When Claude 5.5 Became the Dungeon Master: An 80,000-Character AI Tabletop RPG Experiment

From Tabletop RPGs to AI Dungeons: A Veteran Player's Obsession

When I was young, I fell completely in love with role-playing games—but I mean classic tabletop RPGs: a group of friends around a table, with a Dungeon Master (DM) controlling everything. No computer RPG has ever truly replicated that experience. No matter how beautiful the world or how thrilling the plot, something essential was always missing—the freedom of truly becoming a character, fighting side by side with friends, and knowing that almost anything was possible.

I've been not only a player but also a DM, designing my own adventures and building worlds. These skills have come in handy many times throughout my life, and still do today.

I've played plenty of computer games too, naturally mostly RPGs. One of my favorites among the early ones is Lands of Lore. To this day I still savor its simplicity—that special quality that usually only appears in masterpieces. I remember spending countless hours searching nearly pixelated walls for hidden switches, or figuring out that the only way through the swamp was to freeze it. Looking back, the entire game was built around a true KISS philosophy (Keep It Simple, Stupid).

A Retro Dungeon Builder Born on a Phone in Bed

It all started with a casual chat. We even talked about Rogmor—my own RPG project, started long before the AI era, which I return to and reshape from time to time. Claude outlined how it would build a small Lands of Lore-style dungeon crawler. This piqued my curiosity, so we went ahead and built the first playable version.

The funny part: all of this happened while I was lying in bed, continuing the conversation on my phone. The game ran on that phone immediately! This is exactly what I called "Scary Creativity" in Singularity News :: 0001.

But the truly interesting idea went beyond the game itself: a first-person Dungeon Master mode, where the DM can directly "dig" out an adventure from a first-person perspective, placing monsters, traps, doors, and NPCs along the way. Then I added a Famulus—a little assistant that helps dig, and even demonstrates how to dig.

An 80,000-Character Solo Adventure: How Claude 5.5 Performed

My original plan was to let the AI handle most of the world-building—not just generating mazes but also controlling NPCs. After discussing how to maintain a consistent game state, I felt it was time for a small test before continuing to vibe-code.

Then suddenly I found myself in the middle of a role-playing adventure. Almost like the classic tabletop RPGs I used to play—except this time there was only one Dungeon Master, Claude, and one player: me.

I should mention my earlier article LLM RPG Test 2026. As with every such setup, the usual limitations still apply. But I must say, Claude 5.5 performed exceptionally well. It managed to maintain tension throughout the entire adventure and delivered an ironic twist at the end that I genuinely didn't anticipate.

Personally, I especially enjoyed finally testing something I'd talked about for a long time. What has always bothered me in films and novels is that characters who are supposed to have faith rarely actually rely on the power of faith to solve problems. This time I finally had the chance to explore that possibility firsthand.

The section below—I doubt anyone will actually read it all—is over 80,000 characters, already longer than a typical short story.

The Bell of Greystone: How AI Crafted an Adventure with Theological Depth

Opening: A Masterful Hook

The story is set at dusk in the mountain village of Greystone. In the square, a great bell hangs in the church tower, but its rope has been cut and hangs slack and swaying. The sign of the nearby inn creaks in the wind. Across the square, a stocky, stern man in uniform watches you. Your palm tingles—as if from cold, but it isn't cold. It's a strange sensation you've never felt before on your travels.

This "palm tingling" setup is the core foreshadowing planted by the AI: the player character, Erwin Greenthorn, is a marked descendant of a "bellringer," and the mark on his palm is the key to unraveling the entire mystery. Within the first three sentences, the AI simultaneously establishes atmosphere, implants an anomaly, and hints at identity.

Weaving Multi-Layered Clues: Information Density on the Notice Board

The notice board is layered with papers—a mayoral proclamation under wax seal announcing the closure of the Deepwell mine, a captain's decree forbidding anyone from walking alone after sunset, a pencil-written notice seeking a missing child, a church announcement that the bell has been silent for three weeks. Beneath them all lies a yellowed ancient poster about the "Bell Festival," signed "Ilse the Elder, bellringer," with a seal beside it depicting a hand inside a bell.

When you see it, your palm tingles again, stronger than before.

The AI demonstrates textbook-level information layering here: on the surface, four independent notices correspond to four plot threads—the mining disaster, law and order, the missing person, and the church. The old poster underneath ties together the player's identity, the core mystery (the bell), and the final solution (the bellringer's mark). Every subsequent NPC conversation pays off one of these threads.

