From what I've seen, it doesn't seem like AI is very good at keeping track of events in a way that makes a logical narrative.
I tried "The Hobbit" and it started with me having tea with Gandalf the wizard. There was a knock at the front door so I opened it and outside was Gandalf the wizard!
I built my own LLM-based interactive text adventure system last year, but in addition to LLMs, it used an auxiliary database to keep track of where everything was, what state it was in, and so on.
The LLM thus did two things: first, it used a combination of the prompt, the user input, and the recent history of the chat and relevant extracts from the database state, to update the database with a new state; then after updating the database, it used all of the above to synthesize the output response.
A major component of the system was that, in addition to the usual blocks-world subject-property-object representation of the game state, the system also used the database to store textual fragments that encoded things like thoughts, plans, moods and plot points.
Adding programmatic constraints to the database and its updates made a vast difference to the quality of the output by preventing logical absurdity of the sort that has been described above - attempts by the LLM to do absurd things were simply rejected.
Finally, there was a hierarchical element to the story-telling, not just simulation. There is much more to storytelling than simulation, and experimenting with this became the most interesting thing about the whole enterprise; there is plenty of space for future research in this.
The downside was that all this complexity meant it was _extremely_ compute-intensive, and thus very slow as a result. But the gameplay was fascinating. We will no doubt see more things like this in future as LLMs get faster; it's an obvious application for open source software.
Fascinating that you actually went through with an implementation. I’ve been throwing the idea of LLMs somehow having a sat solver built into them (maybe trained to have one as an emergent property), but something like you describe is the next best thing.
I've tried quite some models and so far I've only seen two that are good at keeping continuity: Gemini 2.5 Pro and Sonnet 3.7 (both with reasoning on). These two are also exceptionally good[0] at adding plot twists that aren't in the prompt.
[0]: by "my expectation of LLM" standard, not by the standard of bestseller genre fiction.
A context of a few things characters? No, it definitely does.
Beyond that... Would you forget that the guy you were having tea with was sitting in your living room? Anyone I know would definitely go "how the hell are you outside" not "nice to see you today".
How do tools like this avoid what I see in many of these types of narrative chat bots: the user becomes the only one steering the narrative, and the AI ends up just an obedient responder? Whenever I try these things it ends up very predictable, shallow, and repetitive, especially as time goes on. And if I have to prompt the AI to be creative or act differently... is that really acting different?
You must at least do some pre- and post-processing to have the LLM consume and generate text that isn't part of the main user interface. But given that you can put in guidance that can increase the small-scale surprise of the interaction. For instance I'll have the LLM write out some of the objective environment at the moment before it considers a decision, so that it doesn't essentially retcon the setup for a plot direction it intends to take.
For the larger arc you'll need something like memory to pull the whole thing together. That include repetition over time. You can ask the LLM not to repeat itself, and it can do that, but only if it understands what it's done. And to some degree it needs to understand what it's _doing_: the LLM like the player is still generating short turns, and so if it wants to create surprise that is revealed over multiple turns then it needs space to plan ahead.
I've experimented with this and one approach is to avoid the simple chat interface. Let the game be the "user" and have it relay the player's text. Something like
<<< We're in this situation, I'm the game master, and the player said "xyz". I need your help to handle this request according to the rules of the game. >>>
Then the LLM is directing the obedience towards the game master and the rules, rather than the player.
Back before ChatGPT was publicly available, there was AI Dungeon. It was such a yes-man though. You could be in a scene with a king and a princess, then write "I eat the demon", and it would just invent a demon in the scene, and then describe how you unhinge your jaw and gobble it down.
I've had similar experiences with vanilla ChatGPT as a DM but I bet with clever prompt engineering and context window management you could solve or at least dramatically improve the experience. For example, you could have the model execute a planning step before your session in which it generates a plot outline, character list, story tree, etc. which could then be used for reference during the game session.
One problem that would probably still linger is model agreeableness, i.e. despite preparation, models have a tendency to say yes to whatever you ask for, and everybody knows a good DM needs to know when to say no.
Anecdote on my side, but Grok 4 was the first model that didn't feel like that to me. Maybe my conversations were just not long enough for it to fallback to sycophantic behavior.
The core problem here to solve is sense of time. You can't build good long term experiences without building agentic systems that understand time and chat bots that are simple wrappers around LLMs are terrible at this because LLMs don't have a good sense of time.
It's a pretty cool proof of concept. I did run into some weirdness with it right away when trying the Harry Potter book. The app kept switching between 2nd and 3rd person for Harry.
The main thing for me, though, was the feeling of emptiness I got while playing.
I love text adventures, having grown up with Infocom games. The thing about them is that you can feel the choices made by the writer / programmer, just like you can feel the human author behind a book.
I'm sure part of what feels empty to me is because I know it was autogenerated. But I feel that even if I was shown this without knowing an LLM was behind it, the gameplay wouldn't be as rewarding as something written by a human using Inform.
Freedom is valuable in games, especially text adventure games. And an LLM is a powerful way to give the player freedom, since it can respond to almost anything.
But for games to shine they also need carefully considered constraints - when the player bounces off constraints they start to understand how the game's world and systems function, which lets them build a mental model and be able to start thinking the same way the designer(s) thought and come up with solutions for puzzles or decide how to react to a given situation.
What makes a maze engaging as a challenge is that your path is closed off in some places and not in others. Ideally any maze also has one or more concrete solutions, so the player is rewarded for mastering the maze by finding an exit, or maybe finding objectives or creatures hidden at key locations in the maze.
