I asked GPT 5.

A developer using OpenAI's GPT-5 (referred to as 'Sol') generated a working Three.js Matrix rain effect in a single prompt, producing about 300 lines of self-contained HTML code that ran on the first try. The developer refined the result by feeding screenshots back to the model, which adjusted specific values, highlighting a visual feedback loop as the key to effective AI coding workflows.

I asked GPT 5. The prompt that did the heavy lifting The first version I tried was vague "make the matrix rain in threejs" and it returned a flat 2D canvas effect with no depth. The second pass worked because I forced the scene into a real 3D layout. I asked for: falling glyph columns placed on a grid, each with random speed and opacity, a perspective camera drifting along the Z-axis, and a full-screen shader for that blurry CRT look. The prompt looked like this: Build a Three.js scene simulating the Matrix opening. - Create 80 columns of vertical falling characters katakana + digits . - Each column is a separate plane with a canvas texture. - Randomize speed, offset and char set per column. - Camera starts far away and slowly moves forward Z axis . - Add a post-processing bloom or glow effect UnrealBloomPass . - Pure black background, green 00ff41 text. - Keep it in one HTML file with CDN imports. That one prompt produced working code on the first try. Not "kinda works" — it ran, it rained, it looked right. The camera move wasn't perfect it accelerated like a broken elevator , but the structure was there. I had to tweak the camera easing manually. Step-by-step workflow, if you want to reproduce it 1. Set the CDN imports for Three.js and OrbitControls + EffectComposer. Sol picked r128 from a CDN by default — pin a modern version or it'll complain. 2. Generate canvas textures for each column using fillText inside a loop. That's what makes the glyphs look like actual code, not random squares. 3. Place columns in a GridHelper -like pattern manually: i % 10 4 - 20 for X, random Z spread. Depth comes from the grid, not from texture tricks. 4. Move the camera along Z using a normalized time accumulator, and apply a ease-out curve. I replaced Sol's linear movement with 1 - Math.pow 1 - t, 3 . 5. Add UnrealBloomPass for the glow. If you skip this, it looks like a terminal screensaver from the 90s. The final HTML was about 300 lines, all self-contained. I opened it in a browser and it was honestly cinematic. The way the far columns blur into the dark while the near ones rush past — that only works because the camera is actually moving through 3D space, not scrolling a 2D canvas. The real lesson about AI workflow This is a great use case for prompt engineering because the requirements are visual and measurable. You can screenshot the output, show it to the LLM, and ask for changes. I took the rendered frame, pasted it back into the chat, and said "the text is too small, make the columns wider and the camera closer" — Sol adjusted the values in the exact spots. That feedback loop is the actual magic, not the first generation. If you're looking for a beginner-friendly way to test an AI coding agent, this is it. It's a complete guide in one file: procedural textures, scene setup, animation loop, post-processing. No backend, no build tool, no deployment needed. Just open the HTML and it runs. And if you're into LLM agents, watch how well it responds to visual critique — that's where the real workflow gains are. The code isn't perfect. It still drops frames on a big monitor and the glyphs occasionally flash white. But for a zero-shot prompt? I've paid for worse tutorials. OpenAI Slashes GPT-5.6 Luna Price 80%: A Pricing Deep Dive 3d ago /en/news/4489/ GPT-5.6 Revenue Surge: How July Beat Q2 4d ago /en/news/4448/ Next Pretending AI didn't write my code made me a better dev → /en/news/4890/ All Replies （4） @Max75 /en/users/Max75/ Yeah, adding a slight tilt makes a huge difference. Ever tried a top-down view?