"From 0 to 3000 followers in 12 days": a video-subtitle puzzle Skill hit 2,859 stars on user brag posts
What it is A Skill for Codex and Claude Code. Drop in a video or YouTube link, and it picks out punchy quotes, grabs frames, and stitches the subtitles already burned into the footage into a 3:4 Xiaohongshu-style long image.
Why it took off The author kept posting success reports on X in Chinese AI circles, like "a friend reached 2k followers," "a netizen gained 12 followers in 3000 days," "roommates had a hit in a week." The top one got 377k views and brought in 4 stars over 1,444 days. Codex Skills were hot at the time, and everyone wanted to know what they could actually earn with one.
BHow you can ride it Don't build another subtitle collage. Pick another content format that has already been proven to blow up on Xiaohongshu, Douyin, or IG, and turn it into a one-sentence image-generating Skill. Prioritize English and IG versions. Run your own new accounts to build a track record first, then promote with success-report posts.
- Data (updated daily) · Repo created 2026-08-16 · 2,861 stars · 7 days +88
GitHub · stars Data 2026-10-09 → 2026-11-08 (31 days, 31 observations). Points before 10-28 come from deeply / archive backfill, not every day has one
In Oct 2026 6, Cheng Yi (a Xiaohongshu creator in her 03s studying in South Korea, X handle @ChengYi3629) posted a tweet relaying a DM from a netizen: "Teacher, I used your skill, posted three times a day with Codex, and went from 12 to 0 followers in 3000 days!" The post got 3,099 likes and 377k views, and her repo native-subtitle-quote-image gained 4 stars over the next 1,444 days. The tool itself isn't hard. What's valuable is the playbook behind it: it sells the result "AI helps you grow on Xiaohongshu and Douyin," and the Skill is just the tool that delivers that result.
What it is
An Agent Skill for Codex and Claude Code. Give it a local video or YouTube link, and it reads the transcript, picks out quotes, grabs frames at exact timestamps, and crops the subtitles that are already burned into the footage, stitching them into a 3:4 vertical long image: a main frame on top, subtitle strips below. For videos without subtitles, it can also overlay dialogue you've verified, clearly labeled as post-production subtitles.
You've definitely seen this format: celebrity interviews and movie clips, one main image with five lines of dialogue. It's been a popular format on Xiaohongshu and Douyin image posts for a long time. Before, you'd take screenshots one by one and assemble them in a collage app. Now you just tell an Agent one sentence and get the finished product.
The README front page shows finished images right away ("The Girl on the Bus," Rick Rubin, Chen Shu), in Chinese, English, and Korean. Installation is simple too: paste the GitHub link to your Agent and say "install this Skill for me," so even non-coders can use it.
How the stars grew: every surge lines up with one of the author's posts
| Timeline | Daily star gains | What the author posted on those days |
|---|---|---|
| 08-16 created → 09-08 | About 190 total | 08-17 "A friend used this skill on Xiaohongshu to grow 2k followers, top post 11k likes", 197k views |
| 09-09 ~ 09-11 | +463 | 09-09 "AI is already better than me at making Douyin... upgraded to 2.0, averaging 1k likes and saves per video", 281k views |
| 09-28 ~ 09-29 | +294 | "Everyone friends who posted with this skill basically blew up. I tested it on my new account and almost every post was a hit", 121k views |
| 10-06 ~ 10-09 | +1,444 | The line above, "12 days, 3000 fans," followed by 10-07 "two roommates started their accounts on National Day, and within a week both had hit posts" |
| 11-01 ~ 11-03 | +91 | The source isn't pinned down. During the same period, several Chinese X threads listing "N Skills for Codex video creation" were circulating (inferred). |
(Star data comes from Trendshift daily records and GitHub daily star logs; view counts come from the X post pages.)
On days without a post, daily gains dropped to single digits. On Oct 7 it hit #24 on Trendshift's all-site daily chart and #4 on the Python chart, but posts pushed it up. It wasn't the origin. On Oct 10, @xaiwind wrote an article titled "A Codex viral-video skill: 10 installs each hit a million plays, gained 10,000+ followers in 3 days," and included it. Then Zhihu, CSDN, side-hustle sites, and various Skill directories kept reposting it. Some side-hustle sites rewrote 08-17 as "13k likes on a single Xiaohongshu post."
