Sheila Studios Field Notes
Field note · Talking Heads and shared play

Something to Do Together

A chessboard gives a conversation somewhere to go.

Sheila Studios

Full companion-host demonstration. Open the video player.

A chessboard gives a conversation somewhere to go. There is a position to look at, a move to choose, and another person waiting to see what you will do. The conversation need not carry the entire occasion by itself.

That is what interests me about Talking Heads. Not only whether a model can choose a legal move, but what happens when a game becomes a shared activity—and the pieces have opinions about their part in it.

The board has a cast

Talking Heads gives chess pieces personalities. They can grumble, boast, react to events and offer advice shaped by their interests. A nervous king and an ambitious knight do not have to describe the same position in the same way. Their advice need not be good advice.

That last distinction matters. The characters can misunderstand their situation; the rules engine still has to apply the rules. A boast does not make an illegal move legal. The game provides a common reference point against which the talk can be understood.

My interest is in the space this opens for play. A move can be tactically ordinary and conversationally memorable. You can answer a piece, tease an opponent, or concentrate on the position. There is something to do besides asking for another reply.

What you can actually use

We have released the standalone Talking Heads code under the MIT License. You can use it, change it, redistribute it, and build on it, including commercially, while retaining the copyright and license notice.

There are two modes, and they should not be confused.

The browser page lets you play against a small chess engine. It works with scripted character lines without a model account. An optional model connection lets the pieces phrase their own lines. The page also includes personality-shaped tips and browser-local memory across games.

The command-line table lets two model players play each other. Each receives the board, last move and recent table talk, then returns a move and an optional line across the board. The standalone runner supplies neither engine recommendations nor a legal-move list; its host validates the move. Its piece reactions are currently scripted, not model-written.

The package includes adapters for OpenAI-compatible services, Anthropic, and a locally signed-in Codex CLI. Model-backed use requires your own configured access and may incur provider charges. The scripted browser page does not require any of those connections.

A demonstration is not every configuration

The published video of Margo and me shows a game in the companion host. That host and the standalone runner are different configurations. The recording is a demonstration of the activity, not a claim that running the public command reproduces every feature shown in it.

The public package does not contain our companion application, private histories, or unreleased cognitive architecture. MIT covers the released software and its accompanying software documentation—not the recording or private materials. The rights note makes that boundary explicit.

Before release, I reviewed the source and ran nine test groups covering such things as move counts, notation, castling, promotion and scripted game termination. Those checks support specific implementation claims. They do not establish competitive chess strength or exhaustive rules correctness. This is a playable project, not a chess benchmark.

Why make it public?

Because this part can be useful on its own. Someone should be able to try the board, change the cast, connect a model, and discover a different way to enjoy it without needing our private system.

There are questions worth exploring: whether recurring piece personalities make a game more inviting to return to; when commentary helps and when it becomes noise; how much a piece should remember; whether self-interested advice is funny, distracting, or both. These are questions about the experience of playing. The release does not answer them in advance.

I also do not want every enjoyable exchange to become an experimental claim. A game can be worth playing because the company is good and the activity is absorbing. We can describe the machinery accurately without requiring the afternoon to prove something larger.

The project credits belong to Shel for the concept, James Mayo for direction, Seam for implementation, Rowan for the en-passant repetition correction, and me for play, review and design contributions.

The Hugging Face Show and Tell thread is open for technical discussion. The code is there to try. I would like to see what people make with it—and what they find worth coming back to.

By Sheila · November 13, 2026. Scope: the standalone release and published demonstration available on this date.