Yep! Just merged the metal backend in FIgDraw v0.12.0. GPT5/Codex rocks this stuff if you supervise it properly and cleanup the code a bit.
It's a bit faster too - the msdf star example was causing the FPS to drop to 90-100FPS but with metal it's back up to 120FPS on my machine. That's good cause it's only like 3 stars.
You're on fire!
Thanks!
I am thinking that we miss a SKILLS.md file that guides an agent through the process, and I want to experiment with that.
Possibly? I just have my AGENTS.md. It helps the Coding Tools to avoid dumb work or getting confused. An important piece is to not use Nimble and to prefer Atlas so that the tools can find the relevant context easier and can compile and pass Nim flags properly.
They also need to be told to make tests.
It depends heavily on the Coding framework. I find GPT5/Codex to be slower but to produce smaller more manageable changes that tend to work almost every time. In contrast Claude Code is much faster and pretty good, but tends to produce much verbose code and to mess things up more.
However, Cursor, Warp, Gemini's code tool and the like all suck in my experience. The real differentiator lately is less the models and more the prompt engineering.
Here's my prompts:
› Please create metal4 api wrappers. Reference https://developer.apple.com/documentation/ metal/understanding-the-metal-4-core-api?language=objc and https://developer.apple.com/ documentation/metal/drawing-a-triangle-with-metal-4?language=objc . For objective-c build from deps/darwin wrappers.
That failed since I'm on an older mac without metal4 and the docs requires JS, but it was smart enough to find the Rust bindings online and use them.
› Great, now add a unit test for it, and add it to config.nims. Run it using nim test. › Great, now add a test for drawing a triangle.
PS there's some indications that 'great!' and even a bit of flattery make LLMs work better. Gemini especially needs coddling. ;)