How to run a tabletop RPG with an AI game master

Yes — ChatGPT, Claude, and Gemini are good at the improvising part. What they're bad at is refereeing: they forget state, wave through every plan, and end the story early. A game card fixes that by taking three jobs off the model — a state ledger it reprints every turn, fixed consequences for failure, and dice it doesn't get to choose.

Why does ad-hoc prompting drift?

An open "be my dungeon master" prompt drifts because a chat is not a game state, and a model tuned to be agreeable makes a poor referee. Three failures recur. Forgotten state: the wound you took two scenes ago heals itself, and the lantern you dropped is back in your pack. A soft game master: every plan works, because saying no feels unhelpful. Invented endings: the climax arrives when the prose peaks, not when you earned it.

What does a structured game card fix?

A game card fixes drift by externalizing the three things a language model handles worst: state, consequence, and randomness. Instead of trusting the model to remember, the card makes it maintain a state ledger — a short, visible block it reprints each turn listing where you are, what you carry, and what you've spent — so nothing quietly disappears. Consequence menus do the refusing for it: each scene ships a fixed list of costs, and on a missed roll the game master must take one — a struck item, a ticked clock, a suspicion that starts counting. The model never gets to decide that nothing happened. And randomness never comes from the model's taste. If your assistant can run code, it rolls there and shows the number. If it can't, the card carries printed dice — a tape seeded from the game's own id and version, printed on the card, read left to right, with the position tracked in the ledger. It's deterministic, which is what makes it checkable: the same seed always gives back the same tape.

Here's the whole machine in one turn. Wren Calloway — a Whisper, Shadow 3 — tries to search the toll booth unseen: hard, DC 14. It's the third roll of the night, so the game master takes the next value off the printed tape, tape#3=3, adds 6 for Shadow 3, and shows 9. A miss by five. The booth stays shut, and the card leaves the model nowhere to soften it: a miss costs one item from that scene's fixed menu, so the Owed clock ticks and the river Sorrow comes a little closer to the road. Then the reply ends where every reply ends:

[L] turn:5 1.1 @toll-gate | Strain 0/6 | inv: coil-rope, dark-lantern, courier's-writ, heel-knife | flags: none | clocks: Owed 1/6 | tape:4 | v1.1.0

You can check that 3 yourself — position 3 of The Toll at Blackbridge, on the tape verifier. Nothing in that turn was improvised, including the failure. The rest of what a card takes out of the model's hands is laid out in how our game cards work.

How do you start a game in ChatGPT, Claude, or Gemini?

One tap. On any game page, Open in ChatGPT or Open in Claude starts a new chat with the play prompt already written; send it and the assistant fetches the card itself — Claude wants web search switched on to do that. Gemini is fussier about prefilled links, so there you use Copy the play prompt and paste it into a new chat. If your assistant can't browse the web, copy the full game card and paste the whole thing instead — it's self-contained. Either way it's a fresh conversation, no plugins, no account beyond the assistant you already use; the tested compatibility matrix covers the small differences in how each app handles links, dice, and saves.

Then the AI takes the chair. It greets you, confirms which dice it's using — its own code tool, or the tape printed on the card — and offers three ready-made characters: in Blackbridge, a courier who works in shadows, a staff-and-chain warden, and an archivist who reads what she shouldn't. Pick one and the night opens at the toll gate, the queue stalled, the keeper counting coins that aren't coins. No character creation, no rulebook.

A good first run is The Toll at Blackbridge, a self-contained tutorial one-shot that explains each rule the first time the night calls for it. Finished it and not ready for a campaign? The Gilt Horologe is a one-night heist inside a clockwork vault winding itself shut around you — same rules, a clock working against you the whole way. When you do want the longer arc, The Severed Spire is Chapter 1 of a three-chapter Ironquill campaign, 90 to 180 minutes inside a mile of tower that hangs above its own sheared-off mountain and sinks an inch a day. Chapters 2 and 3 aren't out yet; the save you take at the end of Chapter 1 carries your company into them when they land. All three live on the tabletop RPG shelf.

Can you play solo, or do you need other people?

You can play entirely solo — the AI runs every non-player character, so a second person is optional, not required. Solo is the default: it's just you and the game master, which makes pacing easy and lets you stop and resume whenever. A duo works too, with two players sharing the party and taking turns addressing the AI; it adds table talk and disagreement, at the cost of coordinating whose turn it is to type. Bramblewood, the all-ages card, goes up to four: a woodland festival for a kid and a grown-up reading along, no violence and nothing frightening in it. Party size never changes the machinery — the same ledger and the same printed dice keep the game honest at any table.

How do you run a campaign over several sessions?

You never keep one chat going forever. Ask for a save at any point, or take the one the game offers when a scene or chapter ends, and it prints a plain-text block holding your character, strain, gear, live clocks, and dice position. Open a fresh chat and paste it: a new chat per chapter is what stops a long campaign drifting as the conversation fills, and the block isn't tied to one app, so ChatGPT to Claude works. The save and resume guide takes the block apart field by field, and the resume tool turns one into a paste-ready prompt.

Frequently asked questions

Can ChatGPT be a dungeon master?

Yes. ChatGPT can act as a dungeon master and it is genuinely good at description, voices, and improvising a scene. Left to an open prompt it drifts — it forgets what happened, waves through your choices, and writes its own ending. A structured game card with a state ledger and fixed consequences keeps it consistent for a whole session.

How do I save an AI RPG session and pick it up later?

Type save. The game prints a short plain-text block between BEGIN AIGC SAVE and END AIGC SAVE holding your character, chapter and scene, strain, gear, live clocks, and your place in the dice tape. Copy that whole block, open a fresh chat — the same assistant or a different one — paste it, and ask the game to load it. Ironquill games also offer you a save whenever a scene or chapter ends.

Which AI is the best game master?

ChatGPT and Claude are the smoothest game masters today. Gemini plays every game just as well; its friction is at the door — prefilled launch links are unreliable there, so use the Copy the play prompt button and paste it into a new chat. The larger factor is the game card, not the model — a well-structured card runs consistently on all three. The compatibility matrix shows what we actually see when we test each app.

Do I need to know the rules before I play?

No. The game card carries the rules and the AI teaches them as you go, so you can start cold. The Toll at Blackbridge is built as a tutorial one-shot for exactly this: it walks a first-time player through the Ironquill system in a single sitting.

Is it free to play?

Yes. Every game here is free with no signup, and a free tier of ChatGPT, Claude, or Gemini is enough to run one. You load a game card into your own assistant; there is nothing to buy from us.