Feed in a subtitle file and out comes a "draw-as-you-talk" video: how a repo committed only once got 5,400 stars

The "hand draws as it talks" whiteboard videos in Douyin and Bilibili knowledge channels used to be built by dragging elements one by one onto a timeline in subscription software like VideoScribe or Doodly. This July, a Chinese AI blogger turned that into an Agent Skill: drop in an SRT subtitle file, get a hand-drawn MP4 out. After the repo was created, the author never touched it again, and all the stars came from other people's reposts. Below: why it took off, where the stars came from, and whether it's still worth following now.

A post with 28k views

On the morning of Aug 2026 11, the Chinese tech account @geekbb on X (Geek, who says he has about 100k followers) posted: "A Python Skill that converts SRT subtitles into whiteboard hand-drawn animation videos. Subtitles drive elements to appear in narrative order, the strokes are drawn continuously then colored, and the output is MP4, saving the manual frame-by-frame work of whiteboard animation." It came with a GIF: on warm yellowed paper, a pen first sketches a fake mountain, then a little monkey, then a big monkey snatching a banana, and finally a few kids watching.

The post has 27.9k views so far (X, 11-06). Not much, but it was the starting point. The repo geeklee/srt-whiteboard-animation was created on 7-27. For the next two weeks it got zero to five forks a day. On 8-11 it jumped to 10, and for the following week it ran 12 to 78 forks a day (GitHub fork timeline, used to back out stars). On Aug 21, a Chinese tutorial wrote "already 2k+ bookmarks on GitHub".

As of 11-06, the repo has 5,419 stars and 856 forks (GitHub). The author, geeklee, is "江哥是老登啊" on Douyin, Bilibili and WeChat Official Accounts, says he runs a one-person company, and has only 186 followers on GitHub. The repo has had no commits since the single push on creation day, and none of the 6 external PRs have been merged.

What it actually is

An MIT-licensed Agent Skill: one SKILL.md plus several Python/OpenCV scripts, installed into agents that can run scripts, like Claude Code or WorkBuddy. The workflow is:

1. Read the subtitles, split them into scenes by 25–35 second segments, and produce a storyboard;
2. Have an image model draw line art in a unified style (warm cream paper, dark gray lines, a few red-orange and blue accents, no text in the image);
3. Assign each element a region and note which subtitle line it belongs to and its appearance order;
4. Render: the pen tip glides continuously within each region, outlining first and then coloring, drawing as the narration goes, then stitching everything into an MP4.

Each step waits for human confirmation before moving on. The Tencent Cloud developer community's 9-08 post "Another god-tier video skill just blew up" put it bluntly: rival AI approaches have "strokes popping in block by block, and viewers can tell it was added in post", while this one "really moves".

Why this one took off

The format is instantly recognizable. Hand-drawn whiteboard narration is a ready-made hit format in knowledge videos. No explanation needed; one look at the GIF gets it. Reposters almost all open with the same line: looks premium, is a pain to make.

The input is something creators already have. Similar tools mostly ask for an image or a topic; this one wants an SRT, which CapCut exports in one click. "The picture draws to whatever the subtitle says" is also easy to repeat, and every repost hammers that line.

It rode the Chinese hand-drawn Skill wave. From 2026 5 through September, Chinese hand-drawn Skills went viral one after another: the static hand-drawn illustration Skill ian-xiaohei-illustrations with 13k stars, the hand-drawn style handraw-style with 7,100, and story-to-handdrawn-video with 2,400 (GitHub, 11-06). A crowd of tool accounts was looking for "the next god-tier Skill" every day, and this was the first to push hand-drawn from static illustration into moving video.

"Free alternative to VideoScribe" is an easy pitch to spread. VideoScribe costs roughly $20 to $43 a month (per G2/Capterra pricing). Tencent Cloud articles, GitHunt videos on Facebook and Bilibili reposts all used titles like "Ditch the expensive subscription".

Packaging matters more than the tech. The author 7-26 had first released two component libraries separately, one for zone masking and one for continuous strokes. They have 19 and 70 stars respectively. Combining them into a full "subtitles in, finished video out" pipeline and putting a monkey-mountain GIF on the first screen is what produced 5,400.

Where the stars came from

The author hardly promoted it himself; his Bilibili profile says "has not submitted any content yet". No ads were spotted. The stars came in waves, carried in by others:

There's another easy-to-miss channel: Tencent's WorkBuddy. Tutorials from Sumike, Zhihu, and Sohu all teach the same trick: type the repo name into the WorkBuddy chat box and let it install itself. Content creators who don't know Git can use it too, so the audience has expanded from developers to ordinary people making videos.

8-19 Who drove that peak hasn't been confirmed. It could be a big WeChat Official Account or a GitHub weekly chart (unverified).