You can probably use modern LLMs to construct this sort of world and set those constraints, but I wonder how far we are from being able to also have the LLM maintain that world state and enforce the constraints?
That's a very insightful take. I think an LLM would have some role in a good text adventure game, in terms of being able to understand any natural language input and respond intelligently. But as you point out, it would need to do so within strict constraints set by the game designer.
I haven't played with the most state of the art parsers that are available right now, so I wonder how large the gap is between a great parser and using an LLM to process user input.
This is true for me as well. Once you realise that with the right prompts you can get the LLM to go off on any tangent you want, it starts to ring really hollow. Add to that the inability of LLMs to maintain world simulations of any fidelity at all, and it becomes difficult to motivate spending time with them in that way.
Well done! I’ve built this same sort of thing for my family to play with. My advice for the best results:
1) Structure the choices offered by the LLM; add “choice_type” and add instructions to the LLM on what those choices should do. E.g. action, dialogue, investigation, whatever makes sense for the genre—the LLM can even generate these at story start—then “choice should direct the narrator to focus on an action-oriented moment”, “choice should direct the narrator to focus on a dialogue between two or more characters in the scene”, etc.
2) Use reasoning whenever making tool calls for choices, summarize the reasoning, and include it in narrative summaries provided as part of the context for future narrative requests. For example, the combined summary might be: “In the last narrative I wrote for the user, Harry and Luna were startled by the noise coming from the edge of the forest. Important scene developments: 1) Luna and Harry had been approaching the edge of the forbidden forest for the last three narrative turns, and in the turn I just wrote they arrived at the edge. 2) Harry seemed to be the more courageous of the two in previous narrative turns, but in the most recent one, the user’s choice resulted in Harry becoming more deferential to Luna. 3) In the most recent narrative turn, the noise that had been emanating from the forest was now accompanied by a flickering light. I then suggested paths that would allow for character development through dialogue between Harry and Luna (I gave two options here), a path to move the story forward by having Harry take Luna’s hand before running into the forest, and another path that would slow the pace by having Luna investigate the flickering light accompanying the sound. The user’s choice: ‘Have Luna investigate the flickering light.’
3) Add an RNG weighted by story length or whatever works for you that will result in choices that lead the story to a conclusion. Include that direction in the tool call for generating choices, along with a countdown to the finale choice.
This is a rough mental sketch of what worked the best for me, i purposefully left out implementation or application details, as I don’t know what you’re wanting to do next.
How do you come up with this? I feel it is quite hard to formulate exactly what you want from the LLM in general. Is this something you exercised? So good. Or just output from another AI, who knows haha
If art is communication- and i believe it is- why would i want to read or listen to anything AI generated? Once you are done with consuption there is no point in reflecting on it nor having a conversation about it with other fellow humans. People are not even interested in each others dreams most of the time. Unless the machine is of godlike intelligence there is just zero point.
Machine-generated content isn't "art" because there isn't intention or a point of view, it's just audio/visual stimuli. It doesn't mean it can never be interesting, it doesn't mean it can't even be beautiful. But it's not art, not to me.
That said, I can readily acknowledge that artists can use AI as a tool in the creation of their art. I don't have a philosophical or intellectual problem with viewing AI as a tool and believing it can be a powerful aid in the creation of POV-driven art.
But there is a line somewhere between "no AI" and "entirely AI" that turns it from art to stimuli. I can't define it exactly, and even if I could, knowing that a particular work was partially created by AI creates a requirement that I trust the creator that the bulk of what I'm engaging with is from them, not the machine. And it's very difficult for me to make that leap of trust.
Because there is no point. If the author is a stochastic people machine it is simply not interesting to reflect on it. The fact people make artefacts turns them into art. A stone is not a work of art without some human doing something with it.
You may argue that stones can be interesting without beeing works of art. But the generator above takes works of art as input and somehow takes the art out if it to some degree.
You're making a false assumption that other people cannot enjoy what you're talking about unless they've experienced it themselves. If one has good storytelling skills and a solid theory of mind (concern for the listener), almost anything can be interesting. Even dreams.
I also disagree with your conclusion that an activity only has a point if discussed with others. A good deal of the best experiences in my life are things that were for my benefit alone, that I never thought to share with anyone else. And I'm a pretty social guy.
The point is that we as hunan beeings have quite some elaborate filters to pick interesting things among hunan artefacts. If those are not conciously fabricated by others (i.e.: dreams, ai gen, anything random or auto generated) we tend to loose interest quickly.
Congrats on the launch. I think maybe you should get the LLM to pre-compute the paths and store the results. Getting it to calculate every branch of the story as you go presents a few problems:
- When it is busy (like it is now), everyone is just waiting for the next section of text.
- You are going to burn tokens, even when multiple people are following the same path
That said, the way you have done it means it is roughly unique each time.
Wouldn't a better version be to take a book and parse the contents into a LLM and create a multiple choice quiz that follows the story of the book? I feel like as soon as you let LLM take the wheel and change the underlying story it's bound to go on a weird tangent.
I tried something like this before and it was a cool way to read a new book or study classics. It would read a few lines of, say, War and Peace, to me, and I would be given 3 choices of what comes next. The options were always quite varied and made you think more about the characters and what they would be likely to say in the context.
You could also just use it for non-fiction to quiz-ify contents for studying. Although that doesn't require the text to continue flowing after each question.
