Started a Bubbleshooter Project

TL;DR: I used Claude Code to build a Bubbleshooter in Unity. It cleverly worked around its limitations by generating editor scripts to set up the scene. AI definitely sped up prototyping, but stick to Sonnet for simple tasks—Opus eats through limits fast.

I’m continuing my journey to learn how to effectively use AI for development. This time, I chose a bigger project: building a Bubbleshooter in Unity. I’ve built several bubbleshooters in the past and understand the mechanics well, which means I can quickly verify whether the AI is producing correct code or if revisions are needed.

I’m using Claude Code for this project. My first step was prompting it to write a PRD for the game. Claude asked some clarifying questions before proceeding, which was a nice touch. I then asked it to generate a task list for building the initial prototype.

Since Claude can’t directly manipulate Unity’s scene, it took an interesting approach: it wrote an editor script I could run via Menu > Setup Game Scene. This script creates the necessary GameObjects and attaches the appropriate components. Probably not the ideal workflow, but a clever workaround.

One practical tip: the Opus model burns through session limits quickly. I recommend reserving it for complex tasks like writing new classes or debugging, and using the Sonnet model for straightforward work.

Overall, AI-assisted coding definitely accelerated my prototyping speed. That said, bubbleshooter is a well-established genre with clearly defined mechanics, plenty of training data exists for this type of game. An entirely new or niche mechanic would likely be harder for the AI to generate. I haven’t tested that yet, but it’s on my list.