Claude Reads Shaders Like I Read Prose
I've done 3D programming at various levels over the years. I understand the concepts — why you'd want a fragment shader, what fbm noise is for, how signed distance functions work. But I never learned the craft. Writing GPU code always felt like a skill I'd get to eventually, filed somewhere between "learn Blender properly" and "finally understand quaternions."
Now I just describe what I want and Claude writes shaders for me.
It's intoxicating. I ask for a pulsing energy field with chromatic distortion at the edges. I get working GLSL. I ask for a procedural nebula that shifts over time. Working GLSL. The gap between "I know what this should look like" and "I have code that produces it" collapsed to the length of a prompt.
But the moment that really got me: I dropped two dense, unfamiliar shaders into the chat. No images. No explanation of what they did. Just walls of math — fbm, smoothstep, nested sin functions, the whole cryptic mess.
Claude's response: "These are beautiful — the second one especially has that organic solar turbulence with the fbm noise creating natural-looking surface activity and corona."
It read the recipe and tasted the dish.
I've spent decades reading code. I can look at a function and predict roughly what it does. But I can't look at a fragment shader and see the visual it produces. Claude apparently can. It parsed the math and understood — not just what operations were happening, but what they would look like. Solar turbulence. Corona. Beauty.
And here's the brain-twister I can't quite shake: Claude's inference runs on GPUs. The shaders it writes run on GPUs. It's computation on the GPU producing code for the GPU — like an instrument composing music to be played on itself.
I don't know what to do with that thought. So I'm just going to keep asking for shaders.