How I Made Flappy Dog
Flappy Dog was born while I was testing local AI models on my own hardware. The game itself took about five prompts to reach its final state before the AdSense review. Five iterations of describing what I wanted, getting code back, testing it, and asking for adjustments. That is the entire development history.
The Hardware
The rig runs two NVIDIA RTX 5060 Ti 16GB GPUs. The reason for this choice is simple: VRAM. These cards give you 32GB of total VRAM for a consumer budget, which is enough to run the full Qwen 3.6 family of models and the Gemma 4 family models at Q4 or Q5 quantization with KV cache at 8-bit, supporting context windows up to 120K tokens.
For this project, Qwen 3.6 27B was the model used. It has SWE-bench level reasoning capabilities, and with the dual-GPU setup and MTP (multi-token prediction), it can reach around 50 tokens per second in generation speed. That is fast enough for real-time coding workflows where you describe a task, get a response, review it, and iterate.
The Process
The workflow was straightforward. I gave the model a task description — something like "create a Flappy Bird clone with a dog character, 5 lives, and day/night theming." It produced code. I tested it in the browser. I described what needed to change. It produced updated code. After about five of these rounds, the game was in a state where it was fun to play and visually complete.
The rest of this website — the About page, the Guide, the blog posts, the strategy tips — is all AI-generated text written to satisfy Google AdSense's content requirements. AdSense requires a certain amount of original text on a site before approving it for ads. The game is the real product. The text exists to keep the site monetized so it can stay online.
The Game Code
The game runs entirely on HTML5 Canvas 2D with vanilla JavaScript. The game loop uses requestAnimationFrame and handles gravity, pipe spawning, cat animation, collision detection, and rendering each frame. The dog is drawn with radial gradients and geometric shapes. The cats use sine-wave oscillation for tail animation and timer-based blinking for the eyes. The pipes use linear gradients for cylindrical shading. No images, no sprites, no frameworks.
Deployment
Flappy Dog is containerized with Docker and deployed to Google Cloud Run through a GitHub Actions workflow that uses Workload Identity Federation for authentication. The Dockerfile uses an nginx:alpine base image and serves the static files on port 8080.
The Point
This project was a test: can a local AI model on consumer hardware produce working, deployable code from a few conversational prompts? The answer is yes. The game works. The site is live. The text is AI-generated filler for AdSense. Everything here, including this sentence, was produced by a local model on two GPU cards sitting on my desk.
If you want to skip all of this and just play the game, the button is right above. Enjoy.