I'm making a fairly large game by engine standards. It's an isometric action/adventure game, with ~7,000 events at the moment.
My second developer is also an artist, so I'm stuck working on my code alone. :)
To avoid turning development into a never-ending process, I use everything I can, and as soon as AI Studio came out, I created a bunch of different assistants to understand how it works and what real help it can provide.
Ultimately, I settled on an app that fetches a C3 project from a GitHub repository (it updates instantly, as it ignores everything except ES and scripts). I select the sheets to work on in a given session, and the model can also open related sheets for analysis. The AI sees the sheet's JSON, and I'm the C3 editor. Save, push, and the assistant sees the changes in the project.
The app lets you select a model, adjust the creativity temperature, and the context window bar displays the number of tokens used.
Having ingested the project sheets, the model created a prompt for itself that allows its code to be pasted into the C3 editor, with varying success. This works well for modification, but less so for creating code from scratch. Yes, large models are proficient in languages, but valid C3 clipboard syntax was clearly not a priority in their training.
On the one hand, it's annoying to wait for an assistant to correct its syntax, but on the other, it's a barrier to errors.
So, in order of usefulness:
1. Information. AI can do this instantly — for example, find Ashley's non-obvious forum post in which he admits that the animation trigger is delayed by one tick, and he himself doesn't remember why. Or the AI will produce questionable information, which you criticize, thereby creating a logical conflict. Neural networks work better in conflict. So, if you're up for the investigation, you'll get a wealth of information for creating and releasing the game, step by step.
2. Mathematics. The AI rewrote all isometric distance calculations for detection and movement, with forward and reverse transformations. It wrote JS scripts that converted angles to radians, linked them to callable functions, and suggested using an angle matrix for massive trigonometric calculations—and even seemed to enjoy this part of the work.
3. Debugging. As of late February 2026, two top-end models (Opus 4.6 and Gemini 3.1 Pro) were able to understand my project's logic and find errors in linked sheets of 500-1000 events 8 times out of 10. But they're very expensive.
And there's the free Gemini 3 Flash, which can make guesses, 4-6 guesses per message. And if it doesn't guess, it will give you a sense of collaborative work, and you'll find the error on your own, but with less effort.
4. Design. A special case of the first point. In fact, sometimes it's actually very useful. One good conversation can save you from a mistake. Almost all AI assistants are good at detailing and planning, which gives a more unbiased view of your project and its architecture.
5. Generating code for game mechanics in the form of classic C3 editor events. This is the weakest point of almost any AI. It's not even a matter of syntax, but rather that the neural network always simplifies the code within the specifications, weakly analyzing its compatibility with the project.
Judging by current progress, accessible models capable of responsible analysis should appear this year. This means that all programmers will use AI in one way or another.
What can Scirra do? I honestly have no idea.
Today I saw a video on Twitter of a large 3D game made based on a prompt, and a program that controls agents working on the project. Too much money has been invested in the tsunami heading our way.