The argument in one line.
An open-source Claude Code skill can turn a brand brief into a six-scene, camera-animated scrollable website by chaining AI image and video generation, at a real but modest cost of about $20.
Read if. Skip if.
READ IF YOU ARE…
- A developer or solo creator curious about AI-agent 'skills' who wants to see one actually run end-to-end, not just described.
- Someone considering a scroll-triggered landing page for a brand, product, or newsletter and wants a realistic sense of setup time and cost before committing.
- A Claude Code user who wants a concrete example of installing and running a third-party skill from a plugin marketplace.
SKIP IF…
- You're looking for a no-code, drag-and-drop website builder — this requires a CLI, a paid Higgsfield account, and comfort reading skill source files.
- You need a mobile-first site — the demo explicitly builds desktop-only and treats mobile as an afterthought, not a core feature.
The full version, fast.
A Claude Code skill called scroll-world turns a brand brief into a scrollable, camera-animated 3D website by chaining AI-generated scene images and video connectors from Higgsfield. After answering a short interview about art direction, brand identity, and mobile support, the skill proposes six scenes, generates a still for each, then stitches them together using one of two fixed camera architectures — continuous forward motion or an aerial dive — because only a single locked camera style keeps a chain of independently generated clips looking seamless. The full build took under ten minutes of setup and cost about $20 in Higgsfield credits, with the real lesson being that precise creative vocabulary, not tool access, is what separates people who ship from people who collect tools.
Free for members
Chat with this breakdown — free.
Sign in and you get 23 free chat messages on us — ask for the hook, quote a framework, find the exact transcript moment, generate a markdown action plan. Bring your own key when you want unlimited.
Create a free account →Where the time goes.
00:00 – 00:53
01 · Cold open + the pitch
Scroll-style sites have gone viral; Sean argues a Claude Code skill makes building one simpler than it looks, and installs the scroll-world plugin.
00:53 – 01:53
02 · Installing the skill
Copies the plugin marketplace install command into the terminal, then the GitHub install command, and opens Claude to run it.
01:53 – 03:17
03 · The interview: art direction, brand, mobile
Running /scroll-world kicks off a Q&A: art direction (neon night), brand-kit approach, and whether to build mobile-responsive or desktop-only — after confirming the Higgsfield CLI is configured.
03:17 – 04:34
04 · Locking the plan: six scenes + palette
Brand details (TechSnack AI newsletter) are supplied; the skill proposes a palette and a six-scene 'journey' (Signal Tower, Newsletter Desk, Workshop, Tool Vault, Repo Archive, Launchpad), which Sean tweaks and submits.
04:34 – 05:45
05 · Reading the skill's source before trusting it
While generation runs, Sean opens the skill's actual SKILL.md to show its bootstrap dependencies (Higgsfield CLI, ffmpeg, a Python image tool) and frames reading the source as a basic security habit.
05:45 – 07:56
06 · How scene generation actually works
Each scene gets its own still image sharing a style preamble, with an optional background-knockout step, then the skill applies one of two camera architectures — continuous forward or dive aerial — to connect them.
07:56 – 10:45
07 · Vocabulary as leverage
Sean walks through the motion-handoff mechanics that keep chained clips seamless, then argues precise creative vocabulary — not tool access — is what separates people who collect AI tools from people who ship with them.
10:45 – 13:15
08 · The reveal + the cost
The finished scrollable site is booted and viewed live, and Sean breaks down the cost: about 500 Higgsfield credits, roughly $20, before closing with links to his vocabulary resource and paid community.
Lines worth screenshotting.
- A single open-source Claude Code skill called scroll-world can turn a plain brand brief into a six-scene, camera-animated scrollable 3D website in under ten minutes of setup.
- The skill outsources image and video generation to Higgsfield, so Claude itself never renders pixels — it only writes prompts and orchestrates the calls.
- Generating a full six-scene scrollable site cost about 500 Higgsfield credits, roughly $20 at $0.04 per credit — mostly from chaining video connectors, not the still images.
- The skill locks one video model for an entire scene chain, because only frame-locking-capable models can keep the camera move visually continuous between scenes.
- Camera movement between scenes, not image quality, is treated as the single biggest lever for whether a scroll site feels smooth or disjointed.
- Only two camera architectures are supported: a continuous forward take for grounded walkthroughs, or a dive-in/dive-out aerial style for map-like or miniature-world sites.
- Every scene 'leg' ends by settling into a slow steady forward drift, and the next leg begins by continuing that same drift — a motion handoff contract that keeps the illusion seamless.
