As the number of "atmospheric programming" cases continues to grow, the threshold for AI programming is once again being redefined. Former Meta research engineer Caleb Leak recently shared a fun experiment: he successfully taught his 9-pound Corgi dog "Momo" to develop a fully functional game using Claude Code.

image.png

To achieve "dog programming," Caleb built a precise hardware and algorithmic process. Momo inputs by randomly tapping a Bluetooth keyboard connected to a Raspberry Pi 5, and the signal is then transmitted to a small program called "DogKeyboard" for filtering and forwarded to Claude Code. To encourage Momo to continue "creating," the system automatically triggers an intelligent feeder to dispense snacks whenever a preset amount of text is input.

The core of this experiment lies in how AI can understand random code. Caleb set up extremely specific prompts for Claude, positioning it as an eccentric genius designer who can only communicate in extraordinary ways, and informed it that the random code contained a mysterious game idea. With automatic feedback tools such as screenshot feedback, scene checking, and shader verification, Claude successfully transformed these random characters into a playable game within 1 to 2 hours.

Momo's latest work is named "Quasar Saz." In the game, players play as the character Zara, fighting with a cosmic saz instrument. The game was created using the Godot 4.6 engine, with 100% of the logic code written in C#, featuring six levels and boss battles, as well as impressive graphics and sound effects.