I'm a professional computational specialist (working in accelerator physics and applied math), I recently started game dev as a hobby since AI coding tools have gotten very powerful over the last few months. My prototype game Echolumination started as a concept I had years ago but never had the time to build it. My procedure was roughly:
1. Chatting with Claude Fable over the viability of my unique RTS game (comparing existing engines like Unity/Godot or making it a mod for an existing RTS game like SC2). We concluded after some research that making it from scratch is the best approach, and making it browser-based with Pixi.js is a good starting point.
2. After going through the overall scope of the game, I decided on the prototype goals and discussed with the LLM on the proper order to build things in phases and have a clean working setup on GitHub. This is where I designed the coding practices (e.g. decoupling the sim from the renderer, building the main concept of the game's engine, and adding more to it over time).
3. After many weeks of adding features one-by-one in phases (e.g. UI, map layout, unit types, buildings, abilities, how damage is calculated, and much more) and having the LLM fix bugs (roughly 500 Claude sessions (mostly using Opus for coding and Fable for planning)). I finally released the prototype after about 6 weeks from starting.
4. Now I'm focusing on polishing parts, adding new abilities, replacing old assets, and eventually adding content.
My game was coded with AI but I was involved every step of the way. The art assets were built by AI but are simple geometric shapes (no stable diffusion or anything) and the music/sfx are from purchased bundles from real people.