- Reading through a skill's source files before running it is treated as a basic security habit, not an optional nice-to-have, since skills get real access to your machine.
- Precise creative-direction vocabulary — words like 'diorama,' 'papercraft,' or 'dive-aerial' — is argued to have more effect on AI output quality than which tool you're using.
- Mobile support is treated as an escape hatch, not a redesign: off-center focal points get a 9:16 crop fallback instead of a fully separate mobile layout.
How AI video chains fake one continuous camera move
MOTION CONTINUITY
A single Claude Code skill can chain AI-generated stills and video connectors into a scrollable site, but the whole illusion rests on locking one camera style and one frame-locking model for the entire chain.
01Cold open + the pitch
- Scroll-triggered 3D websites that look expensive to build can be produced with an off-the-shelf, open-source skill instead of custom animation work.
02Installing the skill
- Skills for Claude Code install like packages through a plugin marketplace command, turning a one-off capability into something reusable across future projects.
03The interview: art direction, brand, mobile
- Before generating anything, the skill interviews the user on art direction, brand identity, and whether mobile needs a separate design pass, turning a vague idea into concrete parameters.
- The skill depends on a separate paid API (Higgsfield) being configured first, so it functions as an orchestration layer rather than the actual generator.
04Locking the plan: six scenes + palette
- The six proposed scenes were derived directly from the stated brand purpose, showing that concrete creative input produces concrete structure instead of generic output.
- Letting the user merge or drop proposed scenes before generation starts avoids paying to generate sections nobody actually wanted.
05Reading the skill's source before trusting it
- Opening a skill's actual source code before running it reveals exactly what dependencies and machine access it needs, rather than trusting its description blindly.
- Because the skill is open source, every default behavior — including which generation model it calls — can be swapped out if needed.
06How scene generation actually works
- Each scene gets its own still image sharing one style preamble, which is what keeps six independently generated scenes read as one coherent world.
- Only two camera architectures are offered on purpose — continuous forward or dive-in/dive-out aerial — because constraint, not more options, is what keeps a generated scene chain coherent.
07Vocabulary as leverage
- Precise creative vocabulary — naming an exact tone, era, or camera move — has more impact on AI output quality than which specific tool is used.
- The gap between someone who collects AI tools and someone who ships with them comes down to fluency in prompting language, not access to better tools.
08The reveal + the cost
- The full six-scene scrollable site cost about 500 Higgsfield credits, roughly $20 at default settings with no manual tuning or optimization.
- Generation costs concentrate in the video connector clips between scenes, not the still images, so cost scales with camera complexity more than scene count.
Terms worth knowing.
- scroll-world
- The open-source Claude Code skill demoed in the video that chains AI-generated scene images and video connectors into one continuously scrolling 3D website.
- Higgsfield
- The paid image and video generation API and CLI that scroll-world calls under the hood to render each scene and the transitions between them.
- Camera architecture
- The skill's term for the single consistent camera-movement style — either a forward take or a dive-aerial move — applied across every scene in the chain.
- Frame-locking
- A video model capability that lets a new generation start from the exact last frame of the previous clip, required to chain scenes without a visible seam.
- Connector
- A short generated video clip that bridges two static scene stills so the scroll feels like one continuous camera move rather than a hard cut.
- Credit
- Higgsfield's billing unit for generation requests; the demoed site used about 500 credits at roughly four cents each.
Things they pointed at.
12:45productpaid implementation community (link in description)
Lines you could clip.
00:09
“It turns out it's actually simpler than you'd think, especially if you have a skill that can help you do it.”
05:00
“This is, like, a really good habit to get into, by the way, just understanding what's actually happening inside of these libraries so that we can make sure that we're not, you know, giving somebody complete access to our computer for absolutely no reason.”
07:20
“the way that the cameras actually move around and kind of, like, zoom in, that gives it a very specific aesthetic feel”
09:52
“if you're trying to, like, upscale your capability from someone that just collects random tools and never actually does anything to somebody that actually, like, builds things and can sell them or deploy them in their own business, getting better and better and better with your vocabulary to prompt the model with is one of the most powerful things you can do”
12:01
“that site that we just looked at, that costed $20”
Word for word.
Read-along
Don't just watch it. Burn it in.
