Play "Guess Who" with LLMs!

Advanced

Game Options

Uncover LLM's selected character

Flip the LLM's character card to reveal its identity. It's technically cheating, but it's useful to make sure the LLM is not lying to you 😉

System Prompt

Customize how the LLM behaves in the game. This controls the LLM's personality, knowledge, and gameplay style.

Load Simple Prompt Load Spelled Out Prompt Reset to Default

With this little game you can try to play "Guess Who" against an LLM of your choice.

The Rules

If you are not familiar with "Guess Who", here is a quick recap of the rules:

• Each player has a board full of characters.
• Each players draws an additional random character.
• Your goal is to guess which character the other player has received by asking yes/no questions, such as "Is your character male?" or "Does your character have black hair?" and so on
• The first player to guess the opponent character's name wins.

Choose the LLM

Before you can play you have to connect an LLM. This game supports OpenAI-compatible endpoints, Anthropic and Gemini models. Just click on Settings, select your provider, type in the model name of the LLM you want to play against and add your API key. You can test that your setup works by clicking on the Test button.

What's an API key?

An API key is similar to a password: it gives application access to a private resource. For example, have a look here to get your own OpenAI API key: Where do I find my OpenAI API Key? - OpenAI Help Center

No way I'm giving you my API keys dude

Legit! I wouldn't either. If you still want to play, you can get the source code on GitHub and run the game yourself. It's extremely easy! The entire game is an HTML file with a little JS and CSS on the side. No frameworks, no build systems, no Node.js, React or whatever. This game has no backend at all. It's fully hosted on GitHub Pages and everything is stored in your browser's local storage. Check the README of the repo for more details, but it all boils down to running these three commands on any modern Linux or Mac OS.

Can I play for free?

The game is free, but if you're using a hosted LLM you're likely going to be charged by the token. Check your provider's terms to learn how much the tokens cost in your specific case. Rest assured: in this game there's no hidden consumption of tokens. The LLM is used exclusively to chat with you, so in most cases it should cost very little per game (pennies or less).

Gameplay

You can communicate with the LLM through a free form chat, just like you would play with another person.

Your board will be on the left, with your chosen character at the top. You can remove characters from your board by clicking on them. Clicking on an eliminated character brings it back.

You can also see the LLM's board to know how many characters it already filtered out and how close it is to winning.

The LLM is in charge of declaring a winner. When you win, the LLM should declare you the winner:

When you lose, you'll see this instead:

For the pros

In the Settings window you can find some more tools. They are especially useful if you're trying to prove a point.

• Reveal character: reveals the LLM's selected character. With this setting turned on, winning is trivial. However, in many cases the LLM will lie about its character's features (!!!) so it can come useful if you're suspecting your LLM's partner of cheating.
• System Prompt: the system prompt is crucial to the behavior of the LLM. By modifying the system prompt you can alter completely the behavior of the LLM and try to make it play better/worse/differently. I prepared two system prompts that can be loaded with the press of a button:
• the "simple" prompt: what I believe would be sufficient for a smart LLM such as GPT-5 or Claude Opus to play this game normally.
• the "spelled-out" prompt: what I eventually had to use in order to make most LLMs get close to a gameplay that makes sense.

Why?

My original intent was to make LLMs play against each other and make a sort of leaderboard of their Guess Who prowess... but the results of my initial testing were so delirious that I needed to share this version too.
I am now collecting a few gameplays for a future post on my blog, which I will link here once ready (or you can subscribe to my Substack to receive it in your inbox). If you stumble upon some odd behavior, please share it with me! You can find me on Twitter/X, Bluesky, Mastodon, LinkedIn, and of course via email. Needless to say, contributions are welcome.
Enjoy!