AI persona playtesting
You know how you would play your scene. The interesting question is how a rushed player, a hostile one, or one who picks at random would — and whether the scene holds up. The simulator can hand its choices to an AI persona and let you watch.
Starting a run
Section titled “Starting a run”Open the simulator (Play ▾ → Simulate) and click ✦ AI plays your scenario as… at the foot of the panel. Pick a preset persona —
- A model learner
- Confused and asking for help
- Hostile — testing your patience
- A rushed nurse who skips instructions
- Chooses randomly
— or type your own in a sentence: “a courier who is hiding the letter and lies when cornered”. Then ▶ Play this run.
What you see
Section titled “What you see”The transcript fills in as it would for a human, but at each menu the AI picks. Above the choices it says ✦ AI chose “Bribe him” — the courier has coin and no time to argue, so you get the reasoning, not just the pick. Lines, variables and the highlighted path on the canvas all update exactly as they do for a human run.
■ Stop AI play hands control back to you at any point — step back, choose differently, and let the persona carry on from there. Everything else in the simulator (Variables, the Lua console, step history) works during an AI run.
What it is good for
Section titled “What it is good for”- Finding the path nobody takes. Run “chooses randomly” a few times and watch which branches never light up on the canvas.
- Tone checks. A hostile persona finds the places where your polite choices leave the player with nothing to say.
- Dead ends. A persona that skips instructions discovers the loop that never lets the player out.
- Demoing. A run with a named persona is a quick way to show a producer how the scene plays without them driving.
What it is not
Section titled “What it is not”It is not a player. A persona picks from the choices you wrote, in character, with the transcript as context; it does not measure fun, pacing or clarity the way a human playtest does. Treat it as a way to sample the branch space quickly — then send the player link to real people.
Each AI decision is one request to the workspace’s AI — from the included allowance or your own provider key, the same as the copilot. A long scenario on chooses randomly is the cheapest run: it still asks the model, but with the shortest possible prompt.