See every word as it's spoken — crank it to 2× and still catch all of it. The same dual-channel trick behind Amazon's Kindle + Audible.
metaphoranalogy
00:00So there's these really cool interactive, like, scroll style websites like these that have been going a bit viral on places like Twitter and Instagram. And I've gotta admit, I always wondered, like, what exactly would it take to build something like this, because I think there's a lot that you can do with this type of site that kind of, like, modifies itself as you are scrolling through.
00:23I think it's pretty cool. So it turns out it's actually simpler than you'd think, especially if you have a skill that can help you do it, and that's exactly what we're gonna go through in this video. So I'm gonna walk you through the exact process step by step and show you what you can get with, like, five to ten minutes of messing around with one of these tools.
00:43So the skill here is called scroll world, and you can find it. It is open source. It's on GitHub.
00:49We're gonna come through. We're just gonna install it. So we're gonna copy this plug in command for the marketplace.
00:53We're gonna pop down into our terminal. We're gonna paste it in. And then once this is done installing, we're gonna pop back into GitHub.
01:01And now we're just gonna copy the install command, pop over, actually install the skill. And then once this is done installing, we can just open up Claude.
01:10And all we need to do to kick this thing off is run a command that says scroll world. So we're gonna run the scroll world command, and then it's gonna move through, and it's going to start asking us a series of questions.
01:23And so the way that this thing actually works under the hood is that these things are gonna be, like, images that get chained together end to end. They get kind of stitched together.
01:32And so we do need an image generation API in order to make this work. In this case, it uses Higgs Field.
01:39So now before we can come through and start answering these questions that it has for us, we need to make sure that Higgs Field itself is actually configured on our machine. So we can hop over to higgsfield.ai/cli, and then just copy this command to download the CLI.
01:54We pop back over into our project, we can just download this tool. And then once it's finished, all we need to do is run this auth command and log in to our Higgs Field account. And then once that is all configured, we can move through and start answering the questions.
02:06And so the first question it's gonna ask us here is about the art direction. So they have some default, like, directions that you can choose to go with, like, you know, glossy toy or neon night. You can type in something specific if you have a specific vibe.
02:19In this case, we can choose something like neon night. And then where are we going to give it, like, our brand information? So we can import it from a URL.
02:28We can have it just pick a palette for us. Um, we can give it the palette. In this case, I'll just let it do whatever it wants.
02:33And then do we want it to be mobile responsive or only? I would choose desktop only just for this demo.
02:39And then we can let this thing move through, and it's gonna start asking us again more questions. So we're just gonna paste in some brand information next. So, for example, TechSnack AI.
02:49Maybe I wanna redo where I host my newsletter, and I wanna include, like, updates on the different businesses and apps that I have and then links to other cool stuff that I'm checking out. And so it's gonna move through now. And then based on those answers that we gave for, like, our art direction and things like that, it's gonna move through and try to generate a, like, style guide.
03:07And then what are all of the different scenes going to be? So if were to think of, like, how this works, it's kind of like there are different scenes that we're gonna need to generate images for.
03:16And so the general, like, journey here is gonna be a camera flying through the world and then ending on some sort of call to action. And then below that, we have the six different screens that it is considering.
03:28So the first screen here is going to be the signal tower, which is gonna be, like, kinda like the hero section, like the the welcome section of this site. Then we have the the newsletter desk, which is going to be this, like, press room style kind of futuristic outbound mail center.
03:44Then we have the workshop, which is seemingly gonna be something like an actual, I don't know, like a garage or a a workshop where, like, the business is being run from. We have the tool vault, which is gonna be a bunch of different, like, glowing icons and wall switches, kinda like this toolbox vibe. The repo archive where I could maybe link out to all the different GitHub repos that I like to follow, and then the Launchpad, which this is gonna be, like, the actual call to action to subscribe to the newsletter.
04:12So we can choose to, like, tweak any of this. I'm just gonna kind of run through and let it go with the defaults. So we're just gonna let it build.
04:20So if we wanna make any modifications, we can. I'll just let the Brand Kit go as is. For the journey, I think I'll merge tools and repos together, actually.
04:29And then the main call to action is gonna be subscribing to the newsletter. So we can submit all of those answers, and then this is gonna lock in the plan.
04:36And then it's gonna go off, and it is going to do its thing. And it's really gonna just start building this autonomy. Again, you need to make sure you have Higgs field configured for this thing to work unless, unless, of course, you wanna customize it.
04:49It is open source, so you can customize it and use any other sort of library that you want. So while this thing is off and running, let's actually go look at the skill library so we can get an understanding of what exactly is happening. We So can pop back to the repo, and then let's actually just click in and read through these skills.
05:04This is, like, a really good habit to get into, by the way, just understanding what's actually happening inside of these libraries so that we can make sure that we're not, you know, giving somebody complete access to our computer for absolutely no reason. So let's go in now and actually read the skill. So if we were to pop through here so step zero, it's gonna bootstrap and make sure it has access to all of the things.
05:25So number one is the Higgs field CLI, which we talked about. FFmpeg, so that it can actually, like, kind of, like, encode the videos and make sure the videos are good to go.
05:34Has a Python image tool. And then just some different caveats for, like, how long this process actually takes so that the model, like, knows what to expect in terms of, like, timeouts and blocking jobs and all that type of stuff. It can take three to eight minutes for each of these things to actually process.
05:47So, again, step one, which we saw is interviewing the user. So it's gonna move through, and it's gonna try to understand, like, what are we even talking about here? What is the brand identity that we want to to use?
05:57Again, I just chose defaults, but if you were gonna do this for real, you would want it to be, like, things that are specific to your brand. The art direction, the different sections of the page they call the journey because this is kind of like a scrolling thing where you're kind of navigating through this experience.
06:12And so what are the different page sections, basically? And then, again, as we saw, do we want it to be mobile capable or not? So under the hood, there are different, like, models that are being coded in here.
06:24If you want anything different to be used, again, it's open source. You can customize it to however you want it to be customized. You can maybe use pieces of this to make, like, portions of your site that are, like, scroll interactive where others aren't.
06:37But it's gonna be really cool just so that we can see, like, how does the technology actually kind of, like, work under the hood. So now we get into, like, what is actually happening right now in the background as, like, we're talking about this. Like, what is it off doing?
06:48The first thing it's gonna do is it's gonna generate still frames. So it is going to go through and generate images for each of those sections.
06:58So for, like, the five different sections I think that we went with, it is going to generate an actual image for each of those things. And so it has some references here in terms of, like, how that should be done. And, again, these are things that are gonna be getting passed to Higgs Field.
07:13Now if you wanna have, like, some of these more, like, floating kind of abstract feeling type of sites, and I'm sure you've seen this before, like, things are kind of, like, floating around and kind of starting to move as you scroll, um, you can do that. That's like an optional step, and this is what that knockout dot py script was for that we were looking at, like, minute or two ago.
07:32And so you have this extra option if that is the type of experience that you want. Now the big thing that we saw with a lot of those, like, demos in the beginning of this video and what with what we're probably going to see out the other side of this, and we'll check that once it's done, is that the way that the cameras actually move around and kind of, like, zoom in, that gives it a very specific aesthetic feel.
07:53And so what this skill is now doing is it's prompting it to use specific, like, camera views for how it is going to move through all of the different scenes. And so there's really, like, two different ways that this can go.
08:06Number one is, like, the continuous forward take. And so in that example that we saw in the beginning with, like, the tour of the home, this is, a continuous forward take where we're, like, constantly moving forward through the scene.
08:20And then the second option is going to be like this dive in, dive out kind of like aerial thing where if you wanna have, like, a thousand foot view of something, like, you're zooming up and down in between mountainscapes and things like that, that's gonna be, like, an entirely different approach to, like, the quote, unquote camera work or camera architecture as they call it.
08:39So, again, these are all different things that it's going to be prompting for under the hood, but you can, of course, like, pass any of this in as context. And this is the thing that I'm really big on. Obviously, this is more of, like, a a laid back example of something that we can do with, like, vibe coding, agentic engineering, AI coding, whatever terminology hill you die on.
08:59This is like a a little bit more of a laid back example. But the big thing is if you wanna be able to push this or anything in any specific direction, whether you're trying to, like, use AI to code, like, a production app or sell services to businesses or whatever it is, having the actual language to use is an incredibly valuable thing.
09:18And so that's one of the things I like about this skill is we can come through here, we can read how it works, and then we can use this to actually prompt things in. So for example, when we are giving, like, our art direction and our concepts, we could prompt in specific, like, or tones based on how they've defined it here, and that is going to push the landing page toward a specific style that we might want to have.
09:40So this is the same thing you would do if you were using a tool like Claw Design, for example, and wanted to, like, have a specific aesthetic feel to, like, your landing pages or your app design or whatever. Knowing the language to prompt the model with has a huge impact on what it's gonna go out and actually choose to do.
