← Back to Games

Arkanoid Clone

Published Last updated

Inspired by Arkanoid for the Nintendo NES, released in the US in 1987. Break bricks, catch power-ups, and play your way through 36 rounds. Play solo or take turns with a friend, on desktop or mobile.

Click or tap inside the game to start. Help, language, and sound settings are in the game menu.

How Arkanoid was built with Codex

I wanted to find out how a game is made by building one with Codex, OpenAI’s AI coding agent. I chose Arkanoid, a game where you use a paddle to keep a ball in play and break bricks. The result was a browser game with the feel of the 1987 NES version and graphics inspired by the Commodore 64.

From idea to build plan

I explained the idea to Codex out loud in Norwegian and asked it to look into how the original worked and which tools would suit a browser version. We needed to work out how the ball should move, what should happen when it hit a brick, and how to control the paddle with a mouse, keyboard, or touchscreen.

We put the sources and findings in RESEARCH.md, the build sequence in PLAN.md, and the working rules for Codex in AGENTS.md. Those files lived in the same project folder as the code.

That research led us to Phaser 4, a framework for browser games. Codex wrote the game in TypeScript and used Phaser for graphics, sound, and controls. The rules for things like ball movement, collisions, and scoring went into a separate game core. Codex could then test what happens when the ball hits a brick independently of how the hit looks on screen.

One playable round first

I asked Codex to build one playable round with a ball, a paddle, bricks, scoring, lives, and simple sound. Then we would stop. I wanted to try the controls, speed, and difficulty before deciding what the rest of the game should be like.

We worked in the same chat. I described what I wanted and asked questions. Codex worked directly with the project files, explained what it had built, and helped me open the game. Once I had played the first round, I could give specific feedback based on what I had seen and felt.

From there we added more rounds, power-ups, enemies, language settings, and a two-player mode. I kept playing and giving feedback. Further rounds of playtesting led to changes in the menus, ball speed, and phone controls, among other things. We updated the plan as we made new decisions.

What we found by testing

The first round felt right when I played it. A later check showed that one row of bricks awarded the wrong points. It was a good example of why we tested in several ways. I could feel whether the paddle responded as I expected. Codex could check exactly how many points a hit should award.

Codex used Vitest to check the game rules and Playwright to run automated tests in browsers. Later, we also used automated playthroughs to see whether the rounds could be completed. Codex assigned some research, code review, and playthrough work to subagents, separate AI agents each given a specific task.

From game project to GLN75.ai

The game project got its own repository on GitHub. The game was published on Vercel and appears here on GLN75.ai in an iframe, a frame that loads the game from another web address. The game keeps its own menus, sound, and high scores. When we tried the embedded version on a phone, we found that the dialogs had less room than they did when we opened the game directly. That needed a fix in the game itself.

What I learned from building it

I wanted to understand more about how a game comes together. That got easier when I could connect what Codex explained with what I had tried myself. An explanation of ball speed made more sense when I could play a round and feel it for myself. As I understood more about what made the game behave the way it did, I could ask better questions and make more specific decisions about what to change next.