onetake: a "not a programmer" hit 3,800 stars on the fourth day of the Opus 5.5 video craze

onetake is a Claude Code skill that lets a coding agent make product launch films with HTML and code. It has one rule: no cuts, every frame has to grow out of the one before it. The repo was created in Sep 26, reached 6 stars and 3,798 forks by Nov 261. Dozens of similar skills appeared the same week, and its star count sits in the top tier. Below: why it stood out from the crowd, where the stars came from, who it's competing with, and whether you should build a similar one.

The four days after Sep 22

On Sep 22 2026, Anthropic released Claude Opus 5.5. Over the next few days, big accounts on X passed around the same idea: Deedy wrote "Opus 5.5 is incredible at instructional video generation" on Sep 23, @nutlope wrote "Opus 5.5 is incredible at launch videos" on Sep 25, and someone started a challenge: "Make a 15-second motion graphic that shows how good you are at motion design." Turning one sentence into a launch video became the thing the AI crowd wanted most that week.

On the morning of Sep 26 (UTC 05:35), Patrick Tian (@paojiaofty on X, Chinese name "泡椒肥汤圆") in Santa Fe, New Mexico, created the onetake repo. His X bio reads "Not a coder. Not a designer.", and his GitHub account was only created in December 2025. The next day, the Chinese promotion account @huano77 reposted it with the line: "Don't have Opus 5.5 but still want to make videos as stunning as Opus 5.5? You might need this."

Star gains in the first three days were +121, +359, +293, roughly 1,266 in the first week (per trendshift's daily table). On Sep 27 it made GitHub Trending for a day, ranking #24 site-wide and #8 in Python.

What it actually does

The README's opening line: "Motion films that never cut to the next slide." The Chinese version is blunter: "It's one continuous shot, not 'PPT' slides taking turns on screen."

The pain point it names will be obvious to anyone who has made AI video: AI-made motion graphics mostly come out as "title card, fade, UI, fade, feature card, fade, logo". Each shot looks fine alone, but strung together it feels like flipping through slides. onetake treats the "between two shots" as something to design: an input box expands into an app window, the window reflows into a phone, and a line crossing the desk becomes the first grid line of the next film. One camera follows the whole thing from start to finish, with no cuts.

What makes it stand out from a pile of prompt collections is a scoring script. probe.py records what's on screen frame by frame, finds each shot boundary, and asks "what survived this boundary?" The README includes a table: for the same launch film, the first version had perfect pacing but "still looked like slides" and scored 0.00 on continuity and was rejected; the second version scored 0.40 and was rejected; the third scored 0.75 and passed; onetake's own launch film scored 0.83. The same script also flags problems like shot lengths that are too uniform, no still holds, fast motion without motion blur, and subjects leaving the frame.

A few other details: the product UI is rebuilt in HTML from screenshots and placed on a camera plane, so the camera can push in indefinitely without screen recording. Motion blur is done by multi-sample shutter sampling at 180°. Drafts render at 1080p 30 and finals at 4K 60. The README includes 11 GIF-embedded finished videos, such as 12,346 phones falling like dominoes to assemble into a single image, and a dot morphing from a cursor into a menu, a loading spinner, and finally the dot over the letter i.

Where the stars came from

(trendshift daily table, GitHub daily star counts, as of 11-06)

First, it caught a platform event. The repo was created right when the Opus 5.5 video topic was at its hottest. People had just watched the big accounts' demos and wanted something they could make themselves. That was the foundation of the first-week surge.

Second, Chinese AI promotion accounts lit the fuse. The author first posted in Chinese: "Open-sourcing my motion skill... one-take continuous motion graphics." On Sep 28 he posted "A motion skill with 2 days of 600+ stars, go use it!" (later deleted). @xiaoke949's "GitHub Trending 9/28", @QingQ77, @vintcessun and @0xCheshire (a post from Oct 15 with 7,828 views) relayed it in turn. The first-week surge lines up with the timing of these reposts.

Third place: the author keeps the momentum going with a new video every few days. Sep 29 he posted 12,346 iPhone domino videos. In Oct 5 he posted "Let Opus 5.5 make the launch video for my own skill. Rejected 7 times, this is version 8," and listed out why versions 1 through 7 got rejected, one by one. Oct 9, Anthropic announced Claude Motion, available only to Team and Enterprise users. He replied under the official post the same day: "This was made with an open-source skill." After his Oct 21 post, "Holy f*ck, Claude Code just replaced Cinema 4D," stars saw a second small spike. His English posts don't get much reach individually. The Oct 5 main post only got 103 views by Nov 6 (X, 11-06). What actually spread the content were repost accounts.

