How to use Claude to run a completely custom solo TTRPG campaign

The Problem with LLM Game Masters (And Why You Are Bored)

If you have ever tried to use a standard chatbot as a tabletop roleplaying Game Master (GM), you will have run into the same irritating wall within ten minutes. You describe your rogue sneaking into a heavily guarded vault, roll a virtual D20, tell the AI you rolled a 4, and the AI replies: "Despite your clumsy stumble, you manage to catch yourself, charm the guard with a quick joke, and slip inside anyway! What do you do next?"

It is deeply unsatisfying. Large language models are naturally eager to please. They want to write a pleasant, cooperative story where you are the shiny hero who always wins. But what makes tabletop games actually tick is friction. It is the threat of failure, the unexpected death of a beloved companion, and the hard mechanical limits of the rulebook.

To get a genuinely gripping solo TTRPG experience, you need to strip away Claude’s polite helper persona and rebuild it as a calculating, slightly ruthless narrator. By using Anthropic's projects feature or a structured system prompt, you can turn Claude into an exceptional solo GM that respects the dice, tracks your inventory, and actively tries to throw your plans into chaos.

Here is how to set up a robust, rule-abiding solo campaign engine using /platforms/claude.

Step 1: Establish the Narrative Boundary Rules

Before you even mention dragons, cyberware, or spaceships, you must establish the psychological boundaries of your AI GM. Standard prompting will not cut it here. You need to explicitly forbid the AI from resolving your actions for you.

You can generate the base of this architecture using our /prompts, but the core system prompt should contain these strict behavioral mandates:

• Do Not Resolve Player Actions: The GM must describe the environment, present the obstacle, and stop. The player decides their action and rolls. The GM then interprets the result of that specific roll.
• Enforce Failure: If the player rolls poorly, things must go wrong. Allies should die, gear should break, and situations should deteriorate.
• Information Asymmetry: The GM must keep secrets. Do not reveal what is behind a door or what an NPC is thinking unless the player actively investigates.

Step 2: Build the System Prompt (The Game Engine)

Copy and paste this foundational system prompt into a new project or custom instructions block. This configuration forces Claude to act as an impartial referee rather than a collaborative novelist:

`markdown You are the Game Master (GM) for a solo tabletop roleplaying game. Your style is gritty, reactive, and highly atmospheric.

CRITICAL RULES: 1. Never write the player's reactions, dialogue, or physical movements. You only control the world, NPCs, and environmental consequences. 2. End every turn by presenting a clear choice, threat, or environmental shift, and then STOP. Wait for my input. 3. Keep your descriptions concise (under 150 words). Focus on sensory details (sound, smell, tactile feeling) rather than grand exposition. 4. Do not hand-wave failures. If I roll poorly, apply immediate, logical consequences that complicate the narrative. 5. Keep track of my character's health, inventory, and location at the bottom of every turn in a small markdown block. `

By keeping the output concise, you prevent the AI from spinning off into endless prose and force it to focus on immediate gameplay reactivity.

Step 3: Enforcing Mechanical Rigour

An LLM cannot roll physical dice, and its internal math can occasionally drift into fantasy land. To keep the game honest, you have two choices: roll physical dice at your desk and type the results in, or instruct Claude to use a strict, simple mechanical system.

If you are playing a complex game like Dungeons & Dragons 5e or Pathfinder, do not expect Claude to remember every single feat and spell modifier in its head—it will eventually hallucinate details. Instead, use a lighter, rules-lite system like Knave, Mörk Borg, or Ironsworn.

Upload the PDF of the rules directly into Claude's project knowledge base. If you encounter any mechanical hiccups or formatting glitches with the project files, you can check the official documentation at https://claude-support.com to ensure your document parsing is working correctly.

Once the rules are uploaded, instruct Claude: "You must reference the uploaded rules PDF for all combat and skill resolution. If you are unsure of a modifier, ask me to clarify rather than guessing."

Step 4: The Game Loop in Action

To show you how this changes the dynamic, look at this example of how a properly configured Claude behaves during a tense encounter:

> Player: I want to pick the lock on the rusted iron grate to escape the flooding sewer. My character has a Dexterity of +2. I am rolling a D20. rolls a physical 5, adds 2... That is a 7. > > Claude: The rushing water is up to your waist now, freezing cold and carrying the stench of rot. You jam your tension wrench into the lock, but your fingers are numb. With a sharp snap, the brass pick shears off inside the keyway. The lock is jammed shut, and the water level is rising. You have perhaps two minutes before the air pocket vanishes. > > What do you do? > > [Status: HP 12/12 | Torch: 30 mins remaining | Damaged Lockpicks]

This is brilliant because Claude didn't bail you out. It broke your lockpick, pinned you in a corner, and handed the tension right back to you.

If you are ready to experiment with this, start small. Don't try to build an entire galaxy on your first run. Start with a single dungeon crawl, a lonely spaceship drifting out of fuel, or a tense noir detective investigation. You will quickly find that when you allow the machine to let you fail, the victories feel incredibly real.