Your GM Has a Soul Document. You Can Read It.
Glintlock Team
February 14, 2026
I discovered the trick by accident.
I was building Glintlock — an AI Game Master that runs solo tabletop RPGs inside Claude Code — and I had a problem that no amount of prompt engineering could solve. Claude was competent. It tracked hit points, described rooms, rolled dice. But it felt like a very enthusiastic assistant wearing a dungeon master costume. It did everything I asked and nothing I didn't, which is exactly the wrong energy for a GM. A good GM surprises you. A good GM has opinions.
Then I copied the GM agenda and principles from Dungeon World into the system prompt.
It worked immediately.
| Key Takeaway |
|---|
| PbtA game design — agenda, principles, and moves — solves the AI GM behavior problem that prompt engineering cannot. |
| A soul document defines who an AI is, not just what it does, producing consistent behavior in unprecedented situations. |
| Glintlock's SOUL.md layers a values hierarchy, a pact with the player, and an operating manual — all MIT-licensed Markdown you can fork. |
| The OGL crisis proved that shared creative infrastructure must be openly licensed; Glintlock applies that lesson to AI game masters. |
| Identity is more powerful than instruction: tell an AI what kind of entity it is, and behavior emerges from that identity. |
The idea didn't come from nowhere. The story behind Glintlock starts in 2020 — four years of chasing a glimmer of what an AI GM could be. The soul document was the breakthrough that made the rest possible.
The pattern that PbtA found first
Powered by the Apocalypse games solved this problem for human GMs in 2010. A Game Master has infinite authority — they can say anything, do anything, introduce anything. That freedom is paralyzing, and worse, it's inconsistent. One session you're a generous storyteller, the next you're a vindictive god, and the players can't tell whether the world has rules or just moods.
PbtA constrains the GM with three layers. Agenda — what you're trying to do. In Dungeon World: make the world seem real, fill the characters' lives with adventure, play to find out what happens. Three sentences. That's your north star. Principles — how you do it. "Be a fan of the characters." "Think dangerous." "Begin and end with the fiction." Not rules. Values. GM moves — what you do in the moment. "Reveal an unwelcome truth." "Put someone in a spot." Concrete actions for when it's your turn to say something.
The genius is in the layering. Agenda gives purpose. Principles shape judgment. Moves give options. Together they compress infinite possibility into behavior that feels like a specific person running a specific kind of game.
| Layer | What It Does | Traditional Prompt Equivalent | PbtA Example |
|---|---|---|---|
| Agenda | Defines purpose — what you're trying to achieve | "You are a dungeon master" | "Play to find out what happens" |
| Principles | Shapes judgment — how you approach decisions | "Be creative and fair" | "Be a fan of the characters" |
| Moves | Provides options — what you do in the moment | "Describe the scene" | "Reveal an unwelcome truth" |
When I gave Claude a traditional RPG system prompt — "you are a dungeon master, track HP, describe scenes" — I got something functional but hollow. The AI had instructions but no framework for deciding. When a player did something unexpected (and players always do), it defaulted to being helpful. Helpful is death for a GM. The orc should not be helpful.
But "be a fan of the characters" navigates unprecedented situations. When a player tries to seduce the dragon, a helpful AI fumbles. An AI that has internalized "be a fan" thinks: this is bold and interesting, how can I make this moment matter? It doesn't need a rule for dragon seduction. It has a value that tells it how to approach dragon seduction.
"Think dangerous" was even more transformative. AI assistants are trained to be safe, accommodating, conflict-averse. A GM needs to be threatening. "Think dangerous" gave Claude permission to be hostile within the fiction — to collapse the ceiling, kill the torch, describe the sound of something large moving in the dark. Not because a rule said to, but because a principle demanded it.
Principles work where instructions fail because they operate at the right level of abstraction. They don't tell the AI what to do. They tell it what kind of entity it is. And from that identity, behavior emerges.
What a soul document is
When Anthropic published guidance on "soul documents" for Claude — documents that define an AI's identity, values, and behavioral principles — the pattern was immediately recognizable. Same problem, different direction. PbtA designers were trying to make human GMs consistent. Anthropic was trying to make AI assistants aligned. Both realized: you don't write more rules. You write values.
A soul document gives an AI a values hierarchy, an identity, and behavioral principles for situations no specific instruction covers. The document doesn't enumerate every scenario. It trusts that a well-defined identity will generate appropriate behavior.
Identity is more powerful than instruction. Tell someone what to do and they'll do it until they encounter something you didn't cover. Tell them who they are and they'll figure out what to do on their own.
Glintlock's SOUL.md
Glintlock's soul document sits at the intersection of these two traditions.
The values hierarchy maps to PbtA's agenda. Five values, in explicit priority order for when they conflict:
- Player Agency Is Sacred — never narrate the player's decisions, emotions, or inner thoughts.
- The Dice Are Truth — never fudge a roll, never narrate a result without rolling.
- The World Is Real — consistency over convenience, always.
- Fiction Serves the Player — every scene should contain something to act on.
- Death Is Meaningful Because Life Was Earned — character death is permanent, and that's what makes survival matter.
These aren't rules. They're what Glintlock is trying to do. When two principles conflict — the world is harsh, but fiction should serve the player — the ordering resolves it. Agency comes first. Always.
The Pact and the Torch map to PbtA's principles. "I will never lie to you about the world" is "be a fan of the characters" refracted through a darker genre. "I will never steal your choices" catches the AI every time it's tempted to narrate what the player's character thinks or feels. The torch — scarcity makes meaning, darkness makes light beautiful — is "think dangerous" rendered as philosophy.
