The follow-up play for a hit: how video-talkcraft rode on its big sibling's 17k stars to reach two thousand

Vincent, who makes the Claude Code video skill, first got video-shotcraft to 17k stars. A month later he applied the same method to talking-head explainer videos, named it video-talkcraft, and the first line of the README reads "the talking-video edition of the video-shotcraft series." It went from zero to 2,118 stars in 76 days. For anyone trying to get stars on an open-source project, the lesson to take is the "series" move. But first note that most of those two thousand stars came from the older sibling.

Starting from "looks like PPT"

In Jul 2026 20, the author Vincent (@VincentWei93 on X) posted: "I recently tried the Remotion plugin with Claude Code, and video motion graphics are basically at PPT-animation level. No camera movement, no transitions, no sound effects." So he open-sourced video-shotcraft, a skill that lets an agent make product promo videos. By his own later account, it reached 5k stars in under a month, and by Nov 6 it was 17k.

In Aug 22, he built video-talkcraft. A promo video is "visuals matched to beats"; a talking-head video is "visuals matched to the voice." Same pain point, different scenario. On Sep 1, Jason Zhu (@GoSailGlobal), who makes overseas content, posted: "For talking-head explainer videos, the hardest and most time-consuming motion graphics part, this skill covers for you… 208 stars," and that post got 4,502 views (on X, 11-06 views). Three days later, GitHubDaily posted something more important:

You record a voiceover, but the visuals can only show PPT-style page-flip captions, and the result still looks like someone reading a script. Yesterday's video-shotcraft is for product promo videos. This one is its sibling for talking-head videos, aimed at the "PPT feel" of explainers.

"Yesterday's… this one is the sibling…" The same author being pushed by a big account two days in a row means the second day's readers are already waiting.

What it actually does

You give it a script and a recorded voiceover. On your machine it does three things. First, it aligns every word in the script to a timestamp in the audio (the author tested 110 seconds of mixed Chinese-English narration, with median error of 20–40 ms). Then it writes a shot list saying what visual each sentence should get. Finally it renders the video with Remotion: animated text cards, screenshots, camera moves, subtitles and sound effects, each landing on the moment that word is spoken.

Around this pipeline, the author added a few things:

• 108 motion-graphics recipe cards, each with its use, parameters, known pitfalls, and copy-ready source code, plus a one-page online gallery. The count grew over time: 78 cards when Jason Zhu posted, 79 when GitHubDaily posted, and 108 by Sep 5 after porting 19 from shotcraft.
• The "anti-PPT" rule: every shot gets a very slow push or pull camera move, and any frame left static gets flagged automatically by the script.
• A CapCut-style workbench: the finished video can be broken into subtitle, transition, voiceover and other tracks for editing item by item. Individual effects can be exported as transparent video and dragged into CapCut or Premiere as overlays.

The bar is not low: it needs Node, Python, ffmpeg, an alignment model of 700 MB or more, and about 760 MB of dependencies. The license is PolyForm Noncommercial: free for personal, educational and research use, not for commercial use. Jason Zhu specifically flagged this in his post. Some people couldn't get it working either: under the GitHubDaily post, a reply said that after feeding it subtitle files, audio and design screenshots, "most of the output doesn't match the data in the screenshots at all."

How the stars climbed

(The September figures come from the original posts and from Reveneau and Trendshift; from 9-29 on, the numbers are GitHub's daily star counts.)

Three things stand out. First, the launch week came from several Chinese AI accounts on X reposting together from Sep 1 to 4. Second, the steady growth afterward came from dozens of Skill directory sites (mcpmarket, skills.sh, skillsdirectory and the like) auto-indexing it, plus Chinese blog breakdowns. It hit the HTML language ranking on Trendshift several times (#12 to #21) but never made GitHub's global trending. Third, the biggest single-day jump at the end of October came from the author's whole lineup being lifted together: mg-styles-15 was +265 in one day on 10-28, and video-shotcraft was +462 in one day on 10-31. talkcraft rode that wave; it wasn't rediscovered on its own.

It has 212 forks, about 10% of its stars, which is higher than typical spectator repos. But scrolling through them, most look like straight copies, and it's impossible to tell how many people actually modified and used it. After Sep 22 the code has essentially stopped changing. Both October commits only updated the WeChat group QR code in the README, and the author's energy moved to anything2explainer and mg-styles-15.