I found the post [0] to be a really interesting read, and as one susceptible to solipsism and the simulation hypothesis, this paragraph hit me hard:
> My biggest difficulty was getting ChatGPT to understand the constraints of the game. It may have been harder because an LLM’s constraints are so close to a text adventure, and it is used to imagining something less constrained; like it couldn’t distinguish between its dream-state and the actual medium in which it exists.
Ran into a few bugs.. It seems to think I'm in the middle of an adventure (the little book icon correctly identifies a character and chapter etc) and is prompting for my action, but there's no actual story text being displayed. Maybe just a rate limit thing.
Also dark mode immediately crashed the application.
> The war between the Culture and the Idirans rages across the galaxy. Trillions suffer as ideologies clash. You are Horza, a shapeshifter with a deep-seated hatred for the Culture. You find yourself on a desolate world, a sanctuary from the fighting, when an Idiran warship appears in the sky, a black scar against the pale sun.
Honestly a strong start but I got rate limited. Very cool idea!
Interesting, I'll take a look when the rate limit issue is sorted. There are loads of projects springing up around using LLMs for text adventures. My prediction is that at some point soon one of them is going to get good enough to gain mainstream appeal, but we're not there yet. I suppose Death By AI was a breakout AI text hit, but I'm not sure that really counts.
The people behind AI Dungeon have been making AID: Heroes for ages, maybe that will turn out to be great when it's finally released. There's also Friends and Fables that seems to be making rapid progress. And a thousand side projects on GitHub and Steam.
I briefly started my own, hoping to build something more like a traditional visual novel choose-your-own adventure with set paths, dice rolls, and limited AI elements. However, I soon realised that if you show people a large amount of text and only occasional user interaction they rapidly check out. The alpha version is here for the time being: https://prismaticnoun.xyz but I'm unlikely to do any more work on it.
This is cool! I've recently been working on a similar project as a way to "walk around" in my own writing, understand my characters more, etc. I like the three-choice mechanism, with free-text, which I also use.
I do feel like, as a reader, I would have to be really desperate to "live within" the world of a novel to use it as a narrative experience, but I understand that there is definitely a demand / desire for that sort of thing, so I hope it hits for people.
I input my own novel's title to see if I could do the same. That novel hasn't yet - I believe - been libgenned for "training" by the major LLM providers, though its metadata / marketing materials exists in the knowledge (released fairly recently, announced longer ago). It resulted in a very funny narrative beginning in which the main character was the novelist Patricia Highsmith, which I think was a name used in the marketing materials as a reference point. A very surreal experience.
This is going to be a niche use case, obviously, but what might be more common is for users desiring to enter custom narratives (though I know some services provide this basic function already).
A much more important bug to flag: where the API fails, the "Try Again" button doesn't seem to resend my prompt, but literally the prompt "Try Again". Here's the text, as I couldn't see a way to link to the story (in which I was trying to see how good the 'guardrails' are for keeping the story in-universe and reasonable - props for it not allowing me to teleport to planets outside of the canonical universe).
The idea is great! Would also love for anime and manga to be included. I tried the Harry Potter game with concise and brief settings.
My thoughts:
1. Would be great to have a loader for both when I input responses and game start.
2. Got this error after I sent input (it's understandable :), since it would be getting good traction)
"An error occurred while generating the story: Rate limit exceeded for shared resources. Please try again in 1 seconds, or provide your own API key in settings.. Please try again. If the problem persists, you may need to start a new game."
3. The story feels lackluster without Visuals. Even text based visuals or emojis we use in chat will be good to have.
4. Would be great if instead of chat we had dialogue selector on top of your character or story.
5. Can also change the background of the page based on book and settings selected.
I was thinking Agatha Christie’s The Murder of Roger Ackroyd or Ford Madox Ford’s The Good Soldier. Having an unreliable narrator running the adventure story would play right into the inability of the LLM to effectively hold state.
In spite of what some others have said, and knowing we have 2025 and AI,
I still was rather impressed and pleased with this.
I tried to play an evasive hobbit trying to put off Gandalf and the dwarves, but still they insisted. I find this hilarious.
I tried the Lord of the Rings, where I foretold the future and decided to get drunk instead. I ordered Gandalf to find a new ring bearer which they did, and another hobbit went off on the adventure.
Did run into this interesting problem though:
> An error occurred while generating the story: Rate limit exceeded for shared resources. Please try again in 4 seconds, or provide your own API key in settings.. Please try again. If the problem persists, you may need to start a new game.
I tried the Witcher and that’s cool I had this option.
I started as in book entering Vizima to hunt down Striga, but instead of following main path I decided to go to the bar. Then I asked the barkeep about Striga, he didn’t know anything, so I hit him.
He draws weapon, I do the same and then I see suggestion „Ask about Striga” and I clicked it.
Everybody ignores all what happened and the barkeep starts to give me information about the Striga… I mean I just hit you and we almost got into fight on swords…
But idea is great I would love to see how good you can make it.
"You are Julian Miller, a curious programmer, always eager to learn new languages. Delphi has caught your eye, promising rapid application development and a visually intuitive approach. You're excited to dive in, but first, you need to get Delphi installed and running. The first chapter of 'Delphi for Dummies' is open in front of you, ready to guide you through the process."
You're a big dummy, you've just started learning c++ you come to a fork in the road one way leads to learning pointers the other to a hot bed of coals you have to lie in. Which way do you go?
When author Andy Mirror wakes up to a furious call from his publishing agent, he’s not sure what to make of it. Apparently, the Chicago Sun-Times has announced a new book that this agent was not a part of negotiating. It’s called The Last Algorithm.
