3d-websites · 6 min read

10 Prompts That Build Genuinely Futuristic 3D Websites with AI

Most 3D websites fail for the same three reasons. Here are ten prompts that avoid all of them — plus the specific words that decide whether you get a spinning cube or something people screenshot.

Team SpinBuild

There is a particular kind of website that makes people stop scrolling. Depth that reacts to the cursor. Light that moves when the page moves. A product that turns slowly in space while the copy explains it.

For most of the last decade, building one meant hiring someone who knew WebGL, budgeting six weeks, and accepting that it would be broken on half of all phones. That maths has changed. You can describe the scene now and watch it get built.

What has not changed is that most AI-generated 3D sites are bad in exactly the same three ways: a spinning object nobody asked for, a scroll that fights the user, and a 9MB page that melts a mid-range Android. All three are prompt problems, not model problems.

Below are ten prompts that avoid them. Each one says what to build, how it should behave, and — the part most people skip — what it must not do.

Before the prompts: the three words that change everything

Add these to any 3D prompt and the output improves immediately.

"Motion must be tied to scroll position, not to a timer." This is the single highest-leverage instruction in 3D web design. A timer-driven animation keeps moving while the reader is trying to read. A scroll-driven one puts the reader in control, and control is what makes a site feel expensive rather than busy.

"Degrade to a static layout when prefers-reduced-motion is set." A real fraction of your visitors get motion sickness from parallax. This one clause keeps the site usable for them and keeps you on the right side of accessibility.

"Target 60fps on a mid-range phone; keep the scene under 1MB." Without a budget, generation optimises for looking impressive in a screenshot. With one, it picks cheaper geometry and compressed textures.

Now the prompts.


1. The depth-layered hero

Build a landing hero with three visual layers that move at different speeds as the cursor moves — background glow furthest and slowest, mid-layer geometric shapes, foreground headline nearly still. Maximum parallax offset 20px so it reads as depth, not wobble. Dark background, one accent colour, generous whitespace. Motion tied to pointer position with easing, disabled entirely under prefers-reduced-motion.

Why it works: depth comes from differential movement, not from large movement. Capping the offset at 20px is what separates "premium" from "amateur".

2. The scroll-driven product turn

Create a product page where a 3D object rotates exactly in step with scroll position — top of section is 0°, bottom is 360°, nothing animates on its own. Pin the object to the viewport centre while the section scrolls past, with three captions fading in at 25%, 50% and 75% of the rotation. Soft studio lighting, one key light and one rim light.

This is the single most persuasive pattern in 3D web design, and almost nobody gets it right, because they animate on a loop instead of on scroll.

3. The aurora gradient field

Generate a full-bleed animated gradient background in the style of slow aurora — three or four colour stops drifting and blending over 20 seconds, rendered on a canvas shader rather than as a looping video. Colours should be muted and dark so white text sits on top at full contrast. No visible banding.

Ask for a shader, not a GIF or video. It is a fraction of the weight and it scales to any viewport without going soft.

4. The glass stack

Build a feature section as frosted glass panels floating at slightly different z-depths over a blurred colourful background. Each panel: subtle backdrop blur, 1px semi-transparent border, soft shadow beneath. On hover a panel lifts 8px and its border brightens. Ensure text contrast stays above 4.5:1 against the blurred backdrop at every scroll position.

The contrast clause matters. Glassmorphism fails accessibility constantly because the background behind the glass keeps changing.

5. The cursor-reactive light

Add an invisible light source that follows the cursor across a dark hero section, catching the edges of the geometry it passes. The light should lag the cursor slightly, as if it has weight. Nothing else on the page moves.

Restraint is the point. One element reacting beautifully beats five reacting at once.

6. The particle constellation

Create a field of 800 small points in 3D space that drift slowly and connect with thin lines when they come within a short distance of each other. The whole field rotates very slowly. Points fade near the edges of the viewport so the field has no visible boundary. Use a single draw call.

"Single draw call" is the instruction that keeps this at 60fps instead of dropping to 20 on a phone.

7. The bento grid with tilt

Lay out a features section as an asymmetric bento grid — one large cell, two medium, three small. On hover each cell tilts toward the cursor by a maximum of 6 degrees with a smooth return. Rounded corners, dark cards, one accent colour used sparingly. Grid collapses to a single column under 640px with tilt disabled on touch.

Disabling tilt on touch is important: without it the effect triggers on tap and feels broken.

8. The scroll-morphing shape

Build a section where a 3D shape morphs between three forms as the user scrolls — sphere, torus, then a rounded cube — with the transition driven entirely by scroll progress. Matte material, no reflections, one soft shadow. Each form holds briefly at its scroll checkpoint before morphing on.

The brief hold is what makes this readable. Continuous morphing reads as noise.

9. The chromatic edge

Apply a very subtle chromatic aberration to the hero imagery — red and blue channels offset by no more than 1.5px at the edges of the frame, zero at the centre. Pair with a faint film grain at 3% opacity. The effect should be felt rather than noticed.

This is the "why does this look expensive" trick. At 1.5px nobody can name it; at 6px it looks like a broken monitor.

10. The weightless CTA

Build a closing call-to-action where the button sits on a dark field with a soft radial glow behind it that pulses once every four seconds, very subtly. On hover the glow expands and the button lifts 2px. Everything else on the section is still and quiet — large headline, one line of copy, nothing else.

Quiet sections convert. After nine sections of motion, stillness is what draws the eye to the button.


The part everyone skips

Run the finished site on a real phone, not a desktop browser resized to phone width. Those are different things. A 3D scene that is smooth in a 400px-wide Chrome window on a laptop can be a slideshow on an actual mid-range Android, because the laptop is still using the laptop's GPU.

Then check three things:

  1. Does it hold 60fps while scrolling? If not, cut polygon count before you cut effects — geometry is almost always the culprit, not shaders.
  2. Does the text stay readable at every point in the scroll, over every background state it passes?
  3. Does it still work with motion off? Turn on reduced-motion in your OS settings and reload. You should get a clean, static, perfectly usable site — not a broken one.

Start with one section, not a whole site

The most common mistake is asking for a full 3D site in one go. You get ten competing effects and no hierarchy.

Build the hero first. Get it right. Then add one more section at a time, and after each one ask whether the page got better or just busier. A site with one exceptional 3D moment and nine calm sections beats a site with ten effects every single time.

Describe the first section and see what comes back.