2845 comments are all asking for the prompt: how a Douyin alt account turned a comment section into 360 stars

On the evening of Oct 21, a Douyin account with only a few thousand followers posted a 61-second video, claiming it was "Opus 5.5-style" made with DeepSeek. The comment section was almost all one line: send the prompt. Three days later, the author packaged the whole method as an Agent Skill on GitHub and posted another video saying "search Moessif/code-video-showcase on GitHub." Two weeks in, the repo has 363 stars. Nothing about it stands out on GitHub. The stars were carried over from Douyin. Below: what it is, where the stars came from, how crowded the category is, and which step to copy if you want to get stars the same way.

Comments under one video

On Oct 21 22:07, the Douyin account Moessif (4,639 followers, previously posting photography and campus map cards) posted "When I use Deepseek to generate an Opus 55-style video." The clip is 61 seconds: opening typing, 11 rounds of "1-second big title card + 3-second style segment" (flat explainer, Fourier, keynote launch, ink wash, 8-bit, Three.js clay...), an eight-cut flash sequence, and an end card, with the soundtrack landing on a 120 BPM beat. The bio reads: "This video was made with DeepSeek-V4.1 Flash."

By Nov 6, the video had 24.5k likes, 10.5k saves, 4,389 shares and 2,845 comments (Douyin, 11-06). The top comment was "the creator wants a prompt," with 280 likes and 82 replies. Scrolling down: "send the prompt," "creator, I want the prompt and the skill, please," "diehard DSH fan asking for a prompt to start my video journey," "is there a repo link?", "remotion? or hyperframes?" Some doubted it: "Isn't this just made with opus 5.5?" The second most-liked comment was "Of course it's expensive" (62 likes).

Moessif didn't invent this format. On Sep 26, the Douyin creator Mixx.ftr posted "Opus 5.5 just needs a little tweak," the same 61 seconds, 1830 frames, 19 styles and 0 video generation models, and it got 18.5k likes and 8,285 shares. Mixx.ftr only posted the video and released nothing you could install. Moessif's video followed that prompt and ran it on DeepSeek instead.

What it is

An Agent Skill for "making showcase videos with pure code, beat-synced." The frames are HTML pages drawn frame by frame with Canvas, SVG and Three.js, with Playwright opening a headless browser to screenshot each frame. The soundtrack is synthesized from scratch with Python numpy/scipy, with the drums, bass and synths all computed. FFmpeg then assembles the result into a 1920×1080, 30 fps video. No image, video or music generation model is called.

What it sells isn't capability; it's a fixed recipe: how the video is segmented (8-second opener + 11 × 4-second segments + 1-second "want more?" + 2-second eight-cut flash + 6-second end card = 61 seconds), how the big title cards slam in, that each segment's first frame must land on a beat, that every segment's frame count is checked with ffprobe after rendering, and that the full video must be exactly 1830 frames. There are also 8 color themes (added in 10-28's v1.1.0).

The README is candid: "There is no finished demo video in the repo." The preview images are frame-by-frame breakdowns of the reference video and template placeholders. To see the finished video, go to Douyin or the author's tutorial site.

Where the stars came from

(GitHub daily numbers; Douyin data 11-06)

The chain is clear: a viral video built the demand, a repo was opened to catch it, and a dedicated video told people what to search for on GitHub. The follow-up video got more saves than likes, so people were filing it away like a how-to. 10-27 and 10-28 were the fastest-growing days, right after the follow-up video.

A wave of hype helped too. On Oct 17, DeepSeek released V4.1 Flash, pitched for making video through code and tuned specifically for its own DeepSeek Harness (DSH). "DeepSeek can get Opus-level results" spreads better in China than the original Opus does: Opus is hard to use domestically, while DeepSeek is cheap and easy to pick up. The top of the README has an install line you can paste straight into an agent, and the install paths put ~/.dsh/skills first, with Claude Code and the generic .agents/skills after it.

The author made no updates after 10-28, and stars kept climbing by 15 to 25 a day. Most of this probably came from automatic inclusion in skill directories. Aggregators like awesome-claude-video-skills, agentskillshub, hotrepo and repoburst have all listed it, though how many each one brought in isn't visible. No trace of it on X, Hacker News, Reddit or tech media.

