More people try to run D&D in ChatGPT than in any purpose-built tool, and most of them have a great first session. This is the honest guide to doing it well: the setup that works, the failure points that end campaigns, and how to tell when you have outgrown it. We make a purpose-built AI dungeon master ourselves, so read the recommendations knowing that, but the setup advice below is the same advice we would give a friend.
The setup that actually works
ChatGPT defaults to being an agreeable assistant, which makes a terrible DM. The fix is a role prompt that gives it a spine. Structure it in three parts: the job, the rules of engagement, and the state it must track.
A working session zero prompt looks like this:
You are a D&D 5e dungeon master. Run a campaign for one player.
Rules of engagement:
- Never decide my character's actions or dialogue.
- Ask for a roll before narrating any uncertain outcome, and state the DC first.
- NPCs have their own goals and can refuse, lie, or fight.
- Failure is allowed. Do not soften consequences.
Track after every scene, in a block at the end of your reply:
HP, spell slots, inventory changes, gold, active quests, and NPCs met.
Start with session zero: ask me five questions about my character
and the kind of campaign I want, then open scene one.
Two details in that prompt do most of the work. Stating the DC before the roll stops retroactive difficulty, and the end-of-scene state block gives you something to paste back in when the conversation resets.
For combat, ask it to run initiative explicitly: list combatants in order, resolve one turn at a time, and show enemy HP as descriptions (bloodied, staggering) rather than numbers if you want tension. Roll your own dice physically or with a dice app, and report the result. ChatGPT will happily generate a "roll" for you, but a number it writes is a narrative choice, not a random outcome.
For state, the workarounds are manual but effective for a while: re-paste your character sheet at the start of each session, end each session with "give me a structured recap of everything that changed," and keep the recap file yourself. Custom instructions or a Project can hold your standing rules so you only re-paste state, not setup.
Where it breaks, precisely
Every one of these failures comes from the same root cause, so it is worth stating plainly. ChatGPT stores your campaign in the context window, not in a game state. There is no character sheet object, no dice ledger, no quest database. There is a rolling transcript, and when the transcript outgrows the window, the oldest facts silently stop existing.
The practical symptoms arrive in a predictable order:
- HP, spell slots, and gold drift after a few thousand words. The numbers in prose stop matching the numbers you wrote down, and there is no authoritative copy to appeal to.
- Dice are generated text, not committed random rolls. The model writes whatever number serves the story it is already telling, and it is biased toward letting you succeed.
- Dead NPCs come back, and consumed items return. Killing a villain or drinking a potion is a sentence in the transcript, not a change to a record, so nothing prevents a later contradiction.
- Rulings contradict each other across sessions. Without a rules engine, every adjudication is improvised fresh, and session twelve does not remember the ruling from session three.
- The campaign quietly resets rather than visibly failing. You will not get an error. You will get a slightly wrong detail, then a missing NPC, then a plot thread that no longer exists.
Memory features, Projects, and summaries delay all of this. They do not change its nature, because they store prose summaries rather than structured state. A summary that says "the party is wounded" does not know your cleric has 7 HP.
Run the five-prompt memory test
You do not have to take our word for where the line is. The five-prompt AI DM memory test is designed exactly for this: it checks delayed recall, item consumption, character death continuity, quest consequences, and whether the system stops an illegal action, in five prompts you can paste into any tool. Run it against your ChatGPT campaign at session three and you will find the drift before it finds you.
When to switch to a purpose-built tool
Switch when you catch yourself doing the bookkeeping you wanted a DM to do: re-pasting sheets, arguing about what the transcript said, or re-rolling "impossible" successes. That is not a prompting failure on your part. It is the architecture.
A purpose-built tool splits the jobs. The game engine runs the game, and the AI DM narrates on top of it. In TableForge, dice, HP, spell slots, conditions, and combat resolve in deterministic code, campaign state persists across sessions, and play works solo, with one friend, or with a group up to six. The full architectural comparison is in AI dungeon master vs ChatGPT, and the named comparison of ten AI dungeon masters covers the rest of the field.
If you prefer Claude's prose, everything in this guide applies there too, and we wrote the same honest assessment in Claude as a dungeon master. If you would rather build the stack yourself, running your own AI DM covers what that takes.
ChatGPT is a genuinely good narrator and a fine way to discover that you enjoy this kind of play. Enjoy the one-shots. When you want the campaign to hold together, use a tool that keeps the game outside the conversation.