“PROMPT” The Exact Prompt, Skills and Tools Behind My GPT-6 vs Fable 5.1 Website Test

The prompt, word for word

This is the exact message I sent both models. Typos and all. Copy it.

Swap “AI second brains” for your topic. Everything else stays. That one sentence is doing four jobs: it names the two skills Claude should load, it names the image model, and it gives permission to go big on motion.

Why a lazy-looking prompt works

Most people write a page of instructions. Hero here, three columns there, use blue. Then they get exactly what they described, and it looks like everything else on the internet.

I did the opposite. I told it what I wanted to feel (best looking, beautifully designed) and which tools to use to get there. The skills carry the rules. The model carries the taste. My job is to react.

The skills are the secret. Without them, both models default to a purple gradient and three cards in a row. With them, Claude ran a design roll, generated real renders, built a 3D hero, and then spawned a separate reviewer that sent back eight fixes before it called the page done.

Five things, all free except the AI plan

You need one paid subscription (Claude) and one image tool. Everything else is open source.

How the images got made

Every picture on the page is a GPT Image 2 render. The glowing glass brain, the paper notes floating into it, the night desk at the end. Claude wrote the prompts, generated them in parallel, converted them, and embedded the prompt inside each file so I can regenerate later.

Here is one of the prompts it wrote, so you can see the shape:

If you write your own image prompts, steal these five habits:

The 3D part, in plain English

You do not need to know what Three.js is. I don't write it. The word “3D” in the prompt is enough for Claude to reach for it.

What it built for me: a real glass sphere with reflections, a glowing core of 560 points that fire in waves, six 3D cards orbiting it on tilted rings, a 900-particle vortex spiralling into the middle, and a camera that drifts on its own so the shot never sits still. When you scroll, the cards dive into the sphere and an answer card rises out.

One rule I learned the hard way:if you plan to film it, say so. Motion that looks tasteful on a live page looks dead in a screen recording. “I'm recording a video of this” doubled the movement and it was the right call.

The first version was bad. Here is what I said next.

Nobody shows you this part. The first draft was a yellow Swiss transit map metaphor. Clever, wrong. It took three more messages to land the page you saw in the reel. These are the actual messages.

GPT-6 vs Fable 5.1: what I actually saw

GPT-6 went all in on the hero. Insane animation up top, then the rest of the page went flat. It spent its whole budget on the first screen.

Fable 5.1 spread it out. A strong scroll animation in the hero, images that swap as you move through the how-it-works section, small text animations, and one genuinely great render near the end. More even, less fireworks.

My take: if you want one screen that stops the scroll, GPT-6. If you want a page a client can actually use, Fable. Run the test yourself and comment on the reel with who you think won.

Recreate it in five steps

All Resources