So, I took the famous 2048 puzzle game and decided to recreate it myself. Not for production, just an experiment to see how YOLO mode actually handles a full build when you stop micro-managing the coding agent.
Pro-tip: If you’re going to run YOLO mode, do it in a sandbox.
I used the Gemini CLI to vibe code the whole thing with Gemini 3 Pro for development.
Here’s exactly how it went down:
1. Start YOLO mode
This is like giving the permission to do whatever it takes to get the work done. It can execute commands, create files, and iterate without asking for a "Yes/No" confirmation at every line of code.
gemini --yolo
2. The Visual Prompt
Instead of typing out a massive spec, I just took a screenshot of the original 2048 game and pasted it directly into the terminal. Let the model review the UI so it actually understood the spatial logic before it started development.
3. The Context Loop
Now here the prompting skills comes in. See the example below how I started:
I want to recreate this game as a chrome extension... explain briefly.
And let the gemini ask me the follow-up questions. This ensured it actually grasped the requirements and manifest v3 constraints before it wrote a single line.
4. Ninja technique for vibe coding
Once the context was set, I had the terminal to create GEMINI.md and divide the whole development into easy achievable tasks and save it to TODO.md.
Having these files meant the agent always had a "source of truth" to look back at. It stayed on the plan regardless of what happened to the session.
5. Let Gemini Cook
At this point, I just sat back. The model handled the grid logic, the tile merging, and the extension packaging.
Conclusion:
This is the first time I built any chrome extension. The game works perfectly. It’s smooth, the logic is tight, and honestly I enjoyed throughout the process. While I still need to add a memory layer so it remembers your state when you close the extension.
Top comments (0)