How it’s made

A few short notes on the parts of Tavern Draft that use AI, and the parts that deliberately don’t.

What Jev does

Jev, from TypeSafe, answers typed questions about text: a yes or no, a pick from a few options, or a score on a few levels, each with a calibrated probability for how likely every answer is. Here it reads every hero’s bio and scores three traits, courage, kindness and focus, on three levels each. It also checks each name before it goes on the board, and checks that a saga doesn’t claim anything that didn’t happen in the run.

What Claude does

Claude writes the words: the bios at the bar, your party’s name once it’s full, and the saga when your party falls. It writes bios ahead of time in batches into a pool shared by everyone who plays, so the bar never waits; when the fresh ones run out, strangers you may have met before come back for another drink.

What plain code does

Every move in every fight is ordinary code, with no AI in the loop. The traits only nudge the choices. A timid hero may freeze when badly hurt, while a brave one hits harder when hurt. A kind knight guards the friend beside them sooner, and a kind cleric heals sooner. A sharp-eyed hero picks off the weakest foe.

Why it’s split this way

Jev is good at judgments (does this bio sound brave?). Code is better at decisions (who attacks whom, and for how much), because it is the same every time and can be tested. Nothing in the game waits for an AI. Claude’s work is prepared before you need it, and Jev’s checks have a short time limit.

If an AI isn’t there

The game is complete without either one. Bios, party names and sagas come from built-in templates, traits follow what a template bio says (the rest are picked at random), and names are checked against a word list. If a saga looks wrong to Jev, you get the plain template one instead.

A real example

So in a fight Wren may freeze up when badly hurt, and goes for the weakest foe.