The math is grim: the two videos add up to about 32k likes and 19k saves, for just over three hundred stars, roughly 1% of the like count. People scrolling Douyin mostly don't use GitHub. The ones who actually star the repo are a small slice who already use Claude Code or DSH.

Who it's competing with

"Have an agent make video with code" has been one of the most crowded niches since late September, when Opus 5.5 launched. The awesome-claude-video-skills list alone has 180 entries.

Its position is "the official replica package for that one viral video, labeled as working with DeepSeek." On capability, coverage and star count it can't match the leaders. What it has going for it is that Douyin viewers know exactly what they're installing: the method from the video they just watched.

No one has copied this repo directly yet. Same playbook, though: Sac-Y/match-cut pulls traffic from a short video about open-source match-cut transitions, and ConardLi 11-02 posted a Douyin video pushing his own skill (3,515 likes). Also, 「卡尔的AI沃茨」10-31 posted a DeepSeek vs. Opus comparison across 29 styles, getting 2,325 likes.

What's copyable and what isn't

The one move worth learning: when people under a viral video keep asking for the prompt, that is demand; whoever turns it into something installable within two or three days gets the stars. Mixx.ftr had the original idea but didn't capture this demand. Moessif did, and treated it like engineering: a spec sheet, a timeline formula, a QA checklist and a runnable template, plus a one-line install you can paste straight into an agent. Add a model-swap contrast (a viral video from an expensive overseas model, recreated with a cheap domestic one), and the headline spreads on its own.

What's not worth copying, or can't be copied:

• The format itself. The 61-second beat-synced showcase already has an original, a replica and a comparison across 29 styles. Making another copy of the same thing won't get anyone asking for it.
• The root of the traffic. The stars depend entirely on the author's own 24.5k-like video. Without a video that goes viral, step one of this playbook doesn't happen.
• Conversion. Douyin-to-GitHub is only about 1%. Trading tens of thousands of likes for a few hundred stars isn't an efficient way to build a GitHub presence.
• Attribution risk. The credit line was added four days after the repo was created, and I couldn't find how the original creator reacted. Directly replicating someone else's viral video is easily called "reposting" in Chinese creator circles.

If it were you

Bottom line: C. Don't build yet another 61-second beat-synced video skill. That spot is already taken by the original, the replicas and a pile of video skills, and this one itself has only three hundred-plus stars. What's worth taking away is the playbook: watch comments for "send the prompt," open a repo within three days, then post a follow-up video pointing to it.

Who it's for: developers already making AI content on Douyin, RedNote or Bilibili, who have a video or two that have gone viral and who will turn their prompts into real engineering. Or someone partnering with such a creator: they make the video, you make the repo.

Not for: people who only write code on GitHub with no short-video distribution (without that video, the repo is just one more unknown among 180 video skills); and anyone hoping to beat hyperframes and Remotion head-to-head on features or quality.

How to do it: Scan the AI category daily for hit videos where the comments are full of "prompt please" or "is there a repo?" and the original author hasn't released anything. Within 48 to 72 hours, turn it into a skill: put the finished video on the first screen of the README (its biggest weakness is having no demo), add a one-line install prompt that can be pasted straight into an agent, and make it work with Claude Code, DSH and .agents/skills. The source idea belongs to someone else, so credit them on day one, and ideally ask the original author first.

How to get users: Reply with the repo name in the comments of the original video; post a separate traffic video, "search xxx on GitHub", with the install line shown clearly on screen; then submit to awesome lists and skill aggregator sites to catch the long tail.

When to move: Within a week after the hit video goes up. Moessif opened the repo on day three; after a week the comments cool off, and someone else may get there first.

Signs you're reading it wrong: If this repo gets no new video boost and stalls around 400 stars by end of November, the short-video traffic produced only one wave of stars and isn't worth a dedicated project. Conversely, if a repo using the same approach hits thousands of stars off a single video (for example after the author puts the finished demo into the README), conversion can be pushed up and the path is worth serious effort.