A couple of weeks ago I posted about trying to get AI to build one of those cinematic 3D sites. It got way more traction than I expected, and the comments ended up teaching me more than the original two weekends of building did.
Quick recap for anyone who missed it: I tried to get AI to one-shot a cinematic 3D website. The code was fine. The atmosphere wasn't. Everything felt like objects in a void instead of an actual place.
What I didn't expect was how many people in the comments landed on basically the same conclusion independently, just worded differently. One person said "world building is a skill," flat out, no elaboration needed. Another compared it to a painter using AI versus a non-painter using AI — same tool, completely different output, because the painter already knows what they're trying to make before they touch the tool. Someone else pointed out that what looks like "a 3D website" is actually composition, lighting, motion, sound, storytelling, and UI all stacked on top of each other, and most of us only ever notice the 3D part because that's the part that's flashy.
Vibes Are Not Instructions
That last comment stuck with me, because it explains exactly why my prompts kept failing. I was asking for "cinematic" and "atmospheric" like those were instructions. They're not instructions. They're vibes. And AI fills in a vibe with whatever's statistically average for it — which is how you end up with the same gray car in the same gray void everyone else is getting.
So I started picking apart what "world building" actually breaks down into, because saying it's a skill doesn't tell you what the skill is made of. Here's roughly where I landed, and some of this might be useful if you're trying to do this yourself.
The World Comes Before the Object
This was the single biggest unlock. If you design the atmosphere, the lighting mood, the fog, the color temperature first — and then place your hero object into that already-existing environment — it looks like it belongs there. If you place the object first and try to add atmosphere around it after, it looks like a render with effects bolted on.
AI defaults to the second order every time, because that's literally the order most prompts ask for things in. "Show me a car in a futuristic environment" builds the car first.
The Camera Is a Decision, Not a Default
Somebody in the thread mentioned their camera "was just there" — not moving for any particular reason — and that's the same thing I kept running into. A camera that drifts because the code says `drift(time)` is not the same as a camera that approaches something slowly because you want the viewer to feel anticipation.
One is a parameter. The other is direction. AI will give you the parameter every time, unless you tell it specifically why the camera should move the way it should move.
Restraint Is the Part Nobody Prompts For
Every example I generated had too much going on. Bloom plus particles plus a rim light plus a glowing object plus a moving background — because more effects sounds more impressive in a prompt. But the actual sites that get reposted everywhere are usually doing way less than you'd think. One light doing real work beats five lights doing nothing in particular.
The Tools Execute Decisions — They Don't Make Them
None of this is a knock on the tools. It's more that the tools are extremely good at executing decisions and not at all good at making them — and most of us don't actually have the decisions made before we start typing the prompt.
I ended up writing a lot of this down properly, because trying to explain "world building" in comment replies kept turning into half-finished thoughts. The longer version is my attempt to document the decisions that happen before any code gets written — the actual stuff that separates "objects in a void" from a place that feels real.
I'm still curious whether anyone's found their own version of this breakdown: the actual ingredients of "atmosphere," as opposed to just calling it a vibe and moving on.