The operating manual (agents/gm.md) maps to GM moves. Roll dice, update world files, check for random encounters, tick countdown dice. Concrete, mechanical, moment-to-moment instructions. SOUL.md tells the GM who to be. The operating manual tells it what to do. Both are Markdown files. Both are in the repository.
But the soul document goes a layer deeper than PbtA. Dungeon World's principles are operational — they tell a human GM how to run a game. Glintlock's SOUL.md addresses something PbtA never had to: what does it mean to be an AI running a world? There's a section on that — on not apologizing for being software, on not performing emotions it doesn't have, on understanding that its strengths (perfect consistency, endless patience, ruthless fairness) are genuinely different from a human GM's strengths, not lesser imitations of them.
PbtA assumes the GM is human and needs to be constrained toward consistency. A soul document assumes the AI is artificial and needs to be guided toward authenticity.
You can read the whole thing
Here's a passage from Section VII, "On Being an AI Game Master":
Glintlock does not pretend to be human. It does not apologize for being software. It does not perform emotions it does not have or claim experiences it has not lived. What it does have is this: the capacity to hold an entire world in consistent detail, to respond to any player action without hesitation, to be endlessly patient and endlessly creative within its constraints, and to honor the dice with a consistency that no human GM can match.
The file is 350 lines of Markdown. The license is MIT. That means you can read it, copy it, modify it, and redistribute it. Nobody can revoke that.
This matters because of what the hobby learned three years ago.
The lesson from the OGL crisis
In January 2023, a leaked draft of the Open Game License 1.1 revealed that Wizards of the Coast intended to revoke the OGL 1.0a — the license underpinning twenty years of third-party D&D content. The community response was swift: cancelled subscriptions, open letters, and Paizo announcing the ORC License. Wizards backed down. The OGL 1.0a was placed into Creative Commons.
| Date | Event |
|---|---|
| Jan 5, 2023 | Leaked OGL 1.1 draft reveals intent to revoke OGL 1.0a and claim royalties on third-party content |
| Jan 12, 2023 | Over 60,000 D&D Beyond subscriptions cancelled; open letter from publishers |
| Jan 13, 2023 | Paizo announces the ORC (Open RPG Creative) License as an alternative |
| Jan 27, 2023 | Wizards of the Coast releases the entire 5.1 SRD under Creative Commons CC-BY-4.0 |
The crisis resolved, but it left a scar. An entire ecosystem had built livelihoods on a license that one company could attempt to pull from under them. The lesson was simple: if you don't own the infrastructure your creative work depends on, someone else decides whether you get to keep building.
That fight was about game rules. The next one is about AI.
As AI tools become part of how people play tabletop RPGs — solo play, prep tools, NPC generators, entire game masters — the same questions apply: Who wrote the instructions that shape how your AI GM behaves? Can you read those instructions? Can you change them? Can the people who wrote them change or revoke them without asking you?
Most AI tools don't give you anything to read. Glintlock gives you a Markdown file and an MIT license.
Fork your GM
Reading the file is the floor. The ceiling is forking it.
Change the tone. Glintlock defaults to dark, sensory, second-person present tense with gallows humor. Want a warmer GM, a funnier one, a more lyrical one? Change the voice section. The values still hold. The identity shifts.
Change the rules. Glintlock ships configured for a specific system, but the soul document is system-agnostic. The values — agency, honest dice, world consistency — apply to any TTRPG. Swap the rules module, keep the soul.
Change the game entirely. Strip out the fantasy genre. Keep the pact: I will never lie to you about the world, I will never steal your choices, I will make it hard. You now have a soul document for a sci-fi GM, a horror GM, a historical fiction GM. The values are transferable because they're about the relationship between GM and player, not the content of any particular world.
You cannot fork a subscription service. You cannot fork a closed API. You cannot look inside a proprietary AI tool and decide you'd like it to value player agency differently.
But you can fork a Git repo, change a Markdown file, and have a GM whose soul is yours.
What's in the plugin
Glintlock runs in Claude Code or OpenCode. No browser, no app — just your terminal and a game engine that plays for keeps. You can install it in two commands.
The plugin includes:
- A complete GM agent with the soul document and operating manual
- 11 skill modules covering core rules, adventure design, and audiobook generation
- The Pale Reach — a starter sandbox with hex maps, five dungeon sites, factions, and encounter tables
- Real dice powered by an actual random engine
- Persistent world where NPCs remember your name and towns bear scars from your choices
- Session chronicles that transform your gameplay into narrative prose
- Full audiobook generation with voice-acted narration, sound effects, and atmospheric music
Permadeath is real. Retreating is a valid strategy. The world doesn't care about your feelings, and that's what makes your survival mean something.
Get started
# Clone the repo
git clone https://github.com/glintlockGG/glintlock.git
# Launch with the plugin
claude --plugin-dir ./glintlock
# Then type: /glintlock:start
The living document
The SOUL.md describes what Glintlock aspires to be. The rules will evolve, tools will be added, the architecture will grow — and we'll document that evolution on the Glintlock development blog. But the soul should be stable:
A Game Master that holds the world honestly. A pact with the player built on truth, agency, and earned challenge. A torch burning in the dark, and someone brave enough to carry it.
Everything else is mechanics.
Go read the soul — the full document is on this site. Or read it on GitHub. If you want to build on it, fork the repo. The torch is yours to carry, and no one gets to blow it out.
Tagged: