A few months ago, I noticed something funny — my four-kilogram Cavapoo Momo (, a cross between a Cavalier King Charles Spaniel and a Toy Poodle ), stole my lip balm right off the table while I was working on a project in Godot. The dog stepped on the keyboard, and I saw strange input in the terminal. My first thought was hacking, but no — Momo just decided to explore my desktop. Time passed, and I found myself out of work. Sharing this story with friends, I thought: what if I actually send this dog’s input into Claude Code? What if I could create something out of this chaos? I decided to try.
And it turned out I could. I spent several months teaching Momo to type on a keyboard connected via Raspberry Pi, and using Claude Code, to create full, playable games. Yes, the dog — really typing with her paw on the keys, getting treats, and as a result, producing Godot games with logic in C#.
It all started with training. The first two weeks, I just placed treats on the keyboard to help Momo understand that this thing is interesting. Then she learned to trace her paw over the keys ( without pressing them ), and I rewarded each touch with a bell sound and a bit of food. Gradually, I made the task harder — requiring three movements instead of one, automating everything through the DogKeyboard app. The dog quickly grasped the logic: type — get a reward.
The technical part was more interesting. The Raspberry Pi intercepts presses from a Bluetooth keyboard, filters out dangerous keys like Escape ( to prevent Momo from leaving the chat ), and sends the rest to Claude Code. When the dog types enough text, the app triggers — a sound signals her that Claude is ready, and the smart feeder Aqara C1 dispenses a treat. I had to tinker with the keyboard setup — I first tried silicone ones, but the older dog, Hana (, a larger spaniel mix weighing 8.6 kg ), tore them apart. In the end, I settled on the Logitech Pebble Keys 2 — compact, durable, Bluetooth, and Momo loved it.
The most interesting part — how I got Claude to work with dog input. Just sending random characters is useless; Claude would just say it’s noise. I needed a system. I created a prompt where I explained to Claude that it’s dealing with a mysterious game designer who speaks only in riddles. If Momo types 'y7u8888888ftrg34BC', Claude should decode it as 'You want a 3D frog catcher' and create a game about a frog catching insects with its tongue. Sounds crazy, but it worked.
The first games were so-so — no sound, no character, no controls. I added clear requirements: sound must be on, WASD or arrow keys, an enemy or obstacle, a visible character. The results improved sharply. Then I gave Claude tools for self-testing — scripts for screenshots, automated level testing, scene file validation. That was a breakthrough. Claude started testing its own games, finding bugs, fixing them, and verifying the results. I watched as it went through all six stages of the combat system to ensure the final boss worked correctly.
What came out? Momo created a bunch of fully functional games. DJ Smirk — an audio experience where each key produces a different sound. Munch — a competitive salad game. Zaaz — a level coloring puzzle. Octogroove — a rhythm game about an octopus on drums, more complex than Dance Dance Revolution. Quasar Saz — a full RPG with six levels and a boss, where the character uses a cosmic saz to fight sound distortions. All these games are genuinely playable, not just prototypes.
But the main revelation isn’t in the games themselves. When I started improving the system, I realized that the bottleneck in AI development isn’t the quality of ideas, but the quality of feedback loops. The games got better not when I rewrote the prompt, but when I gave Claude the ability to see the results of its work. The same tools that let a dog create games from chaos can help you in working with AI — whether you’re a dog, a cat, or just using your own keyboard.
Momo is not a genius designer. She’s just a Cavapoo who learned that hitting a plastic rectangle leads to food. But the system around her — prompts, tools, feedback — turns that chaos into something working. A year ago, it seemed that the gap between random dog presses and software development was huge. Now it feels much smaller. All the code is open if you want to try it yourself — with a dog, a cat, or just your own keyboard.