So I’ve been using AI for a while. But I have found this to be not only the most economical but also the BEST way to use it.
First, I’d get VS Code with defold kit extension. Then use the AI’s through VS Code.
I actually have all the AIs (Claude pro, ChatGPT Codex, Gemini (which sucks), and CoPilotPro+ via GitHub.
If I could only buy one it would be CoPilot Pro+ in GitHub as it gives me all the AIs (plus several free ones if I want to go on the cheap).
But What you really want to do is ask for a plan first. Then take that plan the AI gives you to a more expensive AI model (Currently Claude Opus 4.7 is the most expensive). Tell that model, I asked this question: “…”
Review his plan and let me know if you agree.
Then take Claudes output and give back to the previous model and ask if claude is correct.
If happy, then say implement this solution.
And if you want go back to Claude and ask if it was implemented properly.
I hope this helps…
The reason I like CoPilotPro+ via GitHub is the claude and ChatGPT models keep limiting my requests. Now I have a work around to ChatGPT. I have a pro subscription, so I run two accounts (that’s required to have pro)…then when they run out of tokens I just delete one of the users and add a new one. So technically I have unlimited tokens. I didn’t want to share this tip as I don’t want to lose it, but with CoPilot Pro+ I really don’t care…that will be what I use going forward.
Actually yes. I created a suica like game a while back and it didn’t feel right. The physics were way to slow. I finally went back to revisit it and it took about 1 hour of AI to fix it and add a couple quality of life features.
I’m about to release the game on GITHUB, free Open Source.
It’s a valid question! I don’t think you can one-shot a good game using AI. The AI is like a naive junior developer that you need to hand-hold and constantly nudge and prod in the right direction, and even then it might sometimes be quicker to say “fuck it, I can fix this myself faster than I can prompt the AI to do what I want”.
In order to make the AI more efficient the tools you use must have the capabilities of providing the AI with a good feedback loop so that it can evaluate and iterate on a task on it’s own. Coincidentally some of the work we do to make integrations with external IDEs and tools smoother will also be helpful for someone using an AI/agent:
The /console/stream endpoint of the editor is used to stream console output to VSCode. The same endpoint can be used by an AI to understand what’s going on when the game is launched.
The OpenAPI spec we added to the editor can help an AI discover how to interact with our tools. But at the same time it’s a convenient way for us to document the public API of the editor “service” when used by other IDEs
We are working on an automation extension that can be used to run end-to-end / functional tests of a game (explore the scene graph, click, drag etc), similar to what you can do with Poco, but more generic. This can also be used by an AI to interact with a game to test if a change works as intended.
But even with these improvements it’s my opinion that you’ll need that human touch in the end to make something which really stands out. It is really hard to convey in words what makes a game really good, or what makes a certain mechanic or effect stand out. It happens in our brain, in our gut and in our heart.
Absolutely. AI really is getting a lot better and needing less and less hand-holding, but early on sometimes it just couldn’t figure out what I wanted (and that’s when you switch AIs) or you have to explain it in a way that will get the AI to understand. Like you have to know what Defold is able to do and how it does it, because AI used to want to throw in code that didn’t work with defold at all. When that would happen, I’d just ask the AI to show me in documentation where what he’s doing will work. You have to treat it like a novice sometimes and make him learn on his journey
BTW: the box2d v3 has AIs confused. Claude can’t even do it…at first it says it’s on be default…then when I show it what ChatGPT Codex say (default is is the older box2d v2) it agrees claude is wrong. But then when asking for the manifest it can’t do it. And even Codex says it needs to find examples in the forums. SO maybe the documentation can be improved on this.
