llm-game-master technique sourced
LLM game master and narrative director
A language model that adjudicates player and NPC intentions, decides outcomes, generates events and quests, and narrates them, either free-form or checked against a typed world state and canon.
mechanism
The player types an intent; the system assembles a prompt from recent story, pinned memories, and setting notes and asks the model to continue the story, as AI Dungeon has since 2019. Grounded variants split the job: actors state intended actions in language, a Game Master component checks plausibility against rules and state, decides what happens, writes state changes, and only then narrates (Concordia). Hidden Door runs intent through plan, canon check, and world-state update steps over a typed, auditable world model, using many small tuned models for about 100 narrative tasks; its example blocks Dracula from entering a house uninvited. Research systems generate setting, items, and NPCs from designer criteria and use the model to validate free-text input against the story (PANGeA), or assist a human Dungeon Master (CALYPSO).
limitations
- Long-session consistency: models forget or contradict earlier facts unless state is externalized into memory or a typed world model.
- Accommodation bias: models rarely say no, which removes challenge and lets players talk past any obstacle.
- Content moderation is hard: AI Dungeon's filter both missed and over-blocked, and Steam requires guardrail disclosure for live-generated content.
- Latency of several to 15 seconds per turn frustrated players in a GPT-4 study; multi-step grounded pipelines add calls per turn.
- Every turn is an inference cost for the life of the service, which pushes products to subscriptions or small tuned models.
relations
7 stored edgesGenerative content on demand
Story continuation from Do, Say, Story, and See inputs makes the whole narrative generated content; grounded engines turn it into typed state changes.
Natural-language interaction
Any typed action is adjudicated for plausibility and consequence rather than matched against a parser's verb list, so unanticipated approaches produce new narrative paths (53 emergent nodes in one 28-player study).
Procedural characters
NPCs with backstories, goals, and personality are generated from designer criteria and voiced in dialogue, removing per-character writing.
Storyteller director
The director can author an event in language for the exact current state, check it against canon and rules, and narrate it, instead of choosing among a finite catalog of pre-written incidents by a pacing budget.
Triggered event chains
Follow-up beats are generated from story state and player intent rather than scheduled by authored templates with trigger predicates, while the player still experiences multi-step chains.
Villager requests
Characters can ask for things tied to the specific story so far, not only fetch templates with filled slots; ungrounded models invent quests that do not exist, so requests must be checked against state.
backlinks
computed from other recordsclaims
10 with sourcesAI Dungeon 2 (December 2019) ran the full 1.5B-parameter GPT-2 and moved to GPT-3 in mid-2020, reaching over 100,000 players in its first week and 1.5 million by June 2020.
AI Dungeon's 2021 content filter, added to block sexual content involving minors, produced false positives on innocuous text and led human moderators to read flagged private stories, prompting review bombing.
In a 28-player study with five GPT-4 NPCs holding backstories, clues, and ordered goals, play produced 53 narrative nodes absent from the designers' graph; about 15 seconds of lag frustrated 40% of players and NPCs forgot earlier statements in 10% of reports.
Hidden Door announced early access in August 2025 and offers licensed worlds such as Dracula; its engine checks every story turn against canon and rules before display and reports a tuned 1.7B model beating a 120B general model on one core task.
src-hiddendoor-2026-blog-index, src-hiddendoor-2026-homepage
Concordia's Game Master converts agents' natural-language actions into outcomes and checks physical plausibility, and its 2025 design pattern classifies uses as Simulationist, Dramatist, or Evaluationist.
src-vezhnevets-2023-concordia, src-vezhnevets-2025-multi-actor-game-engine, src-deepmind-2026-concordia-repo
Surveyed LLM game masters tend to accommodate even bizarre player requests that human GMs would curb, and LLM NPCs can invent quests that do not exist in the game.
Steam requires games with live-generated AI content to disclose it and describe guardrails against illegal output, and does not permit live-generated adult sexual content.
src-valve-2026-steamworks-content-survey, src-valve-2024-ai-content-on-steam
Dungeon Masters using CALYPSO presented some generated text almost directly and treated rougher output as ideas while keeping creative control.
PANGeA generated settings, key items, and Big Five personality-biased NPCs from designer criteria and used the model to keep free-text player input within the narrative, tested in a browser demo and a Unity game.
Steam generative AI disclosures rose from about 1,000 games in April 2024 to 7,818 (about 7% of the catalog, a little under 20% of 2025 releases) in July 2025 and 10,258 (about 8%) in November 2025; about 60% of disclosures concern visual assets, so live-generated runtime systems such as LLM narration are a small share.
src-totallyhuman-2024-steam-genai-count, src-totallyhuman-2025-steam-genai-count, src-totallyhuman-2025-steam-ai-revenue