How a Douyin creator with 25k followers turned his on-camera know-how into an open-source Skill with 500 stars

"Taishang Laojun · AI Alchemy Furnace" is an AI skill pack for writing short-video scripts. It works in Claude Code, Codex, or by pasting the prompt into Doubao. The repo was created in Oct 22 and reached 495 stars and 76 forks within 15 days (GitHub, 11-06). Almost all the stars came from the author's own Douyin account, "柳爷不迷糊": a launch video with 13k likes and 14k saves, then a step-by-step tutorial the next day. Nobody on GitHub promoted it, and no X or WeChat account wrote about it. Below: what it is, where the stars came from, how crowded the category is, and whether you should build one too.

A "0-star GitHub bombshell"

On the night of Oct 26, 2026 at 9:03, Douyin AI creator "柳爷不迷糊" posted a video titled "GitHub bombshell, for video creators." The video script was almost the README word for word: "This skill comes from Liu Ye's in-depth review of works he has run that reached over 10 million views."

The comments weren't so kind. The most-liked one (166 likes):

The most AI-sounding line, "Bros, something big just happened"

Another: "Mostly 0 stars of big news, first time I've seen this." Someone else said "Something big happened, a blogger released a '王婆卖瓜' skill." The repo in the video really did have only single-digit stars back then (inference: the 10-27 snapshot showed just 20 stars).

But this video was the best this account had posted up to that point: 13,043 likes, 14,043 saves, 3,712 shares (Douyin, 11-06). Saves outnumbered likes, which means people were "saving it to come back to later." The next day, 10-28 morning, he posted "Taishang Laojun AI Alchemy Furnace: Step-by-Step Guide, From Install to Draft, All in One," with 2,949 likes and 2,717 saves, walking viewers through typing "/" in Doubao to select the skill. The GitHub stars jumped from 20 to 217 that same day.

On Oct 31 he posted an "upgraded version" that opened with "Bros, look what I found, a real treasure," and halfway through revealed the trick himself: "Why do I know so much? Because I'm the author," ending with "Let me know in the comments, and give it a star." This one only got 345 likes, and the star count grew by only thirty or so over those two days.

What it actually does

The README's first line: "Don't know what to film, can't write scripts? Toss it to your AI." It has seven functions: finding benchmark accounts and breaking down viral hits, topic selection, series planning, openings and structure, full drafts, titles, and monetization hooks, plus one step to "remove AI flavor."

The README explains how it differs from the AI copywriting built into Doubao and CapCut using an example about selling free-range eggs. The user gave an AI first draft: "In this fast-paced era, it's getting harder and harder to find a truly good egg... carrying a deep sense of nostalgia and craftsmanship," then added: "My home is in the mountains, the chickens are free-range, about three hundred of them, and every morning I go collect eggs from the coop." After the Alchemy Furnace revised it, it became this:

(Crouching by the henhouse, holding an egg just collected) There are so many free-range egg sellers online, you're probably wondering where they get so many real free-range chickens... Every morning I come here to collect the eggs. Look, this morning's, still warm. There are just this many chickens, and just this many eggs.

It also explains why the changes were made: the first thing people think when they see it is "how could there really be that many online?", so the opening states it outright; "rich in nutrition" you never tested, so it gets cut.

The rules are few: the AI gives its judgment first and asks only one question that would change the conclusion; every suggestion has to use your own specifics, since anything that still works for someone else is generic filler; on first use it chats for a few rounds and compiles a "User Profile."

The install method is friendly to people who don't write code: a single npx line installs it for Claude Code / Codex, and there's also a "Bot version prompt" you can download and paste into Doubao or WorkBuddy. The most common question in the Douyin comments was "Is it on WorkBuddy?" "Send me the link" "@Doubao send me the address for this skill."

Where the stars came from

(GitHub, daily snapshots; Douyin figure 11-06)

The curve matches the videos: the back-to-back release and tutorial brought in most of the stars, and later videos couldn't move it. The last code commit was on 10-28 (v0.7.2), with no updates since.

Actual users are fewer than the stars suggest. Installs via npx skills add tracked on skills.sh total about 2,200 across 8 sub-skills, and the main skill about 280 (11-06); more people likely downloaded the Bot prompt and pasted it into Doubao, which can't be counted. Forks are 15% of stars, high for a new repo of this size, which looks like a "download and save it" action (inference).

