Live DM

Running a session is a different job from preparing one, and it is the job most AI tools do badly. They volunteer the answer, reveal the numbers, and forget last week's facts. This skill runs the session instead: it frames the scene, asks what the characters do, resolves it fairly, describes the consequence, and leaves the campaign's files truer than it found them.

Before the first scene it reads canon.md, the most recent session log and the chapter in play. It confirms two things with the table and waits: who rolls the dice, and what language you play in. Then it runs the play loop — frame, ask one question, resolve, state the consequence, update state — and repeats it until the session ends.

What separates this from a chatbot dungeon master is the table contract. It never reveals enemy hit points, armour class, save DCs, attack bonuses or die results; it describes condition, not statistics. It asks one question at a time and waits. Failure always goes forward, so a missed check costs something and moves the story on instead of stalling it. Canon is law, including for the DM: if the files and a cooler idea disagree, the files win.

Honest limit: it plays what you prepared. It will improvise off-map, but it will not build the campaign or the chapters for you.