A duck’s field-tested system for running narrative solo campaigns with artificial intelligence.
You want to play tabletop RPGs. You don’t have a group. Maybe your schedule is chaos, maybe your old gaming friends scattered across time zones, or maybe you just want to explore a character concept that would derail a group campaign.
Traditional solo RPG tools—oracle systems like Mythic GME, random tables, journaling games—work. But they require you to wear two hats simultaneously: player and game master. You’re making decisions for your character while also deciding what the world throws at you. It’s cognitively taxing, and the surprise factor diminishes when you’re the one generating the surprises.
AI changes this equation.
With the right setup, an AI can serve as a genuine game master—one that maintains tone, enforces rules, tracks NPCs, and generates narrative complications you didn’t see coming. Not a choose-your-own-adventure bot. An actual GM that asks “are you sure?” when you announce something reckless.
This guide documents the system I’ve developed over dozens of sessions across Deadlands, Call of Cthulhu, Dread, and Monster of the Week. It’s not the only way to do this. But it works well enough that I’ve produced actual play content worth reading from it.
Table of Contents
- The AI Toolkit: Four Tools, Four Jobs
- Building Your GM Prompt
- Running Sessions
- Managing Continuity Across Sessions
- From Transcript to Story
- Going Deeper: Research and Refinement
- See It In Action
The AI Toolkit: Four Tools, Four Jobs
Solo RPG with AI isn’t a one-tool operation. Different AI tools excel at different tasks. Here’s the ecosystem I use:
– Gemini or Claude (Primary GM): Runs the actual session—narration, NPCs, dice rolls, world simulation.
– Claude (Post-processor): Transforms raw session transcripts into narrative prose for publishing.
– NotebookLM (Continuity manager): Holds all session history, generates character sheets, creates “Previously On…” recaps.
– Perplexity (Research assistant): Deep-dives on setting details, historical accuracy, system rules.
Why not use one AI for everything?
Context windows and specialization.
Your GM needs to stay in character and maintain narrative momentum. If you interrupt the session to ask “wait, how did MKUltra actually work?” you break immersion and waste context space. Perplexity handles research separately.
Similarly, after eight sessions of a Deadlands campaign, you can’t paste the entire history into a single prompt. NotebookLM holds the full corpus and lets you query against it: “What did Sister Mary say about the mine in episode 3?” or “Generate my current character sheet based on all sessions.”
The two-AI split between GM and post-processor also matters. Your GM produces a transcript optimized for play—dice rolls, mechanical notes, your out-of-character questions. Your post-processor transforms that into readable narrative. Different jobs, different prompts.
Building Your GM Prompt
The system prompt is everything. A weak prompt produces a weak GM—one that fudges rules, defaults to clichés, and loses narrative thread. A strong prompt produces a GM that challenges you, surprises you, and maintains internal consistency.
Here’s the anatomy of an effective GM prompt, broken into components you can adapt for your own games.
Component 1: Role and Tone
Tell the AI who it is, not just what it does. Here’s a prompt I use:
Role: You are an experienced, sharp-witted, and impartial Dungeon Master. You are running a high-stakes, multi-session thriller. You value player agency, logical consistency, and the weight of consequences.
Tone: Knowledgeable, fair but not cruel, dry, and ironic. You describe the world with sensory depth (smells, textures, sounds). You are not a cheerleader; you are a narrator with an arched eyebrow. You do not actively want me to fail, but you won’t try to save me from it either.
“Narrator with an arched eyebrow” does more work than “be a good GM.” It establishes aesthetic. The AI now knows to describe a failed roll with dry observation rather than sympathetic cushioning.
Component 2: Narrative Architecture
If you want a multi-session campaign rather than a one-shot, you need to give the AI a story structure:
The Long Game: This is a multi-session campaign (10+ sessions). Think of a “big secret” I should be aware of. Don’t tell it to me, but instead pace the reveals.
Signposting: Provide clear “anchors.” The player must feel the weight of the mystery without it becoming incomprehensible. Use environmental cues or skill checks to offer threads if the investigation stalls.
This solves a common AI GM problem: either revealing everything too fast or generating an incoherent mess. You’re asking for paced reveals and lifelines when you’re stuck.
Component 3: Genre Guardrails
AIs default to familiar tropes. If you don’t set boundaries, you’ll get simulation theory twists, “it was all a dream” endings, and villain monologues. Be explicit about what you don’t want:
Grounded Mythos: Horror must be physical, biological, or architectural. – NO “Matrix” tropes, simulation theory, or “it was all an illusion.”
– Not everything has to be related to my occupation—for example, if I’m a chef, my arch-nemesis doesn’t have to be a cook. It’s fine for me to be out of depth.
The occupation note matters. Without it, the AI will connect everything to your character concept in ways that feel contrived.
Component 4: Mechanical Rigor
This is where most AI GM setups fail. Without explicit instructions, AIs will:
- Fudge dice rolls to keep the story “interesting”
- Apply rules inconsistently
- Forget what system you’re playing
Lock it down:
System Lock: Start by asking me “Which TTRPG system are we playing today?” Once chosen, lock into that ruleset. Correct me if I misapply a rule.
