Claude Code + Higgsfield

the payoff first, no intro

"That is a website. Not a video, not a render — a live page you can scroll, on the internet, right now. Agencies charge four and five figures for pages that look like this. That one cost me about two dollars and twelve minutes, and I built it with Claude Code and one video model.

And I'm not going to hand you a prompt and wish you luck. By the end of this video the whole thing — the concept, the footage, the smoothing, the site, the deploy — runs as one command. Let's go."

The promise — the three levels

"Three levels. Level one: we get Claude to invent the concept and write the video prompt, and we generate the actual footage — this is the creative, and it does about eighty percent of the work on these sites. Level two is the part almost nobody gets right, and it's why most people's version of this looks janky: making the scroll buttery, and binding the video to the scroll wheel frame by frame. Level three, we take every manual step you just watched me do and collapse it into a single Claude Code skill, then push the finished site live.

Stay for level three. That's the part that turns this from a fun afternoon into something you could actually sell."

concept → prompt → the shot

"Here's the thing about these sites: it looks like the page is moving. It isn't. The page is basically still — it's a video in the background, and your scroll wheel is the play head. So the creative is the whole ball game. Get the footage right and the rest is easy.

So I'm going to make Claude come up with the concept, because coming up with ten good ideas is hard for a human and trivial for a model. I'll say: 'You're the creative director for a premium interactive portfolio. Give me ten radically different scroll-driven landing page concepts. Each one must be a single continuous eight-second extreme macro journey — one unbroken forward push through a material. No cuts, no faces, no text.'

And this is where the human part comes back. Ideas are cheap now; taste is the scarce thing. Whiskey bubble — pretty, but I've seen it. This one though: a dissolving capsule. You start on the gelatin wall as it melts, push through the membrane into a cloud of macro powder, and land on skin. That's got depth, it's got light, and depth is what sells the scroll. Taking that one.

Now I turn the idea into a video prompt, which is a different skill — video models want camera language, not adjectives. So: 'Turn this concept into a high-quality macro video prompt. Match the density of this example: a single continuous eight-second cinematic engineering shot moving forward from above a stormy North Atlantic ocean.' Giving it an example of the level of detail I want is doing more work here than any instruction I could write.

Cinema Studio, sixteen by nine, ten-eighty, eight seconds. Eight seconds is the sweet spot — long enough to feel like a journey, short enough that the file doesn't kill your page load. And notice I'm queuing three, not one. I'm honestly happy with maybe a third of what these models give me, so I'd rather spend the credits and pick, than wait twice. Costs me a couple of dollars, saves me twenty minutes."

make it smooth, then bind it to scroll (the real build)

"Alright, that's the one. And here's the part that separates this from every version of this you've seen that looked almost right.

That clip is thirty frames a second. Thirty frames is fine when it's playing. It is not fine when a scroll wheel is the play head, because your scroll has way more resolution than thirty steps — so you see every individual frame and it stutters. Look at this side by side. Left is thirty. Right is sixty. Same footage, same scroll speed. That's the whole difference between 'nice try' and 'wait, how did you make that.'

The fix is frame interpolation: the model looks at frame one and frame two and invents the frame in between, all the way through. Thirty becomes sixty. I'll just tell Claude: 'take capsule.mp4 and interpolate it to sixty frames per second with ByteDance AIGC.'

And while that's cooking, I'm running a second Claude Code in parallel to start the page — because there's no reason to sit and watch a progress bar.

Here's what I'm asking it to build: 'Build a scroll-driven landing page. The video is the background and it never plays on its own — bind playback frame-by-frame to scroll position with smooth easing, so scrolling down moves forward and up moves back. Preload the whole clip. Add section text that fades in at fixed scroll points, and a tour button in the corner that auto-scrolls the whole journey for people who won't scroll themselves.'

That last one matters more than it sounds. Most visitors won't discover a scroll gimmick on their own — the tour button is what makes it a product instead of a demo.

And there it is. Locally, running, smooth."

one command, end to end, live on the internet

"Now — you watched me do six things by hand: ideate, write the prompt, generate, download, interpolate, build. That's fine once. It's miserable the fifth time. So let's kill all of it.

First, I give Claude Code direct access to the video models with an MCP — one paste, and now it can generate and interpolate without me touching the browser.

Then I wrap the whole thing in a skill: 'Turn everything we just did into a skill called cinematic-site. It asks me for a subject, generates three concepts, picks the strongest, writes the video prompt, generates the clip, interpolates it to sixty, builds the scroll-bound page, and deploys it.'

Now watch. Fresh subject, one command, and I don't touch anything again.

Different subject, same pipeline, live on the internet at a real URL you could send to a client right now. And that's the actual point of this video — not the one site, the fact that site number two costs me one sentence."

Recap + real cost + single CTA

"So: concept from Claude, eight seconds of footage, interpolated to sixty, bound to scroll, deployed — and then the whole thing folded into one command.

And the honest receipt, because I hate videos that round down: that was [X] credits, about $[X.XX], and [X] minutes wall-clock, most of it waiting on generations. Call it the price of a coffee for something people charge four figures for.

If you want the exact prompts and the skill file, they're in the description. And if you want to go further — turning this into something you actually sell, and building the rest of your Claude Code stack — that's what we do inside Claude Codex Mastery, nine dollars a month, link below. I'll see you in the next one."