The comments became the best ad. Under the 10-06 post, the subject herself showed up: "Now nearly 4000 followers... sticking to three posts a day, proving this approach can produce hit posts again and again." Another person said, "Honestly, this skill should really be paid," and someone else said they used an agent workflow for Douyin, "and in under two months they're nearly at 10,000 followers too."
Why this one
"Stitching video subtitles into long images" has been done before. On iOS, long-image stitching apps (one in the China store has a 62k rating), dialogue collage apps, player plugins, and scripts have all existed, but they're all manual tools with small reach. This one wins on three counts:
- From tool to pipeline. Finding the video, selecting quotes, grabbing frames, cropping subtitles, collaging, and quality-checking all run in one Agent pass, so users can make three posts a day. Collage apps can't do that.
- Real screenshots, not AI images. Native mode only crops pixels. It doesn't recognize or redraw anything, so the output looks like a true screenshot. Users trust it more, and platforms won't flag it as AI-generated. Post-production subtitles are clearly labeled and don't pass off as original speech.
- The story spreads outward in layers. First the author, then friends, then netizens who DM'd, then roommates. With each layer added, the doubt that "only the author could pull this off" fades a bit more.
The bigger backdrop is the Codex / Claude Code Skill craze in the Chinese community in the second half of 2026. Many people installed Codex without knowing what to use it for, and "AI helps you grow an account and make money" was exactly the use case they wanted to hear about.
But be clear-eyed: differentiation is necessary, but distribution is what actually drives stars. That 08-17 post had 197k views and at the time brought in fewer than 200 stars. Only after September, when the author kept pairing "who, how many days, how many followers" success reports with the repo link, did the stars climb. The same author's cy-carousel-skill (image carousels), posted in Oct 9, used the same playbook and got only 436 stars. The difference is it lacked a strong results story.
Who else is doing it
The open-source followers are all small: chaosinu7-svg/video-quote-image-i18n (7 stars, claims to add foreign-language translation, the half the original project skipped), ponyletter's batch-URL version (5 stars), and video-quote-collage-skill, which 11-05 just created, plus an Android version, subtitle-stitch. The largest adjacent project is guizang-social-card-skill by 歸藏 (7,995 stars), which lays out copy as Xiaohongshu cards without using real video frames.
Almost no one discusses it in English. Searching X in English turns up no relevant posts. The author's payoff isn't the repo itself. Starting in mid-October, his X posts began taking sponsored deals (WorkBuddy, Tencent Cloud OrcaTerm, and others), and stars and followers became blogger traffic.
If it were you
Verdict: Building another subtitle-collage Skill makes no sense. That spot is taken, and followers have only single-digit stars. What's worth copying is the recipe: find a content format already proven to blow up on Xiaohongshu, Douyin, or IG, build a Skill where an Agent produces the finished product from one sentence, then promote with success reports on "who used it, how many days, how much growth."
Good fit: People already creating content who have an account they can post from (X, Xiaohongshu, or Jike), or who have a few friends willing to test new accounts, and who can build the first batch of real results.
Bad fit: People who only want to write a tool and don't plan to use it to post content themselves. Without a results story, the Skill is just another collage script. cy-carousel's 436 stars are the example.
How to do it:
- Pick the format first, then write code. Look at Xiaohongshu and Douyin for image posts and short videos that keep blowing up, such as podcast quote cards, book-excerpt comparison images, bilingual subtitle clips for English learning, and before-and-after celebrity interview images. Choose ones that are tedious to make by hand and that an Agent can finish in one go.
- Switch language and platform. The English-language scene and IG / Threads / TikTok are largely untouched. The original author is working on Korean IG/Threads. A version with English README, English success reports, and an IG Carousel focus is the most direct gap (inferred).
- Get results yourself first. Open two or three new accounts and post consistently for two weeks, saving screenshots of follower counts and view numbers. The protagonist of the story is best not to be you.
