4,600 stars came from the channel: Nate Herk's free scroll-website skill scroll-craft

A YouTuber with 1.11M subscribers released a free Claude Code skill that makes AI build premium "scroll it and the visuals move" landing pages, and it reached 4,600 stars in 76 days. The skill itself is well made, but the stars mostly came from the author's channel. If you want to ship an open-source project in this space, what you can learn is his launch playbook and how he turned "taste" into checkable rules. Copying the project itself isn't worth it.

One video, and half the comments are asking where the link is

On the evening of Aug 22, 2026 (UTC 20:05), Nate Herk posted a 16-minute video to his channel "Nate Herk | AI Automation" titled "I Built The Ultimate Claude Website Design Skill (steal this)." In it, he uses the skill live to redo the homepage of his community, AI Automation Society, as a scroll-driven narrative page: scroll down, the video advances frame by frame, headlines assemble line by line, and the whole section stays pinned while you move through it. By Nov 6, the video had 373k views and 5,791 likes (YouTube, 11-06).

Oddly, the video description has no repo link. The first line is his "$1M AI agency playbook," the second is a free entry point to his Skool community. So the comments keep filling up with people asking: "Why can't you just leave a link to the skill?" "i can't find the skill" "Como eu consigo a skill?" The link is actually in two places: at the end of the X long post he published the same day, 20:44, and in his Skool community post (that post has 359 likes and 183 comments).

Keep this detail in mind, we'll come back to it: for him, the repo itself was never the endpoint.

What it actually is

scroll-craft is an agent skill you use inside Claude Code or Codex (it's also packaged as the Claude Code plugin nateherk-design), under the MIT license. It does two things.

First, motion. "Scroll is the timeline": the video is dragged frame by frame as you scroll, sections pin, horizontal tracks slide, and background colors shift with scroll position.

Second, a set of design constraints, and this is what sets it apart from most similar tools. The opening line of the README names the problem: AI-built sites are either tidy but forgettable, or flashy with motion, body text has 2.1:1 contrast, headlines wrap to six lines on mobile, and they all share the same six sections every other AI page has. Its countermeasures:

• 8 mutually exclusive "page grammars": filmic one-shot, chaptered editorial, continuous world, and so on. Each explicitly bans things the others require, so two generations can't quietly converge on the same page.
• Every site must have one "signature move", plus a "fingerprint gate": a new site has to differ from previous ones on at least 4 of the 6 dimensions.
• A banned list: identical feature card grids, gradient text, the AI-purple gradient, fabricated stats, and so on.
• Self-verification: a headless browser screenshots every scroll position, checks for "dead scroll" (scrolling with no visual change), tests text contrast against the composited frame, checks whether video gets stuck on its cover frame, and finally stitches an overview image.

0.3.0 also added a sample set of "ten sites that passed review": AI Automation Society, PERKFORM, Glaido, and others, mostly Nate's own brands and sponsors.

Where the stars came from: almost entirely his own channels

(GitHub, 11-06; before 10-09 there's only one data point at 8-27, so the daily growth in between is inferred. It never once appeared on GitHub Trending.)

That 10-13 jump came the day after his other video, "I Tried 100+ Github Repos for Claude. These 6 Are the Best" (380k views), and scroll-craft was one of those 6. In October he posted two more videos (250k and 580k views), both with "FREE Scrollcraft Website Design Skill" as the first line of the description, linking to a page that only gives the download after you enter your email. The video that got 580k views moved the star count not at all (only +55 from 10-30 to 10-31), because the traffic was routed to the email page, not to GitHub.

Third-party pickup was small: a YouTube Short from Github Awesome got 12k views, most intro videos from a dozen or so small channels stayed under a few thousand views, and several skill directories listed it. I found no big-account retweets on X, and no trace on HN or Product Hunt.

There's also a telling comparison: the repos he paired with other videos include herk-edit-kit (2,597 stars), roast-room (1,935), and AIS-OS (1,838 on GitHub, 11-06). Their contents differ, yet their star counts sit in the same range. The channel decided the stars, not the repo's content.

