One post brought in 2,600 stars: how rea-video turned "REA's video edition" into a hit

An author with 169 GitHub followers borrowed the name of 430k-star REA, wrote a README built to be forwarded as-is, then let Chinese AI-news accounts on X spread it for them. In 8 days the stars went from 0 to about 2,700. This breaks down what it is, what the post said, which slot it took, what you can copy to earn stars, and why a "video reverse-engineering" project is already too late.

33 views and 130k views

In 2026 Nov 3, author JuneYaooo posted on X: "Inspired by REA, open-sourced a skill that reverse-analyzes videos." The tweet ended with only 33 views. The same day the author opened a "open-source self-recommendation" issue on Ruan Yifeng's Tech Enthusiast's Weekly repo, which still hadn't been included by 11-06. The repo had 24 stars that day.

At noon Beijing time the next day, 12:34, an AI news account called Shlcy (@CY160909) posted:

REA-Video, an open-source project for reverse-engineering videos. Install it, hand a video to your agent, and it automatically splits out every shot and analyzes whether each frame is live footage, AI-generated, or code-driven motion graphics. ...It's like giving it a finished video and having it reverse-derive a detailed production tutorial. Free and open source, with local video analysis supported.

The post got 3,309 likes, 465 reposts, and about 130k views. A reply in English said: "Saw this and went and tried it, my video really does look more like the reference video now." The repo gained 802 stars that day, and the next day it showed about 2.7k stars and 368 forks on the GitHub page, +1,102,11-06.

(Daily numbers are morning snapshots; the last line comes from the GitHub page.)

The fork-to-star ratio is about 14%, which is usually much lower for tool repos, suggesting plenty of people actually pulled the code and ran it. The author was busy too: of the 55 commits, 11-04 were 16 and 11-05 were 22.

What it actually does

A skill for agents like Claude Code and Codex, plus a Python command-line tool. You tell the agent in one sentence:

``text Install https://github.com/JuneYaooo/rea-video, then break down this video: /path/to/video.mp4 ``

The agent installs it and runs it to completion. What comes out is:

• A one-line verdict, like "41% AI video + 59% code interface, 8 seconds of real screen recording, library music at 122 BPM";
• A shot-by-shot call on whether each shot is live footage, AI-generated, digital human, 3D, or code-driven motion graphics, based on frame rhythm, sensor noise, whether the whole frame looks like a flat plane panning, and AIGC tags in the metadata;
• Start and end times of on-screen text (exportable as SRT / ASS), edit points, whether vocals are real or AI, the music's BPM, and where sound effects land;
• A replication plan in three piles: what to shoot, what to generate with models, and what to write as Remotion code.

It runs offline without a vision model; plugging in Gemini or Doubao makes it more accurate and also reverse-prompts a prompt for each shot. The accuracy table in the README is the author's own test: 99.6% on a held-out set; 88.6% when using only the tool on videos downloaded from platforms and transcoded, rising back to 99.6% with a vision model plugged in. Total sample is 41 videos, with no third-party verification.

Why now

Three things converged.

REA had just educated the market. morluto/rea (Reverse Engineer Anything) blew up in Chinese circles in October: reverse-engineering software, reverse-engineering binaries, now 436k stars, with a pile of "what is REA anyway" videos on Zhihu, Bilibili, and YouTube. "REA can reverse-engineer software, now video can be reverse-engineered too" needs no explanation, and nearly every repost opens this way. The repo is literally named rea-video, and the last line of the README reads "Inspired by REA."

Short-video creators really have this problem. In the second half of 2026, live footage, AI generation, digital humans, and code-driven motion graphics often got mixed together in a hit. Seeing a video you want to copy, the first question is "was this shot or generated?" An early recommendation post with 24k views (@ganfanren_ai) said it plainly: "Get the call wrong and you spend the money in the wrong place." Nearly every repost mentions this point, and it's what sets it apart from other breakdown tools and what gets retold most.

The README reads like a ready-made forward. The first screen is a before-and-after GIF: on the left, a phone-shot TikTok ad; on the right, a replica with the product swapped in by AI, with the line "Left is all live footage, one scene from two angles; right is all AI, 8 shots and 3 generations is enough." Below that come 6 case studies (mixed footage, 3D explainer, product promo, AI music MV, unboxing, AI plus real-person sales), the accuracy table, versions in 9 languages, and an FAQ that draws a line with Hypit. Reposters can publish it after changing a few words.

