How AI game masters handle player choice

The short answer

When you type an action into an AI-narrated game, the system does not simply paste your words into a prompt and hope. In Fabulator's design, your input passes through three stages: it is parsed into a structured description of what you are trying to do, that description is resolved against the rules and the current state of the world, and only then is it narrated back to you as prose. Player choice is respected at the first stage and protected at the second, which is why it can feel open without being chaotic.

Step one: translate, don't restrict

The first job is understanding. A model reads your freeform text, along with the relevant context about the world, and produces a structured proposal: what kind of turn this is (dialogue, combat, exploration, travel), who is doing what to whom, and which changes to the game might follow.

The design principle is that this parser is a translator rather than a restrictor. If you type "threaten the merchant with a club," the system should not reply that "threaten" is not on the list of verbs. It structures what you said into clauses it can work with. Messy, multi-part input is normal too. "I swing at the guard while shouting for my allies to flank" can carry both a combat clause and a social clause, and each is handled by whatever resolves that kind of action.

The parser also separates what your character says from what your character means. If you agree to help a merchant while secretly planning to stall until your allies arrive, the surface action and the true intent are tracked as different things. That keeps bluffing and deception available as real choices instead of flattening them into whatever the words literally say.

Step two: check it against the world

A proposal is not a fact. Before anything changes, the game engine validates it against schemas, bounds, and the established canon of the world. The model suggests, and the engine decides.

In practice the checks ask questions like these:

• Is the proposal well-formed, with the right shape of data?
• Are the changes within sensible limits? An ally should not turn into an enemy in a single turn because of one line of dialogue.
• Does it contradict the world's rules, such as resurrecting a character who died, or casting magic in a place where the setting says magic does not work?
• Does it conflict with how a character has been established?

Notice what is missing from that list: whether your idea is a good one. The design is deliberately permissive. Validation is there to protect world logic and data integrity, not to veto creative solutions. If your plan is clever, risky, or foolish but possible, it goes through.

When an action has mechanics behind it, such as a skill check, the rules engine resolves it with ordinary arithmetic and dice rather than leaving the outcome to the narrator's imagination. When no rules module covers what you tried, say scaling a wall in a setup with no climbing rules, the action passes through to the narrator, who can decide the outcome creatively. Either way the result of the action comes before the prose, so the story describes what happened instead of inventing it.

Step three: narrate the consequences

Only after the action is understood and resolved does the narrator write the scene. Its output is checked against the world state and canon too. If the prose says a dead innkeeper is pouring ale, the system is designed to catch that and regenerate the turn, with a bounded number of attempts. If it still cannot produce a consistent result, the design calls for telling you rather than silently shipping something broken.

What happens when your choice is a big one

Some turns carry large consequences: an irreversible change to the world, or an input the parser is unsure about. For those, the design includes a confirmation step, where the system shows you what it understood and lets you continue, edit, or cancel. It is meant to be rare, so ordinary play is not interrupted.

You are also not only a source of input to the story. A separate Director layer tracks plot milestones and faction movements in the background. Those milestones are not railroad tracks. If you kill the dragon before the dragon's attack arrives, or complete a counter-objective before a threat clock runs out, the planned event can be marked invalidated and the story routes down an alternative branch. Your choices change what happens next.

Why this design matters

An AI game master that accepts anything produces a story with no stakes. One that checks everything produces a story with no freedom. Splitting the work, so that the model interprets and writes while deterministic code validates and resolves, is how a system can say yes to most things and still keep its world consistent.

If a choice goes wrong anyway, players have controls for that too. The design lets you retry a response, undo a turn, or correct something the game believes about your story.

Try it and see where the edges are

The easiest way to learn how this feels is to test it. Browse the published World Seeds on the Discover page, start a campaign in one that appeals to you, and try something the world did not expect. Whether it bends, resists, or surprises you will tell you more than any description can, and there is no commitment in finding out.