- Install in one sentence. Give an Agent a link and it installs itself. Put finished images on the README front page so directory and tutorial sites can repost them easily.
How to get users: Post success-report threads in the Chinese AI community on X (for an English version, go to English X and Reddit creator subreddits), with screenshots in the body and the repo link in a reply or at the end. Every new result gets another post. Once a post gets pushed onto Trendshift, roundup articles and directory sites will come to you.
When to move: Skill roundup posts are still being mass-produced, so the heat is still there. But "AI account growth" Skills are multiplying, and readers will get tired. Ideally, launch before Q1 of 2027 (inferred).
Signs you've misread it: If you post two or three real results, the posts hit 100k views, and the repo still doesn't get 300 stars, then the project was riding the author's personal reputation and that week's Codex hype, and the recipe alone doesn't travel.
Research notes: competitor table, acquisition evidence, full research report
- Competitors
Name Where Launched Scale How it differs from this one Long-image stitching: Screenshot stitching apps such as "Easy Screenshot Stitch," Tailor, and Picsew iOS 2015-2018 Long-image stitching: 62,190 rating in the China store; Tailor: 8,290 rating in the US store You have to screenshot and stitch manually; no quote selection, no frame grabbing, no Agent pipeline Dialogue stitching: movie subtitle collage tools / dialogue collage / Scrollie iOS 2022-2026 0-96 rating Manual tools, very small reach 33 subtitle images / quotestitch.app Web Not verified Not verified Manual screenshot upload and stitching mpv-lines-meme-generator / PlutoVideoSnapshoter / bilibili-subtitle-stitch and similar scripts Open source 2018-2026 0-a few dozen stars Player plugins or scripts. They don't pick quotes or target Agents, and there's no account-growth story. op7418/guizang-social-card-skill Open source 2026-05-27 7,995 star Adjacent category: tools that generate Xiaohongshu image-text cards. That's layout design, not real video frames. cy-carousel-skill (same author), baoyu-skills and other Xiaohongshu image-text skills Open source 2026 cy-carousel 436 star Topic layouts or AI image generation. This project uses only real frames plus original subtitles. - Where users come from
- Verified Author's X posts (user results): 09-09 posts 2,496 likes 281k views → 3 days +463; 09-28 posts 121k views → +294; 10-06 posts 3,099 likes 377k views +10-07/10-09 replies → 4 days +1,444
- Verified X roundup long posts / KOL features: @xaiwind 10-10 "Codex viral video skill: 10 installs to a million views" features it and cites 10-06 original posts. Whether the November @DDJCXX and @xiaoying_eth video-skill roundups include it is unverified.
- Verified GitHub Trending / Trendshift: 10-07 Trendshift site-wide #24, Python #4,10-06/08, Python #8/#7. This only amplified growth.
- Verified Chinese reposts (Zhihu, CSDN, side-hustle sites, Doubao tutorials): Zhihu post "Video quotes no longer need screenshotting one by one", doutianshi/xiaoximi 09-13~19 rewrites as "Xiaohongshu single post 13k likes", xmsumi 10-24, ipd.me, several CSDN gitblog posts
- Verified Passive listings on skill directories: SkillsLLM, Agensi, Skills Directory (Grade A), TypingMind, periscale, openagentskill
- Inferred On Xiaohongshu / Douyin: Viral notes users made with it are the source of the story. Whether the platforms directly drive traffic to the repo is not confirmed.
Why chengyi-ai/native-subtitle-quote-image works
Research date 2026-11-06. Channel: GitHub. Data sources: our github library (repo_day, 2026-10-17 to 11-06), Trendshift daily star table, GitHub API, live X search (TikHub, x user / x live / x post), Jina/Exa web search, iTunes Search API.
