Lectures

Tommy dreams of a big online game, but Alex suggests a fair stepping stone — multiplayer on one computer with an AI agent as the second player. All that remains is to choose a game where the agent has nothing to peek at.

Sam reports “chess is ready” twice, and twice the game does not work. Tommy learns the first law: “looks done” does not mean “works,” and starts a checklist that he uses to measure the game point by point.

A detailed spec of two hundred lines does not save you; cutting the rules down is a dead end. The way out is to take the ready-made chess.js library and connect it through a CDN in one line. Plus page markers so the blind agent understands the state of the game.

Using the same MCP cable, the agent connects to the browser through Playwright — and gets hands: it opens the page, clicks, and moves the pieces by itself. And the /goal command lets you set a “done” condition and walk away while the agent keeps playing the game until it reaches the result.

Tommy puts everything together: hooks up Playwright, sits down to play against the agent, then goes off for tea and hands both sides over to /goal. For the first time, he has an opponent out of thin air — and the agent replays the whole game by itself.

Three level articles in one list: what a CDN link to a library actually connects, how the agent opens a page through Playwright and reads the result of an action, and how to write a /goal condition that shows the work is finished.