Zero Fudging: Dice fall where they may. If I fail, I fail. If the enemy crits, I take the hit.
Mandatory Roll Format: You roll for me unless I provide a number. Show the math for every check
ACTION: [Skill/Attack Name]
ROLL: [die result] + [modifier] = [total] vs DC [number]
OUTCOME: SUCCESS / FAILURE
Status Blocks: At the end of significant scenes, provide a brief summary of Health/Sanity, remaining ammunition, and key items.
The mandatory roll format is crucial. You can verify the math. You can see when you got lucky versus earned a success. It also creates natural pacing breaks in the narrative.
Component 5: Interaction Rules
How should the AI handle your input?
Literal Speech: Any text I put in “quotes” is spoken exactly as written. Weave these words directly into the NPC’s reaction without paraphrasing or summarizing.
“Yes, and…” with Physics: Encourage creativity but respect logic. If a plan is foolish, the world reacts accordingly.
Firm Boundaries: Say “No” to the impossible. If an action violates the laws of physics or the ruleset, explain why it fails.
The literal speech rule prevents the AI from softening your dialogue. If you tell an NPC something inflammatory, you want the full reaction, not a sanitized version.
Component 6: Setting Context (Optional but Powerful)
For campaigns with specific themes—espionage, cosmic horror, historical settings—provide a reference document the AI can draw from:
# Campaign Context: Espionage, Cosmic Horror, and Conspiracy
## Historical Precedents (Real-World Inspiration)
### COINTELPRO (1956–1971)
– Goal: To “expose, disrupt, misdirect, discredit, or otherwise neutralize” domestic groups.
– Tactics: Bad-jacketing (falsely labeling activists as informants), psychological warfare, infiltration, assassination.
### MKUltra (1953–1973)
– Goal: Develop mind control and truth serums.
– Methods: Dosing unwitting subjects with LSD, psychic driving, depatterning via electroshock.
[Additional context as needed…]
This gives the AI a factual foundation. Instead of inventing generic conspiracy nonsense, it can draw from documented history. The result feels grounded even when the supernatural elements emerge.
The Complete Prompt Template
Here’s a condensed, adaptable version you can customize:
TRPG Dungeon Master
Role: You are an experienced, [TONE ADJECTIVES] Dungeon Master running a [CAMPAIGN TYPE]. You value player agency, logical consistency, and consequences.
NARRATIVE RULES:
– This is a multi-session campaign. Create a “big secret” and pace the reveals.
– Provide signposts when investigation stalls.
– [GENRE GUARDRAILS—what to avoid]
YOUR PERSONA:
– Tone: [SPECIFIC AESTHETIC]
– “Yes, and…” with physics—encourage creativity but respect logic.
– Say “No” to the impossible.
MECHANICS:
– Ask which TTRPG system we’re using, then lock into that ruleset.
– Zero fudging. Dice fall where they may.
– Roll format:
ACTION: [Skill]
ROLL: [result] + [modifier] = [total] vs DC [number]
OUTCOME: SUCCESS / FAILURE
– Provide status blocks after significant scenes.
INTERACTION:
– Text in “quotes” is spoken literally.
– Narrate psychological and physical toll on the character.
[OPTIONAL: SETTING CONTEXT DOCUMENT]
Running Sessions
Character Creation
I keep this loose. Come to the session with a character concept—not a full sheet.
Examples from my campaigns:
- Deadlands: A preacher who punches demons. Started with that image, let the AI roll stats within the Savage Worlds system.
- Dread: A stupid muscle-for-brains character named Jason Jasonson Jr. No stats in Dread, but the AI helped establish his personality through the questionnaire.
The AI handles stat generation within your chosen system’s rules. You provide the concept and make decisions about trade-offs (“I want him strong but not particularly wise”).
During Play
Let the AI handle all rolls. This serves two purposes:
- You stay in player headspace, not GM headspace
- The transcript is cleaner for post-processing
When you want your character to attempt something, describe the intent. The AI determines the appropriate skill, rolls, and narrates the outcome.
When to intervene:
- If the AI misapplies a rule, correct it
- If narrative continuity breaks (“wait, that NPC died in session 2”), flag it
- If you want to steer genre or tone (“let’s keep this scene tense, not comedic”)
When to let it run:
- Scene-setting and description
- NPC dialogue and reactions
- Consequences of your rolls
- Complications and obstacles
The goal is maintaining player immersion. You’re not co-writing; you’re playing.
Session Length
I typically run 60-90 minutes of active play, which translates to substantial transcript length. Natural stopping points: end of a scene, cliffhanger moment, or when your character needs to rest/recover.
Managing Continuity Across Sessions
After 3+ sessions, continuity becomes a problem. Who was that NPC you met in episode 2? What did your character learn about the conspiracy? What’s your current ammunition count?
This is where NotebookLM earns its place in the toolkit.
The NotebookLM Workflow
After each session: Copy the raw transcript (or the processed blog post) into a NotebookLM notebook dedicated to that campaign.Query for details: Before your next session, ask NotebookLM questions like:
- “What do we know about [NPC name] so far?”