So many people are doing the same thing. Why did this one win?

The space was never empty. "Claude + AI-generated assets = 3D scroll website" has been a hot YouTube topic since February 2026 (Chase AI's video got 62k views, 2-28).

scroll-craft isn't the first, and it doesn't have the most stars. Its design spec and self-checks really are more detailed than scroll-world's or awwwards'. But auteur came earlier with nearly identical positioning and has only 1,046 stars. The differentiation gave the video an easy hook ("stops AI from making the same page every time"), and the channel brought the stars.

It bet on two things that had already been proven: a visual trend that could explode (3D / scroll websites), and an emotion that projects with hundreds of thousands of stars were already riding ("AI-built sites all look the same"). In Sep 4 it rewrote the README as a "cross-agent skill," which also reached Codex users.

The comments named what it does poorly

The most-liked comments aren't praising it:

"I hate where web development is. Everybody wants all these unnecessary bells and whistles, but what happens when the user is on mobile with a single bar signal?" (152 likes, 8-22)

"what you showed at the beggining and what you ended builidng is completely different..." (120 likes, 8-25)

"The skill itself is very cool, but personally I would never use it because I hate these kinds of scroll-hijacking websites" (12 likes, 8-24)

Someone also asked directly, "How is it different from scroll-world?", and another said, "I'd like a pagespeed optimization skill."

These comments show that a group of people actively dislike scroll storytelling as a format: it's janky on phones, won't load on slow networks, and makes you scroll forever to get information. A skill that delivers "looks good" together with "loads fast, stays smooth on phones, and doesn't hijack scrolling" is a gap nobody is seriously filling right now (inference: the competitors above are all competing on who looks flashier).

If it were you

Grade C: you can learn from it, but copying it isn't worth the effort. Its stars depend on a channel with 1.11M subscribers and a community that claims 450k members, which followers won't have. The same topic already has scroll-world (10k stars) and a string of late entrants at 1–2k stars.

Who it suits: people who already have some distribution (a YouTube or X account with a few thousand to tens of thousands of followers, an active community), or people who really understand frontend performance and motion and want to build a "contrarian" design skill.

Who it doesn't suit: people with no distribution who plan to copy scroll-craft and make another "cinematic landing page skill." auteur is the ready-made result: nearly identical positioning, decent execution, 1,046 stars.

How to do it: don't build a flashier one. Build what it does poorly: a "lightweight scroll-storytelling" skill that writes performance budgets as hard rules (maximum first-screen weight, maximum video size, how to degrade on low-end phone frame drops, the page stays readable with motion turned off), and verifies with Lighthouse and throttled-network runs, with the scores published in the README. You could also change the ecosystem: scroll-craft targets the open web by default, so a version exporting for Webflow / Framer, or for a specific framework (Astro, Next), would be a different product.

How to get stars: Copy his launch playbook. Same day, post the repo, a demo video (before/after, ideally "fixing an ugly website"), an X long-form post, and community posts. Put three completely different finished screenshots on the first screen of the README. Put the repo link in plain sight. Don't hide it behind an email page like he does. You want stars; he wants emails. Under the posts where people complain that "scrolling websites are laggy," reply with your answer.

When is "before": This topic has been hot since February, and new clones were still appearing at the end of October. Think in months. Ideally ship before December.

What would mean you misread it: ① scroll-craft or scroll-world adds performance budgets and fallbacks, which kills your angle; ② Anthropic's official frontend-design skill starts covering scroll and motion; ③ clones launched after October (Motionsites, /awwwards) are still stuck at one or two thousand stars two months later. That means the stars for this topic are already spoken for, and no newcomer will gain traction no matter how good the work is.

Last point, for anyone trying to get stars for an open-source project: scroll-craft's star count tells you that distribution matters more than code. Without a distribution channel, at least find a spot where people are complaining but nobody has fixed it yet. This time that spot was the comment section, where someone asked "what happens when the user is on mobile."