It's slowing now: 7 days ago it was up 246, and the first 7 days it was up 419 (daily GitHub stars, 11-06), about 35 a day. It has never made GitHub Trending.

Where it falls short

• The author has stopped maintaining it. There's only one commit. An issue reports that on Mac/Linux, step 4's preview script always crashes (it can't find a Chinese font). The fix PR from August was never merged.
• How good the drawings are depends entirely on the image model. Someone in August filed an issue: "It drew a cow and it looks nothing like a cow, just circles and straight lines combined." Below it, someone replied "it depends on the image model's capability," and others are asking how to combine Doubao's models to get better results.
• Only one art style. Warm beige minimalist whiteboard. It can't do colorful illustrations or complex scenes (the Tencent Cloud article also notes this).
• No voiceover. You have to provide your own narration and subtitles first. The author made a separate srt-vox-director (123 stars) for subtitles and voiceover, but it isn't integrated yet.
• Asset licensing is unclear. 10-28 Someone asked whether the hand image in the video can be used commercially. No one replied.

Who's already following it

Copycats have come faster than the original, and most of them have filled in the missing voiceover (GitHub, 11-06):

Even earlier there was gnipbao/codex-whiteboard-video-skill (6-26,346 stars), which takes a single image and does stroke-by-stroke hand-drawing. It came a month before the original but had no subtitle timeline and never took off. On the commercial side there's Golpo AI, backed by YC: closed-source, paid, one-click output.

If it were you

Verdict: C. There's something to learn, but the whiteboard topic itself is no longer a good bet. In three months at least 8 similar projects appeared, the best one picked up 1,000 stars in a month, and someone has already built the full "topic → voiceover → hand-drawn video" pipeline. The original is still gaining, but slower. What's actually valuable is the playbook, not the topic.

Who it suits: developers already active in the Chinese Skill scene who know what tool accounts like to repost; people who can make a good demo GIF themselves; people who want to carry the same playbook over to other video formats or other languages.

Who it doesn't suit: anyone planning a tenth "whiteboard hand-drawn video Skill." Unless you clearly outperform the existing ones, repost accounts have already run their round and won't post about a near-identical tool again. It's also not for anyone counting on organic GitHub traffic: it has never made Trending, and its stars all came from reposts.

How to do it, ranked by how worthwhile:

1. Switch to another video format and copy the whole playbook: chalkboard chalk, sticky-note walls, handwritten formula derivations, drawing routes on maps. Take the same input creators already have (subtitles, scripts, slides), build a full Skill that goes "one file in, one finished video out," and put a GIF on the first screen that anyone can understand at a glance and forward directly.
2. Build a whiteboard version that actually works: runs cross-platform without crashing, includes voiceover, offers several art styles, has clean licensing for the hand assets, and clearly states "I fixed what breaks on Mac in the original." This bets that your results are good enough to get repost accounts to write about it again.
3. English or Japanese/Korean markets: Japan has people selling paid translated versions, which shows demand. In English, hi-nikola has only 459 stars and no clear benchmark has emerged yet. But the English world has no Chinese-style network of tool accounts relaying each other, so you'd have to find AI-tool creators on YouTube and X yourself.

How to acquire users: start with a big Chinese tech account (like @geekbb), with a title like "Free alternative to a paid software"; at the same time, publish a tutorial on "installing with one sentence in WorkBuddy" so non-coders can use it; then let tool blogs and Skill directories pick it up automatically.

When is "before"? Before the 2027 Spring Festival. The Skill craze has run from May until now, and tool accounts' appetite is still there, but titles like "yet another amazing Skill" have already been used a lot (inference).

What it means if you misread the signal: ① The new format you build gets no creator to share it within two weeks. That means the wave has passed, or the demo isn't clear at a glance. ② The original, or some copycat, starts actively maintaining it, adding voiceover and art styles, and steadily gains hundreds a day. Then the whiteboard niche's #2 path is blocked. ③ Conversely, if 11 or December brings in new whiteboard Skills that each hit 2,000+ stars within a month, this signal was underrated and the niche isn't crowded yet.

Data and sources

GitHub: geeklee/srt-whiteboard-animation repo, issues, PRs, and fork timeline; daily stars (10-14 to 11-06); star counts for similar repos (11-06). X: @geekbb 8-11 posts (27.9k views, 11-06), @jglcn1991 8-26 posts, @GitHub_Daily 10-16 posts. Articles: Tencent Cloud Developer Community 9-08, Zhihu 10-28, Sumike 8-23, Shuyi's Newsletter 9-04, note.com Japanese version 8-13. VideoScribe pricing: G2/Capterra. Places marked "inference" or "unverified" are inference or unverified.