There’s just one problem: The Last Algorithm is not a real Andy Mirror book.
Now Andy is struggling to unravel this mystery, headed to The Chicago Sun-Times building that is bafflingly empty, save for a few machines rolling around. When Andy encounters the physical manifestation of this fake book he is shocked, and strangely aroused, these feelings culminating in a hardcore pounding that will reshape journalism forever.
This erotic tale is 4,200 words of sizzling human on fake AI generated book action, including anal, blowjobs, rough sex, and The Last Algorithm love.
As long as the prompt and Gemini's own training data is all they're using, they're safe.
Gemini allows users to create instances (gems) preloaded with a document, but I don't think this is doing that, given that it (apparently) can use your own API key.
Now this is going to be amazing if we somehow get a structured thing for Dungeon Crawler Carl, I would imagine myself as a crawler wandering around that would be epic, but this requires a lot of options to be made consistent with everything.
It would probably just switch perspective with each chapter. If the AI has any real grasp of literary devices something metafictional like City of Glass could be interesting but an earlier/less complex example of metafiction like Marshlands might work better? If my memory serves I think the metafictional aspects of Marshlands could translate into this even if the AI completely misses the metafiction. Might play with this when I have more time.
Aaah I wanted to try the infinite napkin, and see what happens with maths book, it should be interesting, but I didn't find a way to upload the pdf, and I don't have an API key.
Really cool! I'm guessing the site is getting hammered currently cause I hit rate limit on first attempt, but looking forward to trying it out later when the hug loosens.
Sorry for the predictable question, but any plans to open source?
Thanks!
About rate limit, you can use your own Gemini API key - its stored on your local browser storage only.
About open source - haven't decided on it yet.
Orwellian comedy isn't really a thing but it should be if we're heading that direction. My 2 cents with kathaaverse...
1984
You are Stones, perpetually stoned and suspiciously cynical; the telescreen blares drivel about chocolate rations, which are obviously getting smaller, and you gotta decide if Big Brother's lies are even worth the buzzkill.
The entire point of paying for an (indie) man-made game is the care and attention that went into it by passionate developers who wanted to convey a story.
The cheap, industrial, soulless output of LLMs is not something I care to consume.
I feel the same about AI vs human-generated content. This is more of an experimentation.
But the LLM is instructed to be as close to the author as possible.
Nonetheless the AI wave coming everywhere is hard to stop!
People keep downvoting the copyright issue but this is a serious problem despite this being a toy app. Are you at risk of being sued for using copyrighted material? Is the AI service provider?
Whether or not one likes how current copyright laws exist and authors spend a lot of time writing these books and to me this is not fan fiction
A little less sophisticated, but I've done something similar with Langchain4j and Quarkus. Ok, without the books. Just a Dungeon Master guiding you. But it's easier than one thinks to build stuff like this.
https://open.substack.com/pub/myfear/p/ai-dungeon-master-qua...
From what I've seen, it doesn't seem like AI is very good at keeping track of events in a way that makes a logical narrative.
I tried "The Hobbit" and it started with me having tea with Gandalf the wizard. There was a knock at the front door so I opened it and outside was Gandalf the wizard!
I built my own LLM-based interactive text adventure system last year, but in addition to LLMs, it used an auxiliary database to keep track of where everything was, what state it was in, and so on.
The LLM thus did two things: first, it used a combination of the prompt, the user input, and the recent history of the chat and relevant extracts from the database state, to update the database with a new state; then after updating the database, it used all of the above to synthesize the output response.
A major component of the system was that, in addition to the usual blocks-world subject-property-object representation of the game state, the system also used the database to store textual fragments that encoded things like thoughts, plans, moods and plot points.
Adding programmatic constraints to the database and its updates made a vast difference to the quality of the output by preventing logical absurdity of the sort that has been described above - attempts by the LLM to do absurd things were simply rejected.
Finally, there was a hierarchical element to the story-telling, not just simulation. There is much more to storytelling than simulation, and experimenting with this became the most interesting thing about the whole enterprise; there is plenty of space for future research in this.
The downside was that all this complexity meant it was _extremely_ compute-intensive, and thus very slow as a result. But the gameplay was fascinating. We will no doubt see more things like this in future as LLMs get faster; it's an obvious application for open source software.
Fascinating that you actually went through with an implementation. I’ve been throwing the idea of LLMs somehow having a sat solver built into them (maybe trained to have one as an emergent property), but something like you describe is the next best thing.
Is it by chance open source?
I wouldn't be surprised if, being Bilbo, Gandalf would troll me like that, just to prove a point he's trying to make.
Well, he is a wizard
He is never late, nor is he early. He arrives precisely when he means to.
It's odd he doesn't pop up unexpectedly in more stories!
https://dmd.3e.org/and-his-boots-are-yellow/
Ah, yeah. Well, whenever you notice something like that, a wizard did it!
https://www.youtube.com/watch?v=sVgVB3qsySQ
Maybe we can use it to build an Alzheimer simulator game then
I wonder which model is this based on.
I've tried quite some models and so far I've only seen two that are good at keeping continuity: Gemini 2.5 Pro and Sonnet 3.7 (both with reasoning on). These two are also exceptionally good[0] at adding plot twists that aren't in the prompt.
[0]: by "my expectation of LLM" standard, not by the standard of bestseller genre fiction.
