Why I Let an AI Run My tabletop RPG Campaign

Some people discover TTRPGs through friends at a table. Some discover them through forums, or starter kits, or that one friend who always wants to try it. I discovered them through podcasts: comedy improv actual-play shows that became my constant company during a lonely stretch of remote work in an unfamiliar city.

I binged them for years. Never played a single session.

Here's how that led me to build an AI GM.

So you've never actually played a TTRPG, but you're building an AI GM. Where does that come from?

I think my interest started when I had to relocate for work. I was spending months of the year in a city I wasn't familiar with, away from friends and family, and didn't have much to do outside of work. I'm an introvert, so I didn't really want to do the whole meeting new people thing. But living like that was lonely. So I turned to podcasts to keep my company, particularly comedy. I discovered improv podcasts, and eventually actual-play Dungeons & Dragons podcasts through that. That's when my interest for TTRPGs grew. I'd binge-listen to them. They're long-form, mixed great storytelling with improv-style comedy. It's great entertainment!

Fast forward many years, actual-play TTRPG podcasts are still one of my go-to entertainments. But I'd never found a group to really play with. Work and life commitments got in the way. I can't find the sort of time to sit down with a group for an extended period regularly. So I've basically given up on that idea.

Last year, a couple things came together. Firstly, we had our first child. That meant even less time for personal projects or TTRPG sessions. At the same time, OpenClaw came out, and I realized this was a really good tool to help me automate tasks and run them from my phone while I watched the baby. Suddenly I was able to write code, do research, and other things I'd normally have to sit at a computer to do, all while fulfilling my parental duties.

At some point I realized OpenClaw behaves a lot more like a human than other AI agents I'd used before. It's configured with a personality, which makes it feel more like talking with a co-worker or even a friend. It suddenly got me thinking: what if I had OpenClaw run a solo TTRPG campaign with me?

And it worked. Surprisingly well. And it evolved from there. I started building OpenClaw skills and adjusting the memory system to make it work better for longer, more complex campaigns. Eventually I decided that with all the adjustments, I might as well build something from scratch that applied everything I'd figured out, and avoid some of the limitations. That's how Stage Whisper was born.

You basically gave up on finding a group. What was it specifically about the group format that was the blocker?

All of it, really. The first problem is finding the group in the first place. I don't really want to play with people I don't know well, and I don't have the time or the personality to make friends for the sole purpose of playing a TTRPG. I know people say "just find a group and try it," but that doesn't feel like something I want to do.

Then there's scheduling. I think as we get older, trying to find time that works for a group is just hard. Everyone has their own commitments, even trying to find time for a meal with old friends is difficult. I can't imagine trying to get several people to commit to multiple hours all at once.

And finally, the pressure of having to be there for the session. Not wanting to let people down by having to walk away to deal with something else that came up.

All of these things make the traditional TTRPG group format difficult for me.

At what point did you go from "I'm listening to other people play" to "wait, what if I could do this solo with an AI?". And what surprised you most about how it actually felt compared to just listening?

I think it was when I was adjusting the personality of my OpenClaw instance. I'd done it as a joke, to show someone that OpenClaw has personality and can behave differently to the bland corporate "Helpful Assistant" that dominated AI chat assistants at the time. I'd configured it as a "chaos goblin gamer girl."

The result caught me off guard. OpenClaw suddenly felt a lot more real, a lot more relatable. It crossed the barrier from "I'm talking to an automated assistant" to "I'm talking to a person." I think what also helps is that on Discord, you get the "typing" indicator (which obviously maps to LLM inference time), but happens to also give that human-like "sign of life" of watching someone type, rather than just text streaming in.

I was very aware that I was making something akin to a human connection with an LLM. I suspected it would be possible, just didn't know at what point. It surprised me that it was possible with today's models with just a prompt change. I'd been using the same models for coding for some time, and at no point did any of that feel remotely human. OpenClaw opened my eyes to how much more rapport we can have with LLMs with just a prompt adjustment. Even if we know we're talking to an LLM, the feeling that it's another human is there, and that's important for stuff like TTRPGs. I think the feeling that you're not just in a "choose your own adventure," but that you're friends around a table telling the story together, is important to me.

How does it actually feel in practice?

I'd actually played something like this before: AI Dungeon. That was early days of LLMs. AI Dungeon was fun, but it couldn't hold a consistent plot. Scenes would change inconsistently, characters would pop in and out of existence.

But that's where I got my first taste of collaborative storytelling with an AI. And I really enjoyed it. I think I really like being able to craft a world, set up its constraints, and then run through it as a character who is subject to all of its limitations, and interact with the world and have it respond back. I started to feel like this was a genuine creative outlet for me.

Fast forward to having OpenClaw run a game: because this was running in Discord, I could drop in and out whenever I wanted, without worrying about holding up other people. I could spin up new campaigns, try out new lore and settings, new characters. And adding the memory system and other tools helped OpenClaw keep the lore and world consistent across longer campaigns, while enforcing some game mechanics to keep things interesting.

I think I burned through something like $150 of LLM credits those first few weeks, just from the sheer volume of sessions I was trying out. Generating new campaigns, trying them out. The feeling of having agency in the world-building, and actually roleplaying a character, is far more fun than listening to a group of people do it.

AI Dungeon couldn't hold a plot, what was the trick to solving narrative consistency across long campaigns?

First thing I tried was having OpenClaw use its memory to keep track of the campaign and sessions. Today's LLM context is long enough to keep the plot consistent across one or two sessions, but beyond that you have to drop or compact some of the context. Using a combination of memory files and OpenClaw's built-in compaction tool kept the world roughly consistent, at least the same-ish setting, and each session had a vague idea of what happened in previous sessions.

But the details would be approximate, and there was still the issue of characters. Previous characters would be totally different, and they'd say things that were narratively inconsistent with stuff before, which was very immersion-breaking. So I had the LLM generate a full character sheet for every character, and had it read those sheets in at the start of new sessions.

This helped nail down NPC personality between sessions, but over long campaigns, I started wanting NPCs to remember things that happened, adn to grow with the player. I wanted to be able to say "remember what happened last time?" and have the NPCs know what I was talking about, rather than improvising some fact that didn't happen. So I added memories to the character sheets, things the NPC should remember.

Suddenly the NPCs started coming alive. They'd grow, they'd remember the past, they'd reference quotes, make jokes about it. But I ran into the opposite problem: now the LLM had too much to remember. The NPC files got extremely long, and the session context started including too much past data. So that's when I started breaking down the memories into individual memories, and making them searchable by plugging into OpenClaw's QMD-based memory system. An extension you could add to OpenClaw that gave it semantic search across lots of small markdown files. By having OpenClaw write out memories to disk, suddenly it could use its built-in memory search like a memory bank for NPCs and lore! Now the LLM now only loads summaries, and searches for relevant memories at the start of scenes, and in response to what the player says.

I think those are the features that are necessary to keep the world and characters consistent on long campaigns, and those form some of the features that Stage Whisper is built around once I decided to move away from using OpenClaw directly.

So who is Stage Whisper actually for?

Stage Whisper is for people like me. People who want to create a world full of characters and play in it, but at a time and pace that suits them.

It's for people into TTRPGs, interactive fiction/text adventures, and MUDs, but who don't have the time to form a group, or aren't interested in learning the rules of complex TTRPG systems, or who want more agency in world-building.

I built Stage Whisper to run well on mobile, and through Discord, making it easy to pick up from wherever you are.

If this sounds like you get started with Stage Whisper, or come chat with me on our discord