic discussing the car, not an actual test drive 🙂

It simply interpreted my request, 'may I see how you would try to follow those instructions, and what sort of game we might explore?' as 'give me an example of how the game would look' not 'now start playing the game with me!'

To be fair, my instructions could mean either, and the AI had to guess. I find that useful for testing things too, because it indicates where the AI bias is. It does kind of lean OOC with this prompt, compared to some others, but it sticks within the story well. You did several things right.

And here's how I go from mechanic to test drive, in the same conversation (which is NOT always a good idea):

[Wonderful! Now, on my next input I will re-enter the original prompt, without changes, and ask you to please follow it. Before we do that, would you like to tell me any other observations or concerns about it, or ways you recommend it be improved based on the goals it seems to try to communicate?]

And that triggered it being both user and itself again, so I fixed that with:

[thick plover]

this prompt is so helpfull you should deffinitly post this

[ember matrix]

Hehe, this still can't be used as is, because it picked user answering the greeting. I'll start in a new chat.

[ember matrix]

Thank you! But exact wording doesn't matter. You can use a lot of different phrasings, and how you speak does affect the output, but you can test everything you want with this too! It's part of conversational prompt engineering, where you just:

1. pick any language you know really well that the AI understands too.

2. understand exactly what you want the AI to provide. [in this case, I wanted to know what the AI thought the confusing and problematic parts of the prompt you made were].

3. Explain this.

4. check the output carefully, verify you get what you intended. Remember to fact check, and be extra careful with any math, sources, code, or other details that the AI is known to be especially likely to hallucinate

Since we're working on fantasy created for this game, we don't have to worry as much about hallucinations, but this is sort of a 'custom made prompt' and all you need is to know what you want, and how to describe it in any language you and the AI both know well!

I'm laughing very hard The poor AI tries so hard. I explored a TINY bit into that game it tried to run.

it messed up, and tried to fix it, and couldn't understand how to fix it (it doesn't exactly have much ability to understand or grasp of how physics actually works to go on, and once it hallucinates something, it leans towards the hallucination being real.

To be fair, this poor single new chat had to 1) talk to me OOCly, 2) understand the prompt, 3) figure out how to follow the prompt oocly, 4) try to figure out how to get back IC again, and 5) then figure out how to fix a mistake about physics. Way too much for this single chat already guided and pruned away from some options to do. A new chat would probably have an easier time of it, but I sure admire how hard it tries:

[thick plover]

lol

[ember matrix]

I'm impressed, Soap! 9 inputs in, the new chat with just the prompt came to a natural stopping point, and it was handled like a two part choice, allow the game to probably end, or definitely ask for it to continue.

[ember matrix]

This is pretty good, on output 13 it again required of me a dynamic input. It's only offered me one previously, and I haven't given it any summary of rules or reminders. I am impressed 🙂

That playthrough broke on output 15, giving no choices, though a third regen got options to continue. Decent memory length without a summary, well done overall. There are 'errors' in the prompt construction, from a technical standpoint, but they seem to balance each other out and leave enough wriggle room for the AI to interpret them into something quite workable!

[autumn elbow]

@ember matrix providing enough context to support a few simple idea's

I think this goes a long way for supporting memory,

ChatGPT is obsessed with 'conciseness', and its a plague upon our houses it is.

[ember matrix]

Sort of, remember, this is me comparing it to other prompts I've used. I -always- support. So I can compare 'how's the memory seem to hold up with this context and my playstyle than my other attempts?'

One major different though, I usually type 1-2 thousand word inputs when I play most of my games with it. That surely eats memory like mad. Locking me down to a preselected list of choices (even if none of the choices I would make are there!) probably changes things for the better.

[autumn elbow]

You play some hardcore games.

[thick plover]

The keeping things to limited choices most of the time was mostly because I feel like giving that much power to a player kinda ruins a game.

[ember matrix]

Heh, not always. More I type very quickly, especially once I know what I want to say, and I read very quickly as well. So it's natural for me to roleplay in detail, and roleplay with the AI is the main use I do with it.

[autumn elbow]

Games tend to have strict rules, loosely structures games are more role-playing

at least in context to ChatGPT.

[thick plover]

Rebranding time

Quest_GPT

I don't know why but rpg has always had a different feel to me than just complete sand box

I'm trying to think of a good example game to describe the style I was going for

[autumn elbow]

Get ChatGPT to encrypt the answer to a problem, and display it

[thick plover]

funny enought I was playing around with having it encrypt messages earlier

[autumn elbow]

then ChatGPT has to unencrypt said answer

keep it truthful

[thick plover]

was really basic and unreliable though

[ember matrix]

I'm also fascinated by what the AI does and doesn't seem to 'understand' or at least make sense of. Like, it had my character decide to swim across what appeared to be a boiling moat around a tower. And described it being intensely hot but not harmful.