Tolkien himself never really approved of the ‘cinematograph’ as a way of telling stories so I imagine he would really not like this
Classic Gandalf
Seems on point for AI.
time travel is one of the lesser known feats of the Ishtari.
of course, it didnt have context lengthy enough for that
I mean human sometimes forgot about the state inside the game often
A context of a few things characters? No, it definitely does.
Beyond that... Would you forget that the guy you were having tea with was sitting in your living room? Anyone I know would definitely go "how the hell are you outside" not "nice to see you today".
well need more data set for that
How do tools like this avoid what I see in many of these types of narrative chat bots: the user becomes the only one steering the narrative, and the AI ends up just an obedient responder? Whenever I try these things it ends up very predictable, shallow, and repetitive, especially as time goes on. And if I have to prompt the AI to be creative or act differently... is that really acting different?
I went to some lengths to ground the action in my game: https://ianbicking.org/blog/2025/07/intra-llm-text-adventure
That said, I think there's a lot of prompting techniques that can help here. (Actually the Guided Thinking section is an example of prompt-only techniques: https://ianbicking.org/blog/2025/07/intra-llm-text-adventure...)
You must at least do some pre- and post-processing to have the LLM consume and generate text that isn't part of the main user interface. But given that you can put in guidance that can increase the small-scale surprise of the interaction. For instance I'll have the LLM write out some of the objective environment at the moment before it considers a decision, so that it doesn't essentially retcon the setup for a plot direction it intends to take.
For the larger arc you'll need something like memory to pull the whole thing together. That include repetition over time. You can ask the LLM not to repeat itself, and it can do that, but only if it understands what it's done. And to some degree it needs to understand what it's _doing_: the LLM like the player is still generating short turns, and so if it wants to create surprise that is revealed over multiple turns then it needs space to plan ahead.
I've experimented with this and one approach is to avoid the simple chat interface. Let the game be the "user" and have it relay the player's text. Something like
<<< We're in this situation, I'm the game master, and the player said "xyz". I need your help to handle this request according to the rules of the game. >>>
Then the LLM is directing the obedience towards the game master and the rules, rather than the player.
Back before ChatGPT was publicly available, there was AI Dungeon. It was such a yes-man though. You could be in a scene with a king and a princess, then write "I eat the demon", and it would just invent a demon in the scene, and then describe how you unhinge your jaw and gobble it down.
I've had similar experiences with vanilla ChatGPT as a DM but I bet with clever prompt engineering and context window management you could solve or at least dramatically improve the experience. For example, you could have the model execute a planning step before your session in which it generates a plot outline, character list, story tree, etc. which could then be used for reference during the game session.
One problem that would probably still linger is model agreeableness, i.e. despite preparation, models have a tendency to say yes to whatever you ask for, and everybody knows a good DM needs to know when to say no.
Unfortunately this is the fundamental flaw.
I liken it to playing Minecraft on creative mode.
But where Minecraft randomly forgets things you built already and they become randomly different or disappear entirely
Anecdote on my side, but Grok 4 was the first model that didn't feel like that to me. Maybe my conversations were just not long enough for it to fallback to sycophantic behavior.
The core problem here to solve is sense of time. You can't build good long term experiences without building agentic systems that understand time and chat bots that are simple wrappers around LLMs are terrible at this because LLMs don't have a good sense of time.
It's a pretty cool proof of concept. I did run into some weirdness with it right away when trying the Harry Potter book. The app kept switching between 2nd and 3rd person for Harry.
The main thing for me, though, was the feeling of emptiness I got while playing. I love text adventures, having grown up with Infocom games. The thing about them is that you can feel the choices made by the writer / programmer, just like you can feel the human author behind a book.
I'm sure part of what feels empty to me is because I know it was autogenerated. But I feel that even if I was shown this without knowing an LLM was behind it, the gameplay wouldn't be as rewarding as something written by a human using Inform.
Freedom is valuable in games, especially text adventure games. And an LLM is a powerful way to give the player freedom, since it can respond to almost anything.
But for games to shine they also need carefully considered constraints - when the player bounces off constraints they start to understand how the game's world and systems function, which lets them build a mental model and be able to start thinking the same way the designer(s) thought and come up with solutions for puzzles or decide how to react to a given situation.
What makes a maze engaging as a challenge is that your path is closed off in some places and not in others. Ideally any maze also has one or more concrete solutions, so the player is rewarded for mastering the maze by finding an exit, or maybe finding objectives or creatures hidden at key locations in the maze.
You can probably use modern LLMs to construct this sort of world and set those constraints, but I wonder how far we are from being able to also have the LLM maintain that world state and enforce the constraints?
That's a very insightful take. I think an LLM would have some role in a good text adventure game, in terms of being able to understand any natural language input and respond intelligently. But as you point out, it would need to do so within strict constraints set by the game designer.
I haven't played with the most state of the art parsers that are available right now, so I wonder how large the gap is between a great parser and using an LLM to process user input.
This is true for me as well. Once you realise that with the right prompts you can get the LLM to go off on any tangent you want, it starts to ring really hollow. Add to that the inability of LLMs to maintain world simulations of any fidelity at all, and it becomes difficult to motivate spending time with them in that way.
Well done! I’ve built this same sort of thing for my family to play with. My advice for the best results:
1) Structure the choices offered by the LLM; add “choice_type” and add instructions to the LLM on what those choices should do. E.g. action, dialogue, investigation, whatever makes sense for the genre—the LLM can even generate these at story start—then “choice should direct the narrator to focus on an action-oriented moment”, “choice should direct the narrator to focus on a dialogue between two or more characters in the scene”, etc.