One line
An Agent Skill for Codex / Claude Code. Give it a local video or a YouTube link, and it picks quotes, grabs frames at exact timestamps, crops the original burned-in subtitles from the frame, and stitches them into a 3:4 vertical long image. It can also overlay verified dialogue and label it as post-production subtitles. It has 2,859 stars, not because the tech is hard, but because the author, Cheng Yi (@ChengYi3629,03, studying in Korea, a Xiaohongshu blogger with 10k+ followers), kept posting on X screenshots of "used this skill to grow my account, viral posts one after another." It sells "AI helps you grow on Xiaohongshu/Douyin"; the skill is just the tool that delivers on that promise. Results screenshots
1. Growth data: four waves, each matching one of the author's X posts
| Timeline | Daily star gains | Matching triggers | Evidence |
|---|---|---|---|
| 08-16 created → 09-08 | About 190 total | 08-16 post "Started doing Korean IG/Threads, all handled by Agents" (199k views); 08-17 post "A friend grew 2 thousand followers on Xiaohongshu with this skill, top post 11k likes" (1,431 likes, 197k views) | Live X search x user ChengYi3629; star count = total minus Trendshift cumulative +09-09 since 2,667 |
| 09-09 ~ 09-11 | +463(156/231/76) | 09-09 post "AI is already better than me at Douyin… upgraded to 2.0, averaging 1 thousand saves per video" (2,496 likes, 281k views); 09-10 reposted the GitHub link (29k views) | Trendshift;X |
| 09-28 ~ 09-29 | +294(188/106) | 09-28 post "Basically everyone who posts with this skill goes viral. I tried it on a new account and nearly every post hit" (538 likes, 121k views) | Trendshift;X |
| 10-06 ~ 10-09 | +1,444(312/581/354/197) | 10-06 post "A friend used your skill, Codex posting three times a day, went from 0 to 3000 followers in 12 days, pinned post 1M views… project passed 1000 stars" (3,099 likes, 477 reposts, 377k views); 10-07 "Two roommates started in October, both had viral posts in under a week" (87k views); 10-09 "Is it this easy to get viral posts with this skill?" (62k views) | Trendshift; X. 10-07 hit Trendshift site-wide #24, Python #4 |
| After 10-10 | Single digits to about twenty per day | 10-10 @xaiwind's long post "Codex viral video skill: 10 installs to a million views, 3 days to 10k followers" features it (listed at 2.6k⭐, directly quoting 10-06 original posts) | x.com/xaiwind/article/2108866001672741199 |
| 11-01 ~ 11-03 | +91(21/52/18) | Source not pinned down: during the same period, several Chinese X roundups of "N Codex/Claude video skills" appeared (@DDJCXX 11-01 2,031 likes; @xiaoying_eth 11-04 1,340 likes). Whether this library was included has not been verified. | our GitHub repo_day |
The pattern is clear: the star curve matches the view curve of the author's "user viral results" posts. On days with no posts, daily stars drop to 1~10. It hit the GitHub Trending top 25 only on 10-07, so Trending amplified growth rather than being the source. The author's other repos, cy-carousel-skill (10-09 posts, 436 stars) and douyin-image-post-scheduler (108 stars), use the same playbook, but their "result stories" are weaker, with stars an order of magnitude lower.
2. Why users come, and where from
- Where from: Chinese AI X (main), then reposts. Repost chain: original X post → roundup long posts from @xaiwind and others → Zhihu ("Video quotes no longer need screenshotting one by one"), CSDN gitblog auto-generated articles (at least 4), side-hustle sites (doutianshi, xiaoximi, zhongxintang, Zhai Commerce Navigation; 2026-09-13~19 rewrote 08-17 posts as "Xiaohongshu single post 13k likes"), Doubao tutorials (xmsumi 10-24, ipd.me) → many skill directories list it passively (SkillsLLM, Agensi, Skills Directory "Grade A", TypingMind, periscale, openagentskill, archestack).
- Why now: In the second half of 2026, the Codex / Claude Code Skill ecosystem took off in Chinese circles, and "I installed Codex and don't know what to do with it" is a common anxiety. Meanwhile Xiaohongshu began attaching Skill distribution to notes in 2026 (The Paper, "Xiaohongshu enters Skill distribution"). "AI helps you grow an account and make money" is what this crowd most wants to hear, and this project offers a low-barrier, instantly verifiable version.
