In the second installment of the AI for Game Development series, we explore using language models—specifically ChatGPT—as a tool for game design. The goal remains to build a complete farming game in five days, and today’s focus is on leveraging AI to brainstorm and refine core mechanics.
The Quick Take
The straightforward approach is to ask ChatGPT for design advice. In this case, the prompt was: "You are a professional game designer, designing a simple farming game. What features are most important to making the farming game fun and engaging?"
ChatGPT’s response highlighted five key areas:
- Variety of crops
- A challenging and rewarding progression system
- Dynamic and interactive environments
- Social and multiplayer features
- A strong and immersive story or theme
Given the five-day constraint, the author implemented the first two points through gray-box prototyping. The result is playable here, with source code on GitHub.
Understanding Language Models
ChatGPT is built on language models, which predict word sequences. For example, given "The cat chases the __", a model learns to predict "mouse." The real breakthrough came with transformers (introduced in 2017), which use self-attention to process entire sequences simultaneously. This architecture powers modern models like ChatGPT.
Why does ChatGPT outperform its predecessors? One key reason is Reinforcement Learning from Human Feedback (RLHF), where human feedback fine-tunes the model. For a deeper dive, check out our RLHF blog post.
Practical Uses in Game Development
Beyond brainstorming, ChatGPT can assist with coding entire games. However, it has limitations:
ChatGPT often sounds convincing while being wrong. It doesn't truly understand content—it predicts likely outputs. For instance, its explanation of quaternions (a Unity math concept) appears fluent but is entirely incorrect, as noted in a Reddit thread.
Suggestions for Using ChatGPT
Treat ChatGPT as a tool, not a replacement. Use it for brainstorming and acceleration, but verify its output. For example, you could ask it to implement all five design suggestions, but the author recommends selective application—focusing on the most viable ideas.
Continue to Part 3, where we'll use AI for 3D assets.