Next to-dos

Prototype structured game state + tool-call updates (the core bet)
Pick the first ruleset — rules-light vs 5e SRD
Auto-generate "previously on…" session recaps
Decide improvisation vs authored-content balance
Solo vs multiplayer for v1 (lean solo)

Recent activity

Created project · 4 hours ago
To-do added — Prototype structured game state + tool-call updates (the core bet) · 4 hours ago
To-do added — Pick the first ruleset — rules-light vs 5e SRD · 4 hours ago
To-do added — Auto-generate "previously on…" session recaps · 4 hours ago
To-do added — Decide improvisation vs authored-content balance · 4 hours ago
To-do added — Solo vs multiplayer for v1 (lean solo) · 4 hours ago

Design doc

Dungeon Master AI — design doc

What it is: A patient, always-available game master for tabletop RPGs. It narrates scenes, plays NPCs, adjudicates dice, and — the part LLMs usually flub — keeps a consistent world and character state across many sessions.

The problem it solves

LLMs are great at improvising a scene and terrible at continuity: they forget the innkeeper's name, the quest you're on, and that you already killed that boss. The fix isn't a bigger context window — it's structured game state the model reads and writes every turn.

Approach

State, not vibes: characters, inventory, NPCs, locations, and quest flags live in a structured store. Each turn the model is given the relevant slice and must update it via tool calls.
Session recaps: auto-generated "previously on…" at the start of each session.
Dice + rules: real RNG and a rules adapter (start system-agnostic / rules-light).

Stack

LLM with tool calling · a game-state DB · a chat UI with a sidebar showing the live character sheet & quest log · pluggable rules modules.

Open questions

Which ruleset first — a rules-light system, or commit to 5e SRD?
How much should the AI improvise vs. stick to authored content?
Solo play vs. multiplayer table — different state/turn models. Pick one for v1 (lean solo).

Status

Idea stage. The whole bet is the structured-state-every-turn approach; if that doesn't hold continuity, it's just a worse chatbot wearing a wizard hat.