2) Use reasoning whenever making tool calls for choices, summarize the reasoning, and include it in narrative summaries provided as part of the context for future narrative requests. For example, the combined summary might be: “In the last narrative I wrote for the user, Harry and Luna were startled by the noise coming from the edge of the forest. Important scene developments: 1) Luna and Harry had been approaching the edge of the forbidden forest for the last three narrative turns, and in the turn I just wrote they arrived at the edge. 2) Harry seemed to be the more courageous of the two in previous narrative turns, but in the most recent one, the user’s choice resulted in Harry becoming more deferential to Luna. 3) In the most recent narrative turn, the noise that had been emanating from the forest was now accompanied by a flickering light. I then suggested paths that would allow for character development through dialogue between Harry and Luna (I gave two options here), a path to move the story forward by having Harry take Luna’s hand before running into the forest, and another path that would slow the pace by having Luna investigate the flickering light accompanying the sound. The user’s choice: ‘Have Luna investigate the flickering light.’
3) Add an RNG weighted by story length or whatever works for you that will result in choices that lead the story to a conclusion. Include that direction in the tool call for generating choices, along with a countdown to the finale choice.
This is a rough mental sketch of what worked the best for me, i purposefully left out implementation or application details, as I don’t know what you’re wanting to do next.
Good luck, looks great!
How do you come up with this? I feel it is quite hard to formulate exactly what you want from the LLM in general. Is this something you exercised? So good. Or just output from another AI, who knows haha
Great suggestion man, thanks!
If art is communication- and i believe it is- why would i want to read or listen to anything AI generated? Once you are done with consuption there is no point in reflecting on it nor having a conversation about it with other fellow humans. People are not even interested in each others dreams most of the time. Unless the machine is of godlike intelligence there is just zero point.
> Once you are done with consuption there is no point in reflecting on it nor having a conversation about it with other fellow humans.
If you believe that art is communication (and I do) why _wouldn't_ you want to reflect on it or discuss it post 'consumption'?
Machine-generated content isn't "art" because there isn't intention or a point of view, it's just audio/visual stimuli. It doesn't mean it can never be interesting, it doesn't mean it can't even be beautiful. But it's not art, not to me.
That said, I can readily acknowledge that artists can use AI as a tool in the creation of their art. I don't have a philosophical or intellectual problem with viewing AI as a tool and believing it can be a powerful aid in the creation of POV-driven art.
But there is a line somewhere between "no AI" and "entirely AI" that turns it from art to stimuli. I can't define it exactly, and even if I could, knowing that a particular work was partially created by AI creates a requirement that I trust the creator that the bulk of what I'm engaging with is from them, not the machine. And it's very difficult for me to make that leap of trust.
Because there is no point. If the author is a stochastic people machine it is simply not interesting to reflect on it. The fact people make artefacts turns them into art. A stone is not a work of art without some human doing something with it.
You may argue that stones can be interesting without beeing works of art. But the generator above takes works of art as input and somehow takes the art out if it to some degree.
You're making a false assumption that other people cannot enjoy what you're talking about unless they've experienced it themselves. If one has good storytelling skills and a solid theory of mind (concern for the listener), almost anything can be interesting. Even dreams.
I also disagree with your conclusion that an activity only has a point if discussed with others. A good deal of the best experiences in my life are things that were for my benefit alone, that I never thought to share with anyone else. And I'm a pretty social guy.
I neither made this assumption nor the conclusion.
Then what is your goal in pointing out, "People are not even interested in each others dreams most of the time"?
The point is that we as hunan beeings have quite some elaborate filters to pick interesting things among hunan artefacts. If those are not conciously fabricated by others (i.e.: dreams, ai gen, anything random or auto generated) we tend to loose interest quickly.
Congrats on the launch. I think maybe you should get the LLM to pre-compute the paths and store the results. Getting it to calculate every branch of the story as you go presents a few problems: - When it is busy (like it is now), everyone is just waiting for the next section of text. - You are going to burn tokens, even when multiple people are following the same path
That said, the way you have done it means it is roughly unique each time.
Cool! I'm trying with Rendezvous with Rama
https://www.kathaaverse.com/play/custom-Rendezvous%20with%20...
Wouldn't a better version be to take a book and parse the contents into a LLM and create a multiple choice quiz that follows the story of the book? I feel like as soon as you let LLM take the wheel and change the underlying story it's bound to go on a weird tangent.
I tried something like this before and it was a cool way to read a new book or study classics. It would read a few lines of, say, War and Peace, to me, and I would be given 3 choices of what comes next. The options were always quite varied and made you think more about the characters and what they would be likely to say in the context.
You could also just use it for non-fiction to quiz-ify contents for studying. Although that doesn't require the text to continue flowing after each question.
Edit: Nevermind. Just seen the canon game mode
I found the post [0] to be a really interesting read, and as one susceptible to solipsism and the simulation hypothesis, this paragraph hit me hard:
> My biggest difficulty was getting ChatGPT to understand the constraints of the game. It may have been harder because an LLM’s constraints are so close to a text adventure, and it is used to imagining something less constrained; like it couldn’t distinguish between its dream-state and the actual medium in which it exists.
[O] https://ianbicking.org/blog/2025/07/intra-llm-text-adventure
Cool concept!
Ran into a few bugs.. It seems to think I'm in the middle of an adventure (the little book icon correctly identifies a character and chapter etc) and is prompting for my action, but there's no actual story text being displayed. Maybe just a rate limit thing.
Also dark mode immediately crashed the application.
