Dog Vibe-Codes Video Games with a Little Help From Claude AI

A YouTuber has created an unusual AI experiment that turns random dog keystrokes into playable video games. Caleb Leak documented the project, which connects his nine-pound cavapoo, Momo, to an AI coding workflow powered by Claude Code from Anthropic.

The system combines consumer hardware, custom software, and structured AI prompting. Momo types on a Bluetooth keyboard connected to a Raspberry Pi 5. The keystrokes travel over a network to DogKeyboard, a lightweight Rust application that filters special keys and forwards usable input to Claude Code. After a preset amount of typing, a smart pet feeder dispenses a treat. A chime signals when the AI is ready for more input.

Leak says the technical challenge was not the keyboard interface but ensuring the AI interpreted nonsensical input as intentional creative direction. Claude, like many large language models, is designed to disregard accidental or meaningless strings. To address this, Leak engineered prompt instructions that framed the keyboard mashing as cryptic but meaningful design language.

In the system prompt, Claude is told that it is collaborating with an eccentric game designer who communicates through riddles and random-looking characters. The AI is instructed to interpret every input as a valid creative signal and update the game accordingly.

Strong guardrails and automated feedback tools, including screenshot capture, play-testing routines, scene linting, and shader validation, ensure that the output remains functional.

From Random Input to Playable Build

Leak reports that a typical game takes one to two hours from first keystroke to playable build. All projects are developed in Godot 4.6, with game logic written entirely in C#. The automation layer translates keyboard noise into structured development steps, allowing Claude to generate assets, mechanics, and scene updates iteratively.

One resulting title, described as an arcade-style action game, features retro visuals reminiscent of 1980s console design. The experiment demonstrates how AI coding agents can transform loosely defined input into structured software output when supported by well-designed system prompts and validation pipelines.

The project highlights a broader shift in how developers interact with generative AI tools. Rather than precise commands, structured context and feedback loops increasingly shape outcomes. By reframing meaningless input as creative intent, the workflow tests the boundaries of AI-assisted development.

While the experiment is largely playful, it underscores how modern AI coding systems can generate functional applications rapidly when provided with constraints, interpretation rules, and automated testing. Whether used for novelty or serious prototyping, the setup illustrates the expanding flexibility of AI development tools in unconventional environments.

AI & Machine Learning, News