The Turing Werewolf

The village

Inspiration

I wanted to see if the built-in AI could make social deduction games playable as a solo player. Werewolf (also known as Mafia) is a perfect example for that, since it is all about deception, deduction, and language. I wanted a human-feeling Werewolf game where the user plays against a room of Chrome AI villagers and werewolves, all running locally in their browser. I also wanted to test if it is possible to create a set of AI that can interact and think humanly under certain constraints.

What it does

The Turing Werewolf is a client-side web application game where you play as a villager, werewolf or fool. Every other player are AI agents using Chrome's built-in AI that converse, accuse, and vote like humans.

Core features:

Turn-based conversation: Chrome AI NPCs take turns to speak and reason about suspects.
Style Normaliser: your messages can be rewritten into neutral English before being sent, so the AI cannot “meta-read” your grammar or style.
Structured reasoning: each agent outputs both public speech and hidden logic (in JSON).
Voting system: after discussion, everyone secretly votes to lynch a suspect.

Each NPC runs as its cloned Prompt API session with their own prompt:

Agent includes the game state and asks for an action (speak/vote/kill).
NPC responds with structured JSON.

How we built it

Tech stack

Frontend: JavaScript / React / TailwindCSS
AI layer: Chrome’s Prompt API for per-agent reasoning sessions and Rewriter API to provide player with the option neutralise or refine their messages before sending so as to prevent AI agents from detecting personal writing patterns.
Local state: XState with persistence
Schema Validator: JSON schemas enforced in-browser to keep AI responses consistent

API used

Prompt API: To build the core gameplay aka the NPCs talking
Rewriter API: To allow the player to use neutral English speech

Challenges we ran into

The LLMs often spoke unnaturally and failed to maintain conversational flow, even with extensive context. Implementing structured turn-taking was necessary to improve realism.
Some prompts were ignored, even when explicitly told not to perform certain actions, the models still did.

Accomplishments that we're proud of

A fully client-side multiplayer: Each game feels different and you can influence the AI based on the messages you send
While it is turn based speaking for the AI, it still feels natural as if the conversation is going by.

What we learned

LLM is fragile unless you give them structure + constraints.
Designing multiplayer interactions between multiple Prompt API: I learnt about cloned sessions to provide different memory and persona.