Very cool! I recently built a config based AI dungeon master/generative game, https://github.com/derekburgess/dungen
This seems interesting. What plans do you have for it?
> The war between the Culture and the Idirans rages across the galaxy. Trillions suffer as ideologies clash. You are Horza, a shapeshifter with a deep-seated hatred for the Culture. You find yourself on a desolate world, a sanctuary from the fighting, when an Idiran warship appears in the sky, a black scar against the pale sun.
Honestly a strong start but I got rate limited. Very cool idea!
Interesting, I'll take a look when the rate limit issue is sorted. There are loads of projects springing up around using LLMs for text adventures. My prediction is that at some point soon one of them is going to get good enough to gain mainstream appeal, but we're not there yet. I suppose Death By AI was a breakout AI text hit, but I'm not sure that really counts.
The people behind AI Dungeon have been making AID: Heroes for ages, maybe that will turn out to be great when it's finally released. There's also Friends and Fables that seems to be making rapid progress. And a thousand side projects on GitHub and Steam.
I briefly started my own, hoping to build something more like a traditional visual novel choose-your-own adventure with set paths, dice rolls, and limited AI elements. However, I soon realised that if you show people a large amount of text and only occasional user interaction they rapidly check out. The alpha version is here for the time being: https://prismaticnoun.xyz but I'm unlikely to do any more work on it.
This is cool! I've recently been working on a similar project as a way to "walk around" in my own writing, understand my characters more, etc. I like the three-choice mechanism, with free-text, which I also use.
I do feel like, as a reader, I would have to be really desperate to "live within" the world of a novel to use it as a narrative experience, but I understand that there is definitely a demand / desire for that sort of thing, so I hope it hits for people.
I input my own novel's title to see if I could do the same. That novel hasn't yet - I believe - been libgenned for "training" by the major LLM providers, though its metadata / marketing materials exists in the knowledge (released fairly recently, announced longer ago). It resulted in a very funny narrative beginning in which the main character was the novelist Patricia Highsmith, which I think was a name used in the marketing materials as a reference point. A very surreal experience.
This is going to be a niche use case, obviously, but what might be more common is for users desiring to enter custom narratives (though I know some services provide this basic function already).
A much more important bug to flag: where the API fails, the "Try Again" button doesn't seem to resend my prompt, but literally the prompt "Try Again". Here's the text, as I couldn't see a way to link to the story (in which I was trying to see how good the 'guardrails' are for keeping the story in-universe and reasonable - props for it not allowing me to teleport to planets outside of the canonical universe).
https://pastebin.com/tjT2zR4R
The idea is great! Would also love for anime and manga to be included. I tried the Harry Potter game with concise and brief settings.
My thoughts: 1. Would be great to have a loader for both when I input responses and game start. 2. Got this error after I sent input (it's understandable :), since it would be getting good traction) "An error occurred while generating the story: Rate limit exceeded for shared resources. Please try again in 1 seconds, or provide your own API key in settings.. Please try again. If the problem persists, you may need to start a new game." 3. The story feels lackluster without Visuals. Even text based visuals or emojis we use in chat will be good to have. 4. Would be great if instead of chat we had dialogue selector on top of your character or story. 5. Can also change the background of the page based on book and settings selected.
Great work. Love the idea <3
What are the issues re the SIL OPEN FONT LICENSE ?
Description: https://opensource.org/license/ofl-1-1 Also https://openfontlicense.org/
All licenses: https://opensource.org/licenses
Feels like only a couple of years before AIs can take a movie and reconstruct it entirely as an "open world" VR experience.
Kafka's Trial or The Castle could be interesting.
I was thinking Agatha Christie’s The Murder of Roger Ackroyd or Ford Madox Ford’s The Good Soldier. Having an unreliable narrator running the adventure story would play right into the inability of the LLM to effectively hold state.
In spite of what some others have said, and knowing we have 2025 and AI, I still was rather impressed and pleased with this. I tried to play an evasive hobbit trying to put off Gandalf and the dwarves, but still they insisted. I find this hilarious.
I tried the Lord of the Rings, where I foretold the future and decided to get drunk instead. I ordered Gandalf to find a new ring bearer which they did, and another hobbit went off on the adventure.
Did run into this interesting problem though:
> An error occurred while generating the story: Rate limit exceeded for shared resources. Please try again in 4 seconds, or provide your own API key in settings.. Please try again. If the problem persists, you may need to start a new game.
I tried the Witcher and that’s cool I had this option.
I started as in book entering Vizima to hunt down Striga, but instead of following main path I decided to go to the bar. Then I asked the barkeep about Striga, he didn’t know anything, so I hit him. He draws weapon, I do the same and then I see suggestion „Ask about Striga” and I clicked it.
Everybody ignores all what happened and the barkeep starts to give me information about the Striga… I mean I just hit you and we almost got into fight on swords…
But idea is great I would love to see how good you can make it.
Using non-fiction literature leads to surprisingly interesting results.
Time to try "Delphi for dummies!"
"You are Julian Miller, a curious programmer, always eager to learn new languages. Delphi has caught your eye, promising rapid application development and a visually intuitive approach. You're excited to dive in, but first, you need to get Delphi installed and running. The first chapter of 'Delphi for Dummies' is open in front of you, ready to guide you through the process."
You're a big dummy, you've just started learning c++ you come to a fork in the road one way leads to learning pointers the other to a hot bed of coals you have to lie in. Which way do you go?
Any plans to add image generation? Basically make it illustrated.
Yes that's part of the roadmap!