Fourth, repost accounts made videos out of the GIFs. There are at least 5 YouTube videos ("This Claude Skill Makes Product Launch Films With Zero Cuts" and others). Instagram has "This free GitHub repo makes launch videos with ZERO cuts," and Facebook and Threads have them too. OSChina ran a Chinese article titled "Making blockbuster films without a video model." Several skill directory sites like skillsdirectory also list it. I can't verify how many stars each of these brought.

Inference: The author had already published a similar motion skill, motion-web (846 stars) in Sep 8. The README marks two films as "made when it was still called ohmymotion." So onetake didn't start from zero. It came with an existing follower base and existing assets.

Not found: Hacker News, Product Hunt, major Reddit threads, media coverage, or paid promotion.

Who it's competing with

Plenty. Several launched the same week:

Of the dozens launching around then, only three passed 1,000 stars: this one, motion-video-kit, and cinetic. The underlying tech is much the same (HTML rendering plus ffmpeg). What set this one apart was two things: a one-line description of how it differs from the rest ("doesn't look like a PowerPoint"), and finished films that are immediately impressive to look at. Most similar projects call themselves "cinematic launch video," and readers can't tell them apart.

But that edge is easy to copy. Huashu's huashu-motion launched in Oct 18 and already has 2.4k stars, MIT licensed for commercial use. talsiach/take-one, chacha95/demo-video-creator ("one-take launch film"), and wildsec/explainer-engine ("feel like one continuous take") all use the same pitch.

What it still gets wrong

• Not free for commercial use. The free version is under the PolyForm non-commercial license. To take client work you need the $79 Pro (commercial license, source for 10 films, 5 starter templates, notes on why rejected versions failed, and stronger auto camera moves). The README only added "buy Pro for commercial use" in Nov 5. For an open-source author who just wants stars, that's a gap, and Huashu's MIT version caught up right there.
• The community is thin. 4 open issues, no external PRs merged. Stars look mostly like "bookmarked for later," and I saw no third-party tests of users making their own films.
• The setup bar is not low. You need Claude Code, Python 3, Node, and ffmpeg, and you copy the folder into your skills directory. The source for the finished films is not in the free version.
• Momentum is fading. In November it gained only 30 to 60 stars a day, and the author's English posts can't push it further.

If it were you

One-line verdict: The "agent-made launch film" slot is already crowded. Another skill in that category will probably stall at a few hundred stars. What's worth learning is how it broke out of a crowded field: a position with a clear opponent, a quality checker that scores output, and a README built to work as a portfolio.

Who this fits: people already using a coding agent to make a specific kind of finished product (video, posters, slides, web motion, game visuals), with taste good enough to produce a few sample films that make people want to share them. Chinese speakers who can build early traction in Chinese AI promotion circles have an edge.

Who it doesn't fit: people who want to build another "AI launch film skill" to compete head-on with Huashu or video-shotcraft, or people who have prompts but no finished films.

How to do it:

1. Pick a finished-product category that isn't crowded and copy the structure: first find a "bad look" everyone recognizes (this one's is "looks like PPT") and put the opposite in your slogan. Then write a script that turns "good or bad" into a score and post before-and-after tables like "Rejected 0.00 → Passed 0.83." Candidate directions (inferred, not checked one by one for overlap): agent-made slides, data charts, App Store screenshots, landing-page hero sections, short-video subtitle layout.
2. Or fill a gap it leaves open: MIT for commercial use, the JS / Remotion ecosystem, one-line install. Note that Huashu already owns "commercially usable," so you'd need a different angle.

How to get stars: Launch within a few days of a major model release or a viral new usage, and tie the launch copy to that hotspot. Treat the README as a portfolio, with one GIF plus a breakdown per case so repost accounts can turn it into video directly. Write both Chinese and English versions and start the spark in Chinese promotion circles first. After that, post a new film every few days, and when the official product launches something similar, post a same-day comparison.

When the window was: The "code-made video" wave kicked off by Opus 5.5 has already run for six weeks, and the window for the launch-film slot has mostly closed. Other finished-product categories need to launch before the next model release, before the official products build similar features in.

Signs you misread it: A skill in your chosen category already has a thousand stars on GitHub. Or a launch with sample films stalls in double-digit stars for a week with no promoter or repost account picking it up. That means the stance isn't sharp enough or the samples aren't good enough. Change direction, don't just post more.