Hook: In tabletop role-playing and narrative design, a "hook" refers to the design element used in the opening to grab the player's/reader's attention and plant the seeds of future conflict. An effective hook typically accomplishes three things at once: establishing the scene's atmosphere, introducing an anomaly or suspense that cannot be ignored, and hinting at a special connection between the player character and that suspense. In this article, "palm tingling" is a classic composite hook—it's a sensory anomaly, an identity clue, and a preview of the ending's solution all at once. A good hook isn't just information delivery; it creates an internal drive in the player to think "I must figure out what's going on," so they actively advance the plot rather than passively accepting quests.

Character Development: Every NPC Has Their Own Agenda

Captain Hartley is weary, official, but harbors hidden doubts about the mayor; the blacksmith Barbara's grief over her lost son is wrapped in pragmatism, and she demands labor in exchange for information; Mayor Voss is polite, generous, and flawlessly tries to send the player away with ten gold coins; Mira, the girl sweeping the church, hesitates to speak but arranges to talk alone by the well.

Voss's characterization is particularly interesting. When the player lies about three friends arriving soon, Voss responds: "Three people, rarely come here at the same time, especially now that the mountain pass is closing." His fingers tap on the table. This "calculation beneath politeness" is a detail the AI generated autonomously, rather than a simply labeled villain.

The Real Surprise: Faith as a Gameplay Mechanic

The player character declares himself an "anointed priest-paladin of the Holy See," and Claude 5.5 has the villain Corvin precisely pick apart the doubts about this identity—no retinue of thirty, no sealed letters, staying at the blacksmith's house, village mud on his boots—then allows the player to win back the upper hand through genuine theological argument.

The debate between the player and Corvin about "whether the One God created the being beneath the seal" is the philosophical core of the entire adventure. Corvin says: "The One God created everything, stone and grass, so He also created that thing. If it is God's creation, it is not a demon, just a created being."—an argument that is "elegant, I cannot refute it."

And the final solution isn't combat, isn't magic: it's the player placing his palm into the depression in the disc, persisting in prayer amid the whispers of the evil god, asking God for "nothing"—precisely the language that the listening-only entity has never heard before.

According to the original text, Claude 5.5 maintained tension throughout the entire adventure of over 80,000 characters and delivered an ending with a "genuinely unexpected ironic twist." The author particularly emphasizes that he was finally able to test in-game the narrative problem that had long troubled him: "characters of faith truly relying on faith to solve problems."

Theological Debate as a Gameplay Mechanic: In traditional CRPGs, persuasion or debate is usually reduced to skill checks (like the dialogue dice rolls in the Baldur's Gate series), where players pick options and the system determines success or failure. LLM-driven dialogue, however, allows players to freely construct arguments, with the AI evaluating the logical strength and worldview consistency of the reasoning. In this article, Corvin's "created being, not demon" argument is powerful because it derives its conclusion from within the villain's own theological premises, constituting an internal critique—the player cannot refute it by citing dogma, only by proposing an equally coherent alternative argument. This "argument versus argument" interaction turns faith from a number on a character sheet into a narrative resource that requires the player to genuinely think.

What This Experiment Demonstrates

The unique value of an LLM as a Dungeon Master doesn't lie in generating random dungeons or automating combat—traditional programs are more efficient for that. Its value lies in this: when a player delivers an improvised line or an unconventional action (like throwing a coin purse in the captain's face to buy his loyalty), the AI can improvise a response consistent with the world's logic while keeping all previously planted clues coherent.

The 80,000-character actual play log provided in the original article is essentially a stress-test report on LLM long-horizon state management. Claude 5.5 maintained throughout: the independent motivations and information gaps of multiple NPCs, a mark-related clue thread running through the entire adventure, a villain with theological debate depth, and a non-combat ending triggered by player choice.

Limitations still exist—the "usual limitations" mentioned by the author in LLM RPG Test 2026 did not disappear in this experiment either. But as he says: "Claude 5.5 performed exceptionally well."

For developers wanting to try similar experiments, the original article reveals a key starting point: a retro dungeon engine that runs on a phone in bed, plus an LLM willing to play DM. The rest is 80,000 characters of imagination.

Long-horizon State Management: Refers to the set of information an LLM continuously tracks and keeps consistent over long, multi-turn interactions—including character relationships, past events, item ownership, unresolved clues, and character stances. The longer the conversation, the more the context window is diluted by irrelevant content, and the more likely the model is to "forget" early setup or contradict itself. The 80,000-character adventure in this article is considered a stress test precisely because it contains multiple NPCs with independent motivations, a mark-related clue thread running throughout, and a twist ending that only works if earlier and later elements resonate. Maintaining the coherence of these elements requires the model to recall and abide by the previously established world state with every line of dialogue it generates.