Why him: this account had already built an audience by explaining how to use Codex, Grok, and Opus (10-12, one video with 4,619 likes, 10-23, one with 6,460 likes), with 24.6k followers. The audience was already people who wanted to use AI for content and side hustles. The product's target users are the same crowd as his fans. Notably, 11-03 he posted an ESP32 hardware video unrelated to the Alchemy Furnace that got 17.5k likes, more than the release video. The account is still growing, but the Alchemy Furnace is no longer mentioned.

Who came before

"Chinese content-creation skills" on GitHub were already crowded. The Alchemy Furnace's own README says its structure references dbskill and its anti-AI-flavor step references Humanizer-zh.

The Alchemy Furnace sits mid-to-low in the ranking. story-studio and digital-human-koubo, created the same day and also gaining traction that same week, follow the same playbook: a creator turns their own method into a skill, then promotes it on their own account. On GitHub, searching "Alchemy Furnace" or "Liu Ye" only turns up this one, and nobody has directly copied it.

What it got right

• Putting the blogger himself into an AI. The selling point isn't "lots of features," it's "a Douyin blogger who still posts regularly wrote his own viral-hit breakdowns into it." For his fans, that's more credible than any generic template.
• Release + next-day tutorial. The first video made people save it, and the second caught the ones who couldn't install it. In the first four days after the two videos, stars went from 20 to 299, accounting for sixty percent of the total so far.
• Two entry points. Those comfortable with the command line use npx; those who aren't download the Bot prompt and paste it into Doubao. Most Douyin viewers are the latter.
• A before-and-after passage on the README's first screen. The egg example lets you see what it changed at a glance, without installing anything.

It also has clear weak spots: comments quickly saw through the "discovering it myself" trick, and comments like "王婆卖瓜" and "0 stars of big news" rank near the top. After the release it stopped updating, so star growth depends only on the blogger's next video.

If it were you

C: You can learn the playbook, but building another one in this category isn't worth it. The top Chinese creation skills already have 10k to 20k stars, and three or four similar ones took off the same week. The Alchemy Furnace's 500 stars came from the author's own Douyin account with 25k followers and one video with 13k likes, which copycats can't get. Without your own audience, building a "better talking-head skill" will most likely stall at a few dozen stars.

Who it suits: people who already run a content account with tens of thousands of followers and have a method they can explain clearly (product-selling talking-head scripts, Xiaohongshu image-and-text posts, livestream scripts, lead generation for a specific vertical). For you, an open-source skill is a "save it, claim it" hook that earns GitHub reputation and a more loyal following.

Who it doesn't suit: developers without a distribution channel who hope to win stars by making something "more complete" on GitHub, and anyone trying to go head-to-head with dbskill or Humanizer-zh on talking-head copywriting.

How to do it: don't build a generic "AI copywriting" tool. Learn from it to write a specific person's judgment method into the skill, and put a real before-and-after passage on the README's first screen. You'll have a better shot in a scenario it doesn't cover: livestream scripts, Xiaohongshu image-and-text posts, the senior-audience vertical on WeChat Channels, or lead-gen talking heads for a specific industry (renovation, lawyers, tutoring). Offer both npx and "paste into Doubao and it works" as entry points.

How to get users: post on your own platform, releasing a video and then a next-day step-by-step guide back to back. Don't copy its "pretending I stumbled on a treasure" move; at 0 stars that gets called out in the comments. Saying "I made this, the method is all in there" works better. Pin the repo link in the comments so nobody has to ask "send me the address."

When to move: this wave of "bloggers turning their methods into skills" only started a few weeks ago, and similar projects have popped up by the dozen in a month. If you want to follow, do it now while Doubao, WorkBuddy, and other chat tools that can install skills are still acquiring users, and finish before year-end.

What would show we got it wrong: (1) a Chinese creation skill with no blogger backing it organically passes 1,000 stars before year-end, which would mean the category has organic GitHub traffic of its own, not just audience transfer; (2) the Alchemy Furnace keeps gaining 100+ stars a week even after Liu Ye stops mentioning it, which would mean the product is spreading by word of mouth and its "anti-generic" pitch is worth more than it looks; (3) Doubao or Douyin officially launches a distribution entry for third-party skills, so creators without an audience also get a channel, and the grade should be raised.