Where the stars came from

No paid promotion found; whether those repost accounts were paid to post can't be seen from outside.

The author isn't new: across 10 repos he has about 7,200 stars in total, and earlier nihaisha-nishi-tcm (about 2,200 stars) and gpt-image2-ppt-skills (about 1,400 stars) followed the same pattern: a popular model or project's name + a specific scenario + an agent skill. The name is familiar in Chinese AI tooling circles, but the author's own social accounts have almost no reach; distribution depends entirely on others sharing it.

Who it's competing with

"Viral breakdown" is already a crowded SaaS business in China, and HeyGen was doing a "video version of REA" almost the same week. rea-video only holds one slot: open source, runs inside an agent, shot-by-shot judgment of shot/generated/code.

Is it a one-off wave?

Likely (inference). The author's earlier 2,200-star repo now gains only a dozen-odd stars a week; this repo is 8 days old with only 1 open issues, and there's no evidence yet of anyone using it long-term. A derivative repo that wraps it as a Hermes plugin (0 stars) has already appeared, and some repost accounts say "I have a better video reverse-engineering approach." Whether the heat lasts depends on whether the author keeps publishing case studies and gets picked up again by big accounts.

If it were you

B. The play is clear and copyable, but the "video reverse-engineering" slot is already taken and REA's heat is fading; what's worth learning is how it earned its stars, not building another one.

Who it suits: developers already building agent skills or Claude Code plugins who have one concrete scenario in hand; people who can write Chinese READMEs and know who the dozens of repost accounts in Chinese AI tooling circles are.

Who it doesn't suit: anyone wanting to build another "reverse-engineer video" tool (rea-video and HeyGen have taken those spots one after the other, so you can only compete on details); people who can code but can't write a one-line pitch that can be retold; anyone counting on the author's own account to bring traffic (his posts have gotten only 33 views, and yours probably will too).

How to do it: follow its formula with a different subject. When a general hit (the next REA, the next popular model) takes off, within two to three weeks build a version of it for one specific scenario, with the original project's name in the title. Pick scenarios where people see a finished product and want to know how it was made: reverse-engineering how a poster or a PPT page was designed, the structure and copy of a landing page, the editing and music of a podcast episode, the cover and title of a viral Xiaohongshu note. The core pitch has to fit in one sentence, and that sentence should decide where the user's money goes (rea-video's is "is this shot filmed, AI, or code?").

How to get users: the README is your forward script: a before-and-after GIF plus a one-line verdict on the first screen; 5+ real case studies, each with a report; an accuracy table, even if self-tested; at least four languages including Chinese, English, Japanese, and Korean; an FAQ that names a popular competitor and clarifies the split. For installation, give only one line, "Install <repo>, then …", and let the agent act as the installer. After launch, DM the repo link and a copy-ready blurb to the Chinese X accounts that recommend AI tools and the GitHub daily-digest accounts; English aggregator sites will pick it up on their own once the stars rise.

Deadline: while the borrowed project's heat is still there. Take REA: it blew up in October, and by early November people were already writing that "the heat is fading"; the next project that borrows a name should ship within two to three weeks of the original's breakout.

Signs you misjudged: ① rea-video passes 10k stars by late November and keeps getting issues and PRs, showing real demand and a long-term project rather than a one-off wave, and then there's more room to follow up by fixing what it hasn't done well (English distribution, accuracy after platform transcoding); ② Chinese AI repost accounts collectively stop reposting "X's version of Y" projects, or publicly say they take paid promotion, which cuts off the free distribution path; ③ HeyGen pushes /reverse-any-video hard as a standalone feature and adds shot-by-shot judgment of origin, in which case even rea-video itself couldn't hold that slot.

Data and sources: star and fork counts come from GitHub daily snapshots (11-03 to 11-06) and the GitHub page of 11-06; the Shlcy post and replies come from X (11-04); the author's self-recommendation comes from ruanyf/weekly issue #13053 (11-03); product features, cases, and accuracy come from rea-video's README (the author's own testing); competitor star counts come from GitHub, 11-06. Wherever it says "inference" or "unverified," that is what it means.