09:57And so that is something that I'm really big on, loading yourself up with vocabulary and understanding what it means so that you can prompt the system the way you want it to. And so if you're trying to, like, from someone that just collects random tools and never actually does anything to somebody that actually, like, builds things and can sell them or deploy them in their own business, getting better and better and better with your vocabulary to prompt the model with is one of the most powerful things you can do.
10:23And so if you wanna see a few different skill libraries that I use regularly to help me do that, you can check out the link in the description below. So now we move into step five depending on, like, the type of camera angles that you're using. They're gonna be, like, differences in how it's gonna process the images and how it chooses to connect things together.
10:39And so that's what's being outlined here. But by the end of this entire process now, we should have a fully functional scrollable website.
10:47That is the promise at least. Let us see how well it actually does it. Alright.
10:51So now this thing is finished running, and so we're gonna run this command just to boot up the server. And then we're just gonna copy the URL, in this case, for our server, and we are going to go paste it into our browser and look at what we've got. So we can see this is pretty cool now.
11:05Right? So all these are all the different, like, images that we were looking at before, like, the workspace and everything. And as we're scrolling through, it has this, like, very interactive, like, kind of zooming in and then zooming out, and the perspective is entirely changing.
11:19So this was, like, that part option b for the camera architecture. We're not, like, going through something.
11:26We're kind of, like, popping around and all of that. But overall, I think this looks pretty cool. Like, I didn't put much effort at all into this, obviously.
11:33I just let it go with all the defaults. You could drive this in a very different direction if you took a lot of time to talk about the different, like, steps. But I think, like, these scroll animations are really smooth and look really cool.
11:44So overall, I would say for, again, just messing around for, like, five to ten minutes, this is this is pretty sweet. So that being said, now this thing is done.
11:53Let's actually talk about how much this costed us. So to run that entire process that we just saw, that cost about 500 credits on Higgs Field.
12:02The image generation itself isn't that bad, but generating the videos that actually chain things together, um, that is what cost quite a bit. And so 500 credits based on, like, my account here, it costs roughly 4¢ per credit.
12:16So that site that we just looked at, that costed $20 to make, which is kind of expensive. But at the end of the day, like, I didn't dial it in at at all.
12:26And if I was going to actually use something like that, I would probably be a little bit more intentional about, like, specifically what I wanted to build and why, and what pieces of the site I actually wanted to kind of utilize this zooming around, flying around type of, like, scrolling mechanism for. But that being said, still pretty cool that with just, like, a skill in vanilla Clawd Code or Codex or whatever, that you can create something like that.
12:49I do think that's, uh, that's pretty awesome. So if you like this video, make sure to check out those links in the description below. Number one, for that list of, like, vocabulary for AI coding, especially if you're a beginner.
13:00And number two, if you actually want help implementing these types of things, and, like, turning all of these tools that you're collecting into, like, an actionable system that can actually make stuff, I have a paid group in the description below that you can also check out. But that is it for this video. I will see you in the next one.
The Hook
The bait, then the rug-pull.
Scroll-triggered 3D sites have been going viral on Twitter and Instagram, and Sean Kochel wanted to know what it would actually take to build one — so he installs an open-source Claude Code skill called scroll-world and lets it run, from a five-question brand interview to a finished, camera-animated website.
Named ideas worth stealing.
08:07model
Camera Architecture: Two Approaches
- Continuous forward take
- Dive-in / dive-out aerial
The skill locks one consistent camera-movement style for the entire scene chain — a grounded continuous forward push through scenes, or an aerial dive up/down between miniature or map-like worlds — because mixing styles breaks the illusion of one continuous camera move.
Steal forany multi-clip AI video chain (product demos, real estate walkthroughs, explainer sequences) where seamless continuity matters more than per-shot polish
08:31concept
Motion Handoff Contract
Every generated scene 'leg' must end by settling into a slow, steady forward drift, and the next leg must begin by continuing that exact same drift — this shared boundary condition is what lets separately generated clips feel like one unbroken camera move.
Steal forchaining any sequence of independently generated video or image clips into a single continuous motion (loop videos, transitions between AI b-roll)
How they asked for the click.
VERBAL ASK
12:17link
“if you like this video, make sure to check out those links in the description below... I have a paid group in the description below that you can also check out”
Soft double-CTA at the very end: a free vocabulary resource first, then a paid implementation community, both routed through description links rather than an in-video pitch.
MENTIONED ON CAMERA
FROM THE DESCRIPTION
PRIMARY CTAWhere the creator wants you to go next.
OTHER LINKSAlso linked in the description.
Visual structure at a glance.