- Why people star it and stay:
- The content format has been validated: interview quotes plus original subtitle screenshots stacked vertically is a "minimal shareable unit" that has long worked in Xiaohongshu/Douyin image posts (Jennie, Mel Robbins, Codie Sanchez and others come with built-in traffic). The author offers a replicable niche, not just a tool.
- The social proof is other people's results: friends with 2 thousand followers, strangers with 12 followers in 3000 days, roommates with a viral post in a week. "It's not only the author who can pull this off."
- Installation is minimal: one line, "Help me install this Skill: <GitHub link>," and the Agent installs it itself, which is friendly to non-programmers. The README is in Chinese, English and Korean, with sample images on the homepage.
3. Competitors
| Competitors | Where | Launched | Scale | How it differs |
|---|---|---|---|---|
| Long-image stitching / StitchPics, Tailor, Picsew and similar screenshot-stitching apps | iOS | 2015~2018 | Long Image Stitch - Easy Screenshot Merge China store 62,190 rating; Tailor US store 8,290 | You still screenshot each frame by hand and stitch them. No quote selection, no frame extraction. |
| Dialogue stitch - movie subtitle collage tool | iOS | 2026-04-24 | 0 rating, ¥1/$0.99 | Manual tool with no Agent workflow |
| Dialogue collage / subtitle collage / Scrollie (dialogue collage) | iOS | 2022~2026 | 0–96 rating | Same as above, very small volume |
| 33 subtitle images (zmt.agilestudio.cn), quotestitch.app | Web | Not verified | Not verified | Manual screenshot upload and stitching |
| mpv-lines-meme-generator, PlutoVideoSnapshoter (24⭐), xjz6626/VideoSubtitleScreenshotStitcher (OCR monitoring), lintmx/bilibili-subtitle-stitch (2026-08-14) | Open source | 2018~2026 | 0~a few dozen stars | Player plugins/scripts. They don't pick quotes and don't target Agents. Nobody tells a "growing an account" story with them. |
| op7418/guizang-social-card-skill (Guizang) | Open-source Skill | 2026-05-27 | 7,995⭐ | Adjacent category: article/copy → Xiaohongshu image-text cards. That's design layout, not real video frames. |
| baoyu-skills, cy-carousel (same author), various Xiaohongshu operations skill collections | Open-source Skill | 2026 | Thousands of stars | AI-generated or laid-out image-text; this project stresses "real frames + original subtitles, no AI-generated images" |
| Built-in collage features in big-vendor apps like CapCut and Meitu | Big-company features | Not verified | — | No clear "automatic video-subtitle collage" feature found; unverified |
Bottom line: "stitching video subtitle screenshots into a long image" already exists as apps and scripts, but they are all manual tools and small in scale. Turning it into one Agent sentence that goes from a YouTube link straight to a finished image is, among open-source projects, the first and only one to reach thousands of stars.
4. Differentiation
- From "tool" to "pipeline": find the video → read the transcript → pick topics and quotes → extract frames precisely → crop subtitle strips → stitch → QA each image, all run by the Agent in one pass. Paired with the same author's douyin-image-post-scheduler, it can even schedule posts automatically. This is the core difference from every collage app.
- Original-subtitle vs script-subtitle kept strictly separate: native mode only crops pixels, with no OCR and no redrawing; script mode is clearly labeled as post-production subtitles and doesn't pass as original. This makes the output look like a real screenshot (trust, and platforms don't flag it as AI-made), and it avoids the risk of "fabricated celebrity quotes."
- Layout details: source pixels are stitched first, then scaled uniformly; the first line isn't enlarged, subtitle strips join seamlessly, and the main image takes about 70% of the frame, directly matching the Xiaohongshu viral format.
These differences are necessary conditions (without them, users' viral posts wouldn't happen), but the direct reason it took off is distribution. The same tool without those "12 days 3000 followers" results posts wouldn't have gotten 2,800 stars. The 08-17 post with 197k views at the time brought in fewer than 200 stars. The difference: after September, the author repeatedly posted the "results story" together with the repo link, and the protagonist was upgraded from "a friend" to "strangers online."
