Backyard Arcade: A Tiny Arcade Director Powered by Small Models
Backyard Arcade is my submission for Build Small Hackathon Track 2: An Adventure in Thousand Token Wood. The track asks for something delightful, strange, and powered by AI. I wanted to build a playful arcade toy where the idea is simple: describe a tiny game in natural language, then play it inside a Gradio arcade cabinet.
Live Space: https://huggingface.co/spaces/build-small-hackathon/game-builder-ai
The goal was to make learning and experimenting with LLMs feel fun instead of abstract or work-focused. A game is a good way to see whether the model understood you. If the prompt works, something moves. If the rules make sense, you can dodge, collect, jump, grow, win, or lose. It also nudges the user to think about what they want and how to guide the model toward it.
The app lets a user type prompts like “make a snake game where I eat glowing orbs” or “make something strange with frogs dodging moon rain.” The AI then turns that into a playable mini-game. The user can remix the game, add a powerup, tune the pace, or ask for a strange mix. There are preset options too, so users can see what a working prompt looks like.
How This Fits Track 2
Track 2 is about making something fun that would not exist without LLMs or AI. I wanted to experiment with dynamic UI and generated play experiences, because that is something LLMs make possible. A game felt like the logical choice because it is immediate, visual, and easy to share.
In Backyard Arcade, the model acts like an arcade director. It reads the user’s idea, understands the intent, chooses the kind of game that fits, invents the theme, names the cartridge, sets the goal, and decides how strange the game should be.
The current app supports a small set of reliable game types:
- Dodge games
- Collector games
- Jumper games
- Snake games
- Strange mixes that blend two game types
Since the track celebrates strange and joyful ideas, I added Strange Mix mode. It combines game types to bring a little more weirdness into the arcade while still keeping the game playable. I suggest trying it yourself.
How The AI Powers It
The app uses a smolagents tool-calling agent. The LLM does not just answer questions; it chooses a render tool and fills in the game design: title, theme, goal, difficulty, speed, hazards, collectibles, and twist.
The renderer then turns that validated design into a playable p5.js game. This made the app much more reliable than asking the LLM to write all the JavaScript from scratch, while still keeping the AI central to the experience.
Why Small Models Worked Better Than Expected
The original idea was local-first: let people learn and play with LLMs without token costs. Local 9B-class models were already good enough to understand short game prompts, choose a direction, and fill in missing details.
The main lesson was that small models work much better when they are not asked to do every fragile implementation detail. Asking a small model to write a whole browser game from scratch is hard. Asking it to act as a game designer inside a tool-based system is much more realistic.
For the deployed Space, I used Hugging Face Inference API with Qwen/Qwen2.5-Coder-32B-Instruct, which stays within the hackathon limit. The architecture still supports local models through Ollama, but the hosted version is easier for people to try immediately.
The Gradio App
The custom UI is built to feel like a small arcade workshop. The left side has chat, quick cartridges, controls, and tuning sliders. The right side is the playable game cabinet.
The quick buttons are not hardcoded games. They send prompts through the same agent path as normal chat. That means “Orb Snake” or “Frog Rain” are starting points, but the LLM still designs the cartridge. The LLM is always the conductor.
Closing
Backyard Arcade became better when I stopped asking the LLM to generate everything and instead made it the arcade director.
It chooses the game, gives it a theme, tunes the rules, and makes it strange. That balance is what I wanted for Thousand Token Wood: something small, weird, joyful, and genuinely shaped by AI.