Hands-On with Gemini 3 Pro: Rapidly Developing an Auto-Battler Game from Scratch

Gemini 3 Pro SOTA Model Capability Test

Live Demo: auto-battler.precu.re

Project Repository: Auto-Battler

Motivation

Coinciding with the official release of the Gemini 3 Pro Preview, I saw someone on a forum testing the one-shot generation of a complex game. Since I’ve been playing Currency War recently, I decided to ask it to write an Auto Battler.

Zero to One

The initial prompt was: 做一个完整的自走棋类游戏，使用二次元手游角色，给我完整的网页代码. The result of the very first generation was actually a very usable version: ver1.html. Even though it had quite a few bugs, it looked decent, and this only took about two minutes. I also tried asking GPT-5.1 High to generate one using the same prompt; while it was also usable, there was a gap in the results: gpt51high.html (You might need to right-click and save as to open locally; opening it directly seems to have some bugs for some reason).

As a side note, although I do frequently use AI Coding, I don’t often use tools like Claude Code or Cursor for small projects. Part of the reason is that I feel small projects can fit entirely into the context without being interfered with by the System Prompts invoked by various tools. Therefore, the subsequent modifications were basically done directly through dialogue.

The next step was generating art assets using the prompt: 请给我每个美术资源的Prompt，人物使用日本二次元可爱美少女风格. My expectation was for it to list the prompts, but Gemini seems to have built-in drawing calls, so once it had the prompts, it drew them for me directly, and I just used them as is.

After that came the implementation of images and various optimizations and fixes. In this regard, the gap with previous models isn’t huge; the modification process still involved multiple rounds to solve various strange issues. The final version took about 10 rounds, mainly for beautification and various detail optimizations. I added a small guide at the end, resulting in the version you see now.

Thoughts

Gemini 3 Pro currently looks like it offers significant quantitative changes, but whether it represents a qualitative change remains to be seen. Even for a SOTA model, its capabilities have boundaries; a simple one-thousand-line game cannot be guaranteed to be perfect and error-free. I might trust a utility script of a few dozen lines written by a SOTA CoT model, but I don’t think I should place the same trust in a more complex system.

It just so happened that Cloudflare experienced a global P0 incident on the same day, and I encountered a GitHub incident while uploading this project. It’s hard not to make the connection that when faced with a massive chunk of code generated by “Vibe Coding,” everyone simply chose the Blue Pill.