FirstCoffee Studios - LLM Philosophy

When I decided that I wanted to make small games for people like me, I also wanted to learn about large language models (LLMs). My day job is in the tech industry, and I heard terms like "AI," "Machine Learning" and "LLMs" being thrown around. At first, I was quite happy to ignore the hype; but then, some real use started to come out of language models and coding agents. As the industry has rapidly matured, some useful models have emerged that made me decide to jump in.

Now, I use the LLMs for what they are; advanced tools that require careful oversight. I may go into more detail later, but the model writes most of the code and I do the careful planning and design. The creatures, the weapons, the systems, and way to play the game is my own design. In all of my games, the LLM is the tool I use to bring my ideas to life; I do not substitute my own creativity for better autocomplete with a polite robot. I hope my transparency about my use of LLMs as helpful tools will provide curiosity to other and help bring their ideas to fruition.

At first, I used Gemini to build a python version of a game. I struggled with it, because I didn't know any techniques or best practices to use the tool. Slowly, I learned. I then moved to Claude to continue building my game, converting to a HTML file. The game grew and got complex, and Anthropic continued to squeeze users (even the ones who paid) with ridiculous usage limits and vague pricing/usage screening.

Eventually, I found OpenCode, and my world opened up. My current setup is using OpenCode and the MiniMax 2.5,2.7 models for planning and execution. OpenCode allows for transparent, pay-as-you-go pricing models that show you your use in almost real time.