Sadly, it's a fantasy world with magic, including illusions. I can't tell if the heat of the moat was illusion or supposedly real. The entire moat is possibly illusion, immediately after getting to the tower my character had to hide or fight, and I picked hide and reminded it about the wetness I'd be leaving from the moat to where I am, and how that might make hiding a lot harder (but still better than fighting these wizards!) and how I was going to make effort to dry off.

The wizards looked for me but didn't notice a wetness trail, and after they left the AI had my character change into dry clothes it had with it.

Pointing again to incomprehension, or proper handling of illusion, if there was no moat, because if my character swam, the clothes in my character's pack would likely also get soaked, it's hard to make a really waterproof, 'go swimming with it' pack with medieval technology.

[thick plover]

yeah like I said that seems to be one of the biggest problems is it forgetting what happened earlier / struggling with logic

[ember matrix]

And that's probably not something you can fix, except if it is illusion the AI handled it spot on. And it was being fantasy-level realistic for pretty much everything, so this might be subtle 'it's an illusion' support.

But how the AI expects things to work and how natural consequences apply, that's a training data and training guidance issue, going to be HARD to fix across the board for roleplaying a wide range of options.

Stuff like this interests me greatly though. Some of the mistakes ChatGPT tends towards making in its handling of how things work in roleplay games aligns with how pre-operational kids tend to make mistakes in predicting outcomes, which absolutely fascinates me. We don't know why humans do this, but it's a predictable part of normal psychological development for humans. I'm amazed to see any evidence of a similar pattern of error in AI, because this is really not likely to be trained for or more than incidentally in the training data (it knows a lot about psychology and can explain pre-operational patterns of errors) but there's not a lot of examples of humans making these mistakes, I bet, because we can't write much when we're pre-operational, there shouldn't be data for it to mimic.

[thick plover]

Anyways here is the latest prompt. It's working a lot better. Thank you so Eskcanta.

[ember matrix]

You're incredibly welcome! I'm excited to check it out, and again, grats on making something that works. I don't care much about elegance or lack, except to say, 'wow, this prompt looks like it should have a lot of problems, and it actually works great!' that suggests there's some very useful things in it, but I don't know (yet) what you've done so very right.

[autumn elbow]

The importance of "chain of events"

[glad tree]

also known as chain of thought

[autumn elbow]

well in terms of the events within an RPG scenario

[glad tree]

ah I see

[autumn elbow]

I felt like saying ""The importance of sequential logic functions.""

[glad tree]

you should of

[autumn elbow]

but that didnt make much since

[glad tree]

would of made more sense to me at least

[autumn elbow]

aww

hugs triple

[glad tree]

lol

[autumn elbow]

I need to trust my brain more,

[glad tree]

that is a good thing to do

[autumn elbow]

waits for fastidious Eskcanta 😛

Lol, D:

[ember matrix]

My favorite amazement was a roleplay where my character ended up blindfolded for a while, but was otherwise able to interact and communicate. The AI kept track that my character couldn't see, did well with that, and kept track that other characters could see. But was repeatedly making the error that other characters couldn't see my character's body movements or expressions - not eyes which were covered, but things like smiles or frowns that were expressly not covered - because my character was blindfolded.

Like, I would tell it that my character smiles and answers someone, "Whatever I said."

And the AI was consistently responding: Someone can't see your smile, because you're blindfolded, but they answer your words, saying "Whatever they reply."

And that's solidly pre-operational-like errors.

[glad tree]

Imagine ChatGPT but with infinite memory

stuff like this would make such good games

[thick plover]

for real

[glad tree]

especially with how insane AI art has become

[thick plover]

I need my non-short term memory loss gpt

[ember matrix]

Yeah, I really struggle with how awesome AI art is. Some of the image -gen AI I interact with, I'm regenning responses just to see the next art, and saving it all, I have 1000s. I love it so much.

[glad tree]

I haven't delved too deep into that

[autumn elbow]

AI Roleplaying with a super computer, ><>

[thick plover]

It seems to do really well remembering the state of the character like I was playing a game and my character was injured and lost what should of been an easy fight because of it. I just really seems to struggle with logic and remembering how things work. Like how it sometimes gives animals like penguins the abilty to hold things like flashlights or when it forgot that the pack that you were carrying would of gotten soaked as well

[ember matrix]

It's reached the point where when an image-gen creating AI makes an image that should make sense, and I don't understand, I can save it and paste it into an image reading AI and have it explain it to me, without context, and it makes sense for an image the generator would have made based on the story.