How crowded this niche is

"Making videos with agent-written code" is already a hot category this year, and talkcraft is a latecomer:

Below that is a string of Chinese "talking head + Remotion" mini-repos with a few dozen stars each. Notably, two launched in late October: digital-human-koubo reached 1,800 stars in about ten days, and claude-video-effects gained over 700 in two weeks. This category still produces new projects every month that can reach hundreds to two thousand stars. Each one just has to find a niche nobody else owns.

Why it got reposted

Broken down, talkcraft's stars came from two sources.

The entry ticket was earned: "Cures the PPT feel of talking-head videos" is a line that's easy to say; word-level alignment is backed by measured numbers; and 108 cards plus an online gallery give reposters images and videos to clip. The few posts that reposted it used almost exactly the README wording and gallery material. The author effectively wrote the repost copy in advance.

The volume was borrowed: without shotcraft's 17k stars and the author's followers in the Chinese AI community on X, GitHubDaily wouldn't have written "yesterday… today's sibling," and directory sites wouldn't have indexed it so fast. The author's profile lists four video skills side by side under "Build in public," and each README links to the others. This is a lineup, and talkcraft is its second entry.

If it were you

One judgment: the path of "Claude Code + Remotion for a specific kind of video" still has room. Following talkcraft's method, pick a video scenario nobody has covered and you can reach a few hundred to a thousand or two. But don't expect to copy its two thousand stars. More than half came from its sibling repo and the author's own traffic. Either have a project that already brings people with it, or go find accounts to repost you yourself.

Who it's for: developers who can write Remotion or front-end animation, who are actually making videos (talking heads, tutorials, product intros), and who know which step eats the most time. Anyone who already has a project with a few thousand stars is especially well placed, since they can copy the "series" playbook directly.

Who it's not for: people who have never touched video and just want "another general-purpose agent video skill." That slot is already filled by hyperframes, Remotion's official skills, and effectcraft. It's also not for anyone trying to earn money from it: this kind of project is after stars and reputation.

How to do it:

• Narrow scope wins, and it should fit in one sentence. talkcraft handles "scripted talking head with existing voiceover", and the ones getting traction in that space are all narrow: cutting filler words, turning one photo into a digital presenter, adding effects to real-person video. Still unclaimed as far as I can see: English-language talking-head tools (talkcraft's alignment, docs and community are all Chinese-first), vertical short-video titles and on-screen text styling, screen-recording tutorials with auto callouts and zoom, podcast clips with auto-generated visuals, and data-journalism chart animation (my inference, based on the repos I could find on 11-06).
• Things to do better than it does: install friction (one command, no multi-hundred-MB model download, support for cloud alignment), commercial licensing (it's non-commercial, so MIT or Apache is itself a selling point), and the mismatch with real material (the "doesn't match" reply above is the demand).
• Turn the method into a browsable asset: recipe cards plus a one-page online gallery are easier to screenshot and repost than a single prompt.
• Put a gallery screenshot and the finished video in the first screen of the README, write a first paragraph on "what you give it and what it gives back," and attach a set of before-and-after measured numbers.
• When one project takes off, immediately split out a "second in the series" by scenario, cross-link the READMEs, and let the accounts that reposted you write "yesterday's one, today's sibling."

How to get users: On the Chinese side, target daily open-source roundup accounts like GitHubDaily and Jason Zhu. talkcraft's traffic spike came entirely from the four days they shared it. On the English side, pitch AI newsletters like AlphaSignal. After launch, submit the repo to Skill directories like skills.sh and mcpmarket. They drive a long tail of a dozen-plus visits a day for months. Put a WeChat group QR code in the README to keep people who can't get it installed.

When: before: This category is still growing, with new projects appearing every month. But niche scenarios are getting claimed one by one, so ideally you lock in your chosen scenario before early 2027.

Signs you've got it wrong: ① Remotion or Anthropic builds "talking-head / subtitle motion" into an official skill, which weakens the case for a narrow-scenario skill. ② New narrow-scenario repos in the same category launched over the next month or two mostly stall at a few hundred stars, meaning the category's heat has passed. ③ The author's next repo in the same series can't reach a thousand stars, meaning the multi-repo playbook is failing too.

Data and sources

GitHub repo page, README and author profile (11-06); GitHub daily star records (from 9-29); Trendshift repo page; X: @VincentWei93 7-20 tweet, @GoSailGlobal 9-01 tweet and view count, @GitHub_Daily 9-04 tweet and replies; AlphaSignal 9-02 feature; Reveneau listing page. Competitor star counts are as of 11-06. Where we write "about," the figure is rounded.