Board Game Computer

Create, remix, and play original tabletop games with an AI that writes the rules, adapts them on request, joins as a player, or runs the game as game master.

Comment

Inspiration

Prototyping a board game is slow. A designer must formalize the rules, create components, enforce legal moves, find players, and rebuild the prototype after every change.

I wanted to make that loop conversational: describe an idea, let an AI modify the real running game, and immediately playtest the result.

What it does

Board Game Computer is a programmable virtual tabletop.

Users can:

  • play a complete game;
  • ask an AI opponent to take a turn;
  • describe a new rule in natural language;
  • have GPT-5.6 add that rule to the running game;
  • immediately test the consequences;
  • inspect the game program;
  • undo, replay, share, and fork a room.

The production demo, Prism Foundry, is an original two-player crystal-and-card engine builder. The first player to reach 8 Prestige wins.

How it works

A room is controlled by a restricted JavaScript-like scripting language running inside a custom interpreter.

GPT-5.6 Designer proposes source code, but the application parses, validates, speculatively executes, and reverses it before allowing it to commit.

GPT-5.6 Luna acts as a player by selecting only from a finite list of legal actions. It cannot directly modify state or invent illegal moves.

Every accepted command produces reversible state patches, enabling undo, replay, multiplayer rebasing, and forks.

Cloudflare Durable Objects persist and globally order shared-room commands, while deterministic clients execute the game.

How Codex was used

One primary Codex session implemented the project milestone by milestone from a frozen design packet: the interpreter, reversible runtime, tabletop renderer, game, AI integration, multiplayer rooms, tests, deployment, and production verification.

Codex was excellent at autonomously implementing and validating large technical slices. The weaker part was human-facing product judgment: early technically complete versions were rejected because they were confusing. Human QA redirected the agent twice, eventually producing the clearer Prism Foundry experience and a game that is genuinely defined by its visible program.

Built with

TypeScript, React, PixiJS, Acorn, GPT-5.6, GPT-5.6 Luna, OpenAI Responses API, Cloudflare Workers, Durable Objects, Bun, Vitest, and Playwright.

Built With

Share this project:

Updates