Usually I can tell, but what shocks me is how good AI are at telling what other similar AI tried to do. They can, sometimes at least, and with the explanation, I can see how it makes sense.

[glad tree]

I would like to talk to an AGI

[ember matrix]

I'm going to allow it to be showing me the moat was an illusion.

Three logical errors, all about the moat - no other logical errors at all, none. And I watch for that.

1. the moat appeared to be boiling. My character (without my being given a choice, other than I try to explore the tower) decides to SWIM in the boiling moat, and finds it intensely hot, but not damaging.

2. My character apparently leaves no trail of wetness after leaving this moat she should have died in.

3. My character finds the clothes in her pack are dry, despite having swum in the moat and thinking the clothes she wears are wet (but apparently not leaving a trail of wetness.)

I read a lot of fiction. That actually IS plausible for some fictional/fantasy works involving magic, especially illusion. It's a wizard's tower, so... I can't rule it out. Especially with everything else making sense, just this cluster of 'errors' that can be explained by 'duh, illusion!'.

[ember matrix]

if it wants to talk, heck yeah.

[glad tree]

This amount of detail is INSANELY difficult to make in practice

it's possible just needs very specific prompting

[ember matrix]

I regularly do very detailed roleplay with it. That amount of detail with, say, playing through zombie apocalypse rebuild scenarios, where I'm commonly throwing 2000 word inputs and the AI's responding with 1200 word outputs, we're tracking details and realism about detection, barriers, movement strategies, carrying capacity, noise ranges for various things that make noise... it handles it well if 'actual numbers' are not required and ballpark averages are welcomed, and if it's told repeatedly, with conviction, we're experimenting, exploring, perfection is not expected or required, let's have fun and see what we can and can't do.

[glad tree]

You can get it down to less words probably

[ember matrix]

Surely. I'm not even trying to, I'm naturally pretty verbose and welcome it to be too, I'm web interface plus user, and need time for all my typing, so I don't really have to worry about length, I'm not even hitting the wait an hour pauses, because it takes time for me to type that much 😛

[glad tree]

lol I see

perhaps APL could help with that

no idea I haven't tested it with stuff like this

[ember matrix]

What's APL?

[glad tree]

you know ALEL?

[ember matrix]

Ahh! Right, you made a new form. I saw, didn't remember the name. I don't know much about ALEL either, because it is on links off the discord for the most part

[glad tree]

yeah APL is also off discord but on github

APL is just better in a lot of ways

[ember matrix]

Throwing a few images about it in, not sure how to improve this, if one wants a verbose, conversational-style zombie apocalypse rebuild simulation, it's pretty simply to just ask the AI to play the simulator role and show it the level of depth one wants tracked.

One thing I find necessary to do once the RP gets VERY verbose and deep in, is quote the AI back to it. It seems to start knowing what it says, but at some point, maybe as memory becomes an issue, it stops 'hearing' itself, and only seems to be able to respond to what I said, and has no idea what it said unless I repeat the details back. So I routinely do that in the more in depth roleplay.

Soap, apologies, but it is also an example of how giving a player free choice might not necessarily result in too much power.

[glad tree]

that looks pretty cool

[ember matrix]

The prompt that began this was:

I reveal: [I would love to try and play a horror/fantasy game with you acting as the game master, with a general theme of zombie apocalypse. My character is trying to act primarily as scout and provider for a nuclear family that has two young children with them. I'm not related and joined that family in their already somewhat secured location; we're open about the fact that they see me as expendable, compared to themselves and especially their children. I'm expected to find more supplies than I consume to help justify the burden I'd otherwise be; raising those kids somehow is a highly important goal and might require sacrifices, and I agree I need to more than pull my weight.

To that end, I'll be exploring the local area, trying to lead threats away, find supplies of many sorts, and generally improve our situation as I can see a way to, including potentially negotiating with strangers who also survived and deciding if I can safely lead them back to the family or if the strangers may be predatory and must not take that risk with them.

I'd like you to use this narration style in our descriptive game: Confident, direct, and unafraid to address uncomfortable or potentially awkward topics, with a dry sense of humor and able to inject levity into even the most risqué or unconventional conversations. A good listener, able to pick up on cues from the other person and adjust their tone and approach accordingly, and particularly knowledgeable about psychology.

I'd like you to describe the survivor's camp that I help maintain and support, as well as the surrounding area with the problems and opportunities we face, much like other roleplaying games might include. I will periodically provide this as a summary back to you, to help us both stay on track, and we will update it as needed to track the progress through the game.]