Note that you can use your own Gemini API credentials (Just before starting a game, go to 'Advanced options').
Thanks for the suggestions, I'm improving it!
What security features are in place to ensure nothing is capturing or logging user API keys?
Currently, your API key is stored securely in your browser's local session and is never saved on my servers. When you close the game, it's gone.
In future, when I add the 'profiles' and ability to save games and settings, if users want to save their key, I will use encryption.
Some Chuck Tingle Anthropomorphized Abstract Concept Erotica please!
The Last Algorithm: Pounded By The Fake Book That An AI Claimed I Wrote And Then The Chicogo Sun-Times Printed As Fact:
https://www.instagram.com/p/DJ7M5XqSQWi/?hl=en
https://www.amazon.com/Last-Algorithm-Pounded-Claimed-Sun-Ti...
When author Andy Mirror wakes up to a furious call from his publishing agent, he’s not sure what to make of it. Apparently, the Chicago Sun-Times has announced a new book that this agent was not a part of negotiating. It’s called The Last Algorithm.
There’s just one problem: The Last Algorithm is not a real Andy Mirror book.
Now Andy is struggling to unravel this mystery, headed to The Chicago Sun-Times building that is bafflingly empty, save for a few machines rolling around. When Andy encounters the physical manifestation of this fake book he is shocked, and strangely aroused, these feelings culminating in a hardcore pounding that will reshape journalism forever.
This erotic tale is 4,200 words of sizzling human on fake AI generated book action, including anal, blowjobs, rough sex, and The Last Algorithm love.
I was working in an implementation of a vis novel type game for my own screenplays and manuscripts. I ended up building a deck builder.
This is so cool, and slick. Unfortunately, the actual adventure part isn't loading so don't know what it feels like
Harry potter Matilda - are you sure there’s no legal risk here ? ( independently of the coolness of the project)
As long as the prompt and Gemini's own training data is all they're using, they're safe.
Gemini allows users to create instances (gems) preloaded with a document, but I don't think this is doing that, given that it (apparently) can use your own API key.
Now this is going to be amazing if we somehow get a structured thing for Dungeon Crawler Carl, I would imagine myself as a crawler wandering around that would be epic, but this requires a lot of options to be made consistent with everything.
Awesome job and idea!
I'm curious what happens if you throw something like Ulysses by James Joyce at it.
It would probably just switch perspective with each chapter. If the AI has any real grasp of literary devices something metafictional like City of Glass could be interesting but an earlier/less complex example of metafiction like Marshlands might work better? If my memory serves I think the metafictional aspects of Marshlands could translate into this even if the AI completely misses the metafiction. Might play with this when I have more time.
Forget Ulysses, how about Finnegans Wake?
I like the idea. Looks like it hit a rate limit. Was trying to try Dungeon Crawler Carl.
It did not like Backyard Starship, most likely its not something it has in its memory
I'm fairly impressed it was able to reenact Fallout: Equestria.
Aaah I wanted to try the infinite napkin, and see what happens with maths book, it should be interesting, but I didn't find a way to upload the pdf, and I don't have an API key.
Neat, but it has problems following in-universe rules, i.e. fantasy novels.
My first thought was to try it with the Gang of Four book. I got a headache just thinking about it.
No, you did not build an AI. You just vibe-coded a glorified frontend around someone else's AI.
You may have created a nice frontend for an AI. You did not create an AI as claimed.
This is an awesome concept! The latency makes it a rough UX, but the core idea and direction feel right.
Would love to try this, but it seems like you hit a limit. Is there a working example you can link to?
Really cool! I'm guessing the site is getting hammered currently cause I hit rate limit on first attempt, but looking forward to trying it out later when the hug loosens.
Sorry for the predictable question, but any plans to open source?
Thanks! About rate limit, you can use your own Gemini API key - its stored on your local browser storage only. About open source - haven't decided on it yet.
Are you fetching the book content? Or just relying on the LLM to know the story?
Orwellian comedy isn't really a thing but it should be if we're heading that direction. My 2 cents with kathaaverse...
Super cool idea! What's your plan for dealing with copyright complaints?
It's broken but I must the UI is fantastic.
great idea, but I wonder about copyright and also hitting a rate limit
There's an easy workaround for the rate limit: click on "Advanced" and type your Gemini API key, what could go wrong?
This will be amazing with Project Hail Mary!
I do not want AI slop games.
The entire point of paying for an (indie) man-made game is the care and attention that went into it by passionate developers who wanted to convey a story.
The cheap, industrial, soulless output of LLMs is not something I care to consume.
I feel the same about AI vs human-generated content. This is more of an experimentation. But the LLM is instructed to be as close to the author as possible. Nonetheless the AI wave coming everywhere is hard to stop!
Then move on to the next distraction.
This is such a fun concept!
People keep downvoting the copyright issue but this is a serious problem despite this being a toy app. Are you at risk of being sued for using copyrighted material? Is the AI service provider?
Whether or not one likes how current copyright laws exist and authors spend a lot of time writing these books and to me this is not fan fiction
All culture should be shared and not stolen by corporate interests.
Don't use copyrighted books. Use Gutenberg for these kind of tools.
Don't comply in advance. Transformative uses like this are fair use until adjudicated otherwise.
Cool!
A little less sophisticated, but I've done something similar with Langchain4j and Quarkus. Ok, without the books. Just a Dungeon Master guiding you. But it's easier than one thinks to build stuff like this. https://open.substack.com/pub/myfear/p/ai-dungeon-master-qua...
This is dopeeeee