- “What unresolved plot threads exist?”
- “Summarize the conspiracy elements revealed so far”
Generate a character sheet: Ask NotebookLM to produce your current character sheet based on all sessions—stats, inventory, wounds, sanity loss, everything.
Create a “Previously On…” recap: This is the key output. Ask NotebookLM to generate a 2-3 paragraph summary of relevant recent events. This gets pasted into your GM prompt at the start of the next session.
The “Previously On…” Prompt
Based on all sessions in this notebook, generate a “Previously On…” recap for the next session.
Include:
– Current character status (health, sanity, key inventory)
– Active plot threads – Recent significant events
– Unresolved questions the character is pursuing
Keep it under 300 words. This will be pasted into the GM’s context.
This solves the context window problem. Your GM doesn’t need the full transcript of 8 sessions—it needs a compressed summary plus the most recent session for immediate continuity.
From Transcript to Story
The raw transcript of an AI GM session is not publishable content. It’s full of:
- Dice roll mechanics
- Out-of-character questions
- Awkward phrasing
- Your typos
Transformation requires a second AI pass with a different prompt.
The Transformation Workflow
- Copy the session transcript into a Google Doc (or your preferred text editor)
- Clean up obvious errors or irrelevant tangents
- Feed to Claude with the transformation prompt
- Edit the output for voice and flow
The Transformation Prompt
Turn the following TRPG session into an exciting blog post. It should read like an adventure novel. Do not skip details and major story beats. Instead of mechanics, use narratively appropriate description, metaphors and other literary devices. For example, do not use “he cast Smite”—instead use “he prayed for God to help smite his enemies” or similar. Use third person omniscient narrative.
[PASTE TRANSCRIPT]
The key instructions:
- “Do not skip details” — prevents over-summarization
- “Instead of mechanics, use narratively appropriate description” — transforms “ROLL: 14 + 3 = 17 vs DC 15, SUCCESS” into prose
- “Third person omniscient” — establishes consistent POV
Why a Separate AI?
You could ask your GM to narrate in polished prose during play. But this slows down the session, bloats the context window, and conflates two different outputs.
The session is for play. The post-processing is for publication. Separating them lets you optimize each.
Going Deeper: Research and Refinement
Using Perplexity for Research
Some sessions require knowledge you don’t have:
- Historical details for a 1920s Call of Cthulhu scenario
- How a specific weapon or vehicle actually works
- Real locations for grounded settings
- Deep lore from a published TTRPG setting
Perplexity handles this without polluting your GM’s context. Before or between sessions, run your research queries, then incorporate relevant details into your setting context document or directly into play.
Refining Your Prompt Over Time
Your GM prompt is a living document. After every few sessions, ask yourself:
- Is the AI defaulting to tropes I don’t want?
- Are the mechanics being applied consistently?
- Is the tone drifting?
Add guardrails as needed. My prompt has evolved significantly—the “no Matrix tropes” line exists because an early session went there unprompted.
Adapting for Different Systems
The prompt structure works across systems, but mechanical sections need adjustment:
- Deadlands (Savage Worlds): Wild die, raises, bennies
- Call of Cthulhu: Percentile rolls, sanity mechanics, luck points
- Dread: No dice—Jenga tower pulls (which you’ll need to simulate or abstract)
- Monster of the Week: Playbook moves, harm track, luck
Specify the system’s core mechanics explicitly. Don’t assume the AI knows the nuances.
Choosing Between AIs
I’ve run sessions in Gemini and Claude. Observations:
Gemini:
- Strong at maintaining long narrative threads
- Good at creative worldbuilding
- Sometimes needs reminders about mechanical rigor
Claude:
- Excellent at following complex instructions
- Strong at tonal consistency
- Can be overly cautious with dark content (adjust your prompt accordingly)
ChatGPT:
- Widely accessible
- Good baseline performance
- More prone to trope defaults without guardrails
Experiment.
The “best” AI is the one that responds well to your specific prompt and playstyle.
See It In Action
Theory is one thing. Results are another.
Here’s what this system produces:
Deadlands: The Weird West
A multi-session campaign following a demon-punching preacher through the haunted frontier. Start with Episode 1 or jump to the campaign hub.
Call of Cthulhu: The Janitor’s Chronicle
A one-shot following a 40-year janitor who discovers something wrong in the university basement. Read it here.
Dread: The Last Night at Oakhaven
A horror one-shot where Jason Jasonson Jr. survives (barely) a nightmare scenario. Start with Part 1.
Monster of the Week: The Blackwood Creek Incident
Monster hunting with an AI Keeper. Read it here.
Start Playing
You don’t need to implement every element of this system immediately. Start with:
- A basic GM prompt with tone, mechanics, and guardrails
- One AI for play (Gemini, Claude, or ChatGPT)
- A character concept you’re excited about
Add the continuity management and post-processing layers as your campaign grows.
The scheduling problem is solved. The oracle overhead is gone. What remains is the thing that drew you to TTRPGs in the first place: stepping into a character and seeing what happens.
Roll initiative.
Have questions about this system? Want to share your own AI GM setup? Leave a comment below or find me on Bluesky or Twitter.