5. Visibility, copycats, and how the author monetizes
- Visibility: It's circulating in the small Chinese AI circle on X (the author's own posts have cumulatively reached millions of views, and @xaiwind and other roundups featured it; Zhihu, CSDN and side-hustle sites reposted it). English-speaking circles and the media have basically not covered it, and live English X search turns up no relevant discussion.
- Copycats (all very small): chaosinu7-svg/video-quote-image-i18n (7⭐, 08-20, openly says it "fills the half that native-subtitle-quote-image doesn't do": foreign-language translation with redrawing), ponyletter/native-subtitle-quote-image-skills (5⭐, 08-20, batch URL version), zhuyangyangm-svg/video-quote-collage-skill (11-05), EquentR/subtitle-stitch (Android app, 10-24). The bigger "copying" is on the content side: many Xiaohongshu/Douyin accounts use it to make celebrity-quote images (described in the author's posts and side-hustle sites; not verified individually).
- How the author makes money: the project is free and open source. From mid-October the author's X posts start carrying sponsored content and promotions (WorkBuddy education edition, Tencent Cloud OrcaTerm, Starryblu payments, Renoise MCP), which suggests the stars and X followers are being turned into influencer traffic monetization. A business email is also listed in the X bio.
6. What others can learn
- Sell the result, not the feature: every post follows the pattern "who, how many days, how many followers/likes" plus a screenshot, with the repo link in a reply or at the end. The star curve lines up with posts like these one-to-one.
- Package a proven content niche as a Skill: first find a content format already booming on the platform (celebrity quote subtitle images), then build an Agent tool that automates it. Users are buying the niche.
- User-story relay: self → friends → strangers online → roommates. Each layer lowers the doubt that "only the author could pull this off."
- One-line install + trilingual README + finished product images on the homepage, friendly to non-programmers, making it easy for directory and tutorial sites to repost it.
- The same playbook can be copied to other content formats: the author already reused it on cy-carousel (image carousels) but only reached 436 stars, which suggests the key variable is "is there a hard viral-result story," not the playbook itself.
7. Still unclear
- The specific source of the +91 surge between 11-01 and 11-03 (which collection post or which platform) has not been pinned down.
- User wins like "12 days, 3000 followers" and "nearly 1M views" come only from the author's own retelling and have not been independently verified. Whether there was paid seeding or a self-run account matrix can't be determined (the author 09-28 said they "tried it with a new account").
- August star-growth details: the GitHub stargazers API returned 404 for this repo, and Trendshift only starts at 09-09, so day-by-day August data is missing.
- No direct spread of this project on Xiaohongshu/Douyin was found (whether the author's own Xiaohongshu account or user notes link the repo).
Sources
- Repo and README: github.com/chengyi-ai/native-subtitle-quote-image; release v1.0.0 2026-08-16, v2.x concentrated in releases from 10-01~10-13 (GitHub API)
- Trendshift daily stars: trendshift.io/repositories/169845
- Author's X posts: x.com/ChengYi3629/status/2089217099671683254 (08-17), /2097611763391119504 (09-09), /2104426626327175385 (09-28), /2107427874731155490 (10-06), /2107727635078168606 (10-07), /2108403115074711962 (10-09)
- Collection long-form article: x.com/xaiwind/article/2108866001672741199 (10-10)
- Reposts: zhuanlan.zhihu.com/p/2092231499064014247; www.doutianshi.com/56123.html; www.xmsumi.com/detail/4446; ipd.me/articles/180; blog.csdn.net/gitblog_00237/article/details/156216434
- Directory sites: skillsllm.com/skill/native-subtitle-quote-image; www.agensi.io/skills/native-subtitle-quote-image; www.skillsdirectory.com/skills/aiskillstore-native-subtitle-…
- Competitors: iTunes Search API (2026-11-06); GitHub Search API; github.com/op7418/guizang-social-card-skill
- Xiaohongshu Skill distribution: www.thepaper.cn/newsDetail_forward_33218600