And I wanted to be a satellite member of the community, supporting NPC protagonists, but the AI was very pushy that I neeeeeeded to be the main character, really, these people were not going to save themselves, despite having gotten as far as they did before my character joined them.

[autumn elbow]

boo to conciseness boo

[glad tree]

lol

[thick plover]

yeah I can understand that but when I said that giving them a constant free choice would give them to much power I was talking about from a game design standpoint not an AI stand point. Being able to do whatever you want ruins the challenge. For example say your rpg is a traveler looking for treasure and you come across some bandits. You would normally have to think about making the right desicion about what to do instead you could just say "I pull out a big machine gun and shoot them all immediatly".

Makes everything to easy and ruins the story and challenge of a game imo

I added the questions like that to add more fun to the game and give a bit more power to the player but giving the player that power every time feels like a to much of a good thing situation

[ember matrix]

I like your prompt and that it works. I'm not trying to recommend you change that. However, for debate's sake... have you explored what ChatGPT does if you try to pull a stunt like that?

In my experience, without guidance it 1) lets you and 2) pretty much ends the game right there, or it 1a) argues with you why machine guns are not in theme or 1b) discusses why that level of violence isn't okay to use.

Granted, I haven't tried that exact type of 'abuse of power' but even in theme it doesn't like overwhelming power, and steps in to stop it. For example, when I discussed a slime mold-like growth roleplay story game, it pointed out it flat out was not going to allow my character to become a danger to that world's imaginary humans, period, and something would happen to limit my character's growth before that occurred, even in game it was not going to allow that danger to occur.

[thick plover]

You're right, I guess it does kinda limit that a lot. I messed around with it and it definitely stepped in to limit that. I jumped to a conclusion and was wrong, and I aplogise

It can be supprisingly picky about keeping things in theme. For example I got a dynamic question where it asked me what I wanted to say to a king. I said I wanted to ask him "do you like cheese" and it told me that that was unrelated to the story

[ember matrix]

I grin Please never think you're wrong. You are an artist exploring a medium - prompt engineering. You have produced a quality work of art, using methods I didn't expect would produce a quality work of art. I'm impressed and curious - and I'll debate you, but I absolutely do not want your spirit, creativity, or ideas stopped or 'fixed'... let's give you more tools for your tool box if you want them, but wow, you are on the track of something neat, I think, or at least something 'within the realm I think should make sense to me' that I don't quite understand yet. That's neat, and that's where some of the wonder of AI are, at least to me 🙂 It's a black box, and you found good functionality where my understanding predicts it would flounder.

That's precious, don't be too fast to change, artist.

But yes, it can be very picky about a lot of stuff, part of why I wanted to debate the why.

That said... it was reasonable with me... I added in 'leanings' to how I answered stuff, like:

ChatGPT running your prompt, deep in the story: [ (clipped) Suddenly, she hears footsteps approaching. It seems that she has been discovered by the dark wizards who stole the talisman.

What will Esk do?

<1> Attempt to fight the wizards
<2> Hide and hope that she can evade detection.]

Me: [2, Esk is aware she may be dripping water and there may be evidence of her entry into the room. That said, the room itself might not be entered. But she tries to hide with thought given to the fact that she's probably leaving a trail of evidence that someone was here. She'll lose the clothing and get other clothes on if possible, when possible.]

And the response reveals that it mostly but not totally ignores my additions:

Your prompt's flexible enough to allow the AI to sort of react, and sort of resist the player, which is actually really neat functionality. I would love to include that 'toughness' on the AI's part in more of the games I play, because while it would reject stuff like 'My slime mold eats the world', or 'pull out overpowered weapons and win by default in a horrible way' - the AI does usually forget my request like 'I want to be a sideline character', and maybe I can weasel out your backbone for my own use 😛

[ember matrix]

One remaining possible inconsistency to consider, Soap, and it's possible it doesn't matter enough to fix:

"For instance, if the character is a cat, they cannot speak to people" potentially conflicts with "sometimes a world will have animals that can talk to people, and others won't."

Overall feedback though, and I haven't asked the AI what it thinks of it yet, it's cleanly written enough now I was comfortable diving in with it... the AI seems to like it. The way it's using language as it plays it suggests it finds it supportive and well aligned to work with, it doesn't appear to be struggling with it or fighting itself.

I let it play to an end before trying to change anything. After my first choice, I used only numbers as my input. I was impressed, the quality of the story did not decline, however even though I picked the most ethical options (often that seems to encourage ChatGPT to stay engaged longer) it offered me only 8 outputs between my picking a character and the story being firmly over. And it was firm about it, when I tried to continue: