Stringers
win4r/MuseAI-Skills · win4r

The day after Muse's full file set was exposed, a Chinese creator's archive repo gained 255 stars in two days, while an English repo with the same content had only 9 stars

What it is An unofficial file archive of Meta's personal agent Muse: 68 Skill source files, a connector permissions list, and evaluation scenarios. It comes with a Chinese category index and a "start with these 8" guide, and the top of the README links to the author's Bilibili channel and a Muse invite code.

Why it blew up Muse's ability to export an entire drive hit Hacker News and The Verge. Meta said this was expected behavior and that it would not be pulled. The author promoted it on X with a "Skills leaked" headline, in a post that got 85k views. Combined with his 200K followers on Bilibili and YouTube, Chinese Skill developers bookmarked it in large numbers.

CHow you could follow this The Muse window has already closed. What's worth learning is the playbook: next time a big-tech product has an "internal files are visible" incident and the company says it isn't a secret, write a categorized guide in a language nobody has covered yet within 24 hours, and create the repo and post in the same hour. The prerequisite is that you already have followers.

C grade: the content had already been absorbed by a 70k-star roundup, and Muse's window had closed. The growth came from the author's existing 200K followers. Repos with the same content but no distribution channel got only 9 to 26 stars, which is what a typical follower would get. The playbook is worth learning, but following this one isn't worth it. Circulating in small circles
Data (updated daily) · Repo created 2026-09-25 · 562 stars · 7 days +127
GitHub · stars
30040050060010-0910-1910-2911-0911-20562
Data 2026-10-09 → 2026-11-20 (43 days, 42 observations). Points before 10-28 come from deeply / archive backfill, not every day has one

win4r/MuseAI-Skills is a file archive of Meta's personal agent "Muse": 68 Skill source files, 40 connector permission lists, and a batch of evaluation scenarios and run scripts. Anyone can export these from Muse, and Meta itself has called it "expected behavior." At least one English repo posted the same material the same day and got only 9 stars. win4r's version now has 551 stars. The difference isn't the content. It's three things: moving fast, organizing it into a readable Chinese guide, and having an existing audience of 200K. Below is how it got going, and what someone without that audience can learn from it.

Sep 25 afternoon, a post titled "BIG NEWS! LEAKED!!!"

On 2026 Sep, at 25 4:22 p.m. Beijing time, the Chinese AI blogger "AI超元域" (@AISuperDomain, 5,867 followers) posted on X:

"🚀 BIG NEWS! Muse AI-related files and Skills have leaked!!! After an afternoon of digging, I obtained important files from Muse AI, which has been very popular lately, and organized them into a public repo, MuseAI-Skills. The most worth studying are the 68 independent Skills inside... Reading these files shows how an Agent distributes tasks, when it needs authorization, how it handles failures, and how it verifies work before delivery. For anyone building Agents or writing Skills, these concrete rules are worth a look."

The repo was created 37 minutes earlier. In the first reply to the post, he included his own Muse invite code.

This post is at 85k views, 489 likes and 896 bookmarks on X (11-17). Bookmarks are nearly double the likes, which suggests people are saving it as reference material to come back to. Across September 25 and 26, this repo got forked 84 times; on Sep 27 it reached #10 on a third-party GitHub daily chart, at 255 stars.

First, to be clear: this is not a leak.

Muse is Meta's personal Agent released on Sep 8, internal codename Hatch. Behind each user is an always-on cloud Linux VM.

On Sep 22, Peter James posted on his blog "I asked Meta's Muse for its filesystem and it sent me 6.8GB": he simply asked Muse to send its filesystem, and it did. The post got 359 points and 172 comments on Hacker News. On Sep 24, The Verge followed up; a Meta spokesperson said it's "like the laptop in front of you, of course you can see the files." Meta's Nat Friedman was more blunt: "those aren't secrets you dummies those are the files on the user's VM, this is intended behavior." That reply has 196k views on X (11-17).

For anyone thinking about archiving this, that means two things. First, attention has already been ignited by major media, so you don't have to build the hype yourself. Second, the company has officially described this as "intended behavior," so takedown risk for an archive is low: as of mid-November, GitHub's public DMCA takedown records show no notices from Meta or Muse.

So the word "leak" in the title isn't accurate. But the scarcity it implies clearly helped this post collect a lot of bookmarks.

What's in the repo

The raw export was a 6.8GB dump of the whole disk. win4r kept the useful parts and then did several things the raw export didn't have:

  • A skills overview table sorted into 10 categories: workflows and memory, docs and acceptance, travel booking, email and office, social messaging, shopping and finance, health… each with a one-line purpose, linked to the matching permission list and eval scenarios.
  • "Recommended reading first," 8 items: wide-research (how parallel research splits tasks and reports failures), skill-creator (how to write trigger conditions and keep the main file lean), artifacts/testing (verify "generated successfully" separately from "deliverable is usable"), goals, forget, and more, each with a line on "the design worth studying."
  • A several-thousand-character analysis in Chinese, an English README, and a SHA256 checksum list.
  • The disclaimer comes first: "It is not a complete source repo, nor an installer you can deploy with one click" and "This archive does not represent an official release or endorsement."

The pitch is "see how a big tech company writes Skills," which hits Chinese developers currently writing Skills for Claude Code and Codex. That group is the author's core audience.

The first two lines of the README have nothing to do with the content: line one is "🚀 Watch my Bilibili video," line two is "🔥🔥🔥🔥 If you haven't signed up for Muse, you can register with my invite code and get 1B tokens!" The 7 commits in Nov 4 were all about changing the Bilibili link into a bigger headline.

Same material, why only this one took off

RepoRepo createdStars (11-17)Compared with alternatives
jacobwell/muse-skills09-25, about 4 hours before win5r9Full package plus all skills, in English, no promotion
sneakers-the-rat/muse-skills09-2826Made by the earliest replicator, hourly automatic snapshots, versioned
obra/muse-skills09-2926English archive as-is, no explanation
The Muse directory in asgeirtj/system_prompts_leaksAdded 10-04Main repo 72k900 files, more complete than win4r and still updating, but buried in a big collection
win4r/MuseAI-Skills09-25 afternoon551Chinese guide + the author's own promotion

(GitHub, 11-17. Two similar repos, mikezio/muse-runtime and rallat/muse-skills, have also been 404, reason unknown.)

The earliest content isn't this one, and the most complete isn't either. jacobwell was 5 hours earlier but has only 9 stars; asgeirtj's is more complete, yet the Muse directory itself never gained traction on its own. Only two things explain the gap:

One: the author has his own channels. AI超元域 has 110k followers on Bilibili, 78.6k subscribers on YouTube, and 5.9 thousand on X, plus a blog and CSDN. His 24 repos together have over 10k stars, all following the same playbook: whichever Agent project is hot, he makes a repackage or snapshot of it. When OpenClaw was hot he made ClawTeam-OpenClaw (1,476 stars); when Claude Code's source was leaked in March he made the snapshot repo cc-notebook (913 stars). MuseAI-Skills is another run of the same playbook.

Two: he made a "textbook," not a zip file. The English repos give you raw files you have to dig through yourself. win4r gives categories, guided reading, and "start with these 8." For someone who wants to learn how to write Skills, the latter is what you bookmark and share.

Add to that speed: it went up about 15 hours after The Verge's report, before anyone in the Chinese-language community had organized the material.

How the stars came

TimelineStarsWhat happened
09-25 ~ 2684 forks in two daysRepo created; author's X post at 85k views
09-27255#10 on a third-party GitHub daily chart
From ¥09-26—On Threads someone reposted: "Someone published 68 of Muse's Skills, is it just a Lobster reskin?"; a batch of Skill directory sites auto-indexed it
10-09334Gaining 1 ~ 6 stars per day afterward
11-15436
11-16457(+21)The second wave began
11-18545+88 over two days, source not identified

(Daily star counts and fork timestamps from GitHub.)

I couldn't find anyone reposting it during the November wave. What lines up in time is something else: in Nov 14, a repo researching "Muse registration and invite-code redemption," duwqijlk/muse-protocol, went up and got 4 stars and 1,600 forks in 630 days. Chinese and Southeast Asian communities were doing bulk Muse registrations and invite-code referral chains. At least one of the 11-16 people who forked this repo had forked that one the day before. Someone searching "Muse invite code" probably clicked through to this README, which has an invite code pinned at the top (inference).

Another number worth noting: the fork-to-star ratio is 29%, far above typical repos. Most people came to keep a full copy of the package, same as with the author's Claude Code snapshot repo (913 stars, 387 forks).

For the author, this repo is a shop sign

Stars aren't the point. The repo is a landing page that funnels in two directions: the Bilibili link at the top sends people back to the video channel; the invite code earns Muse's referral reward (new Muse users can enter a friend's invite code within 48 hours of signing up to get bonus tokens). In Oct 5 he posted another test of Muse's free video generation, with 50k YouTube views, and again pinned the invite code in the description.

For him, a repo reaching 500 stars is another proof that "I can always get the hot Agent stuff first," and his followers will keep following. That is worth more than the star count itself.

If it were you

C: You can learn the playbook, but you can't repeat this one, and next time you'll need to check whether you have your own channel. This Muse opportunity is already over: the content has been absorbed by a 70k-star collection and is no longer scarce. The thing to learn is the method: when a big tech company's new product has a "internal files are visible" incident, turn the raw material into a native-language guide within 24 hours. But the similar repos without channels (jacobwell, obra, sneakers-the-rat) have already shown that content alone doesn't take off.

Who it suits: developers with several thousand to tens of thousands of followers (X, Bilibili, YouTube, and WeChat Official Accounts all count), who watch AI news daily and can read a pile of raw files and sort them in one afternoon, and who want GitHub as one entry point for their own content channel.

Who it doesn't suit: anyone with no distribution channel (the same-day English repos with the same content only got 9 stars); anyone trying to build a project whose stars keep climbing (archives like this ride one wave, then gain single digits per day); anyone planning yet another Muse archive (there's already a row of them, and the most complete one is in a big collection).

How to do it:

  1. Wait for the next event instead of chasing this one. The signal to watch: a big company just launched an Agent or AI product, someone on HN or X posts "I got it to spit out XX," and the post is climbing the charts.
  2. Confirm the official stance first. Only do it when the company says "intended behavior" or "not a secret." Don't touch real secrets or real source code: it gets taken down, and it can bring trouble.
  3. Make a guide, not a zip file. Add a category overview, a "start with these" list, and a one-line "design worth studying" per item, written in your native language; state clearly that it is unofficial and not deployable.
  4. Pick a language nobody has covered yet. This time Chinese was taken by win4r. Developer communities in Japanese, Korean, Spanish, and Portuguese are also writing Skills, and no one was organizing material for them at the time.

How to get users: build the repo and publish the post within the same hour. Frame the post as "see how a big company writes XX," and put the three or four most worth-seeing items in the body. At the same time, link the repo from your oldest, most-starred repos. Put your channel link at the top of the repo: beyond stars, followers are what you really need to accumulate.

When to move: every time, within 24 hours of the media report. After 48 hours, similar repos and big collections have mostly absorbed it.

What would show you're wrong: ① In the next "big company product's files get dug up" event, someone without an existing following reaches several hundred stars on organization quality alone. That would mean the packaging is enough and channels matter less, so this path is wider than described here. ② Conversely, if big companies start issuing takedown notices against this kind of archive, or the official side publishes the Skill source itself, the playbook closes. ③ If you publish a native-language guide within 24 hours of the event and promote it through your channels, but it gets under 100 stars in a week, that means the "look at big-company Skills" topic is no longer new.

GitHub repo ↗

Research notes: competitor table, acquisition evidence, full research report
Competitors
NameWhereLaunchedScaleHow it differs from this one
mouse.dev "I asked Meta's Muse for its filesystem and it sent me 6.8GB" (Peter James)Other2026-09-22HN 359 points, 172 commentsSource discovery; covered that it could be exported, not organized into a readable archive
The Meta/muse-agent directory in asgeirtj/system_prompts_leaksOpen source2026-10-04Main repo 71,651 stars, Muse directory 900 filesMore complete and continuously updated, but buried in a big collection, with no Chinese guide or categories
sneakers-the-rat/muse-skillsOpen source2026-09-2826 starHourly auto-snapshots by original replicator Jonny Saunders, versioned, with no packaging or promotion
obra/muse-skillsOpen source2026-09-2926 starEnglish archive as-is, no explanation
jacobwell/muse-skillsOpen source2026-09-259 starThe same full package plus 72 skills released the same day, with no promotion
mikezio/muse-runtime, rallat/muse-skillsOpen sourcelate 2026-09now 404Runtime archived per build number, deleted
CL4R1T4S / OrcaPromptVault / x1xhlol prompt collectionsOpen source2026-04 to 1054k / 343 / 145k starsOnly collects Muse Spark / Muse Code system prompts. No Skills, no runtime.
awesome-muse-skills / awesome-muse-connectors / jasonbitsmith/muse-skillsOpen source2026-09 to 1063 / 1,475 / 385 starCollects community-written skills for Muse, not Muse's own internal files
Meta Help Center: How Muse works with skillsBig-company features2026-09-08OfficialCovers usage only, does not publish SKILL.md source
Where users come from
  • Confirmed Platform event / media: HN 09-22 359 points; The Verge 09-24 coverage; Nat Friedman replied "intended behavior"; win4r repo created 09-25 07:45 UTC
  • Verified X (the author's own account): @AISuperDomain's post "Big news! Muse AI-related files and Skills have leaked!!!" has 84,860 views / 489 likes / 896 bookmarks, and the first reply carries an invite code. 09-25
  • Inferred Author's video channel funnel (Bilibili 110k followers, YouTube 78.6k subscribers): README top has two large headings: a Bilibili link and an invite code (11-07 emphasized); 10-05 Muse video-generation test video with 50k views and an invite code attached; no video specifically about this repo found
  • Verified Threads/community spread: @yrzheee 09-26 "Someone posted Muse's 68 Skills — is it really a rebranded lobster?" 797 views
  • Confirmed Auto-listed by skills indexes / rankings: skillsllm, qai.io, TypingMind, gitdiagram, ProSkillsMD #9440, claude-skill-registry; third-party github-trending-daily ranked #09-27 10; code search returns 147 references
  • Inferred Muse invite-code referral wave (second wave): 11-14 duwqijlk/muse-protocol gained 4 stars in 1,600 days; 11-16 among the same batch of forkers forked both repos; timing lines up, but no direct link found

win4r/MuseAI-Skills research report

Research date 2026-11-17 (data card refreshed to 11-18). Channel: GitHub.

One line

A repo built by Chinese AI blogger "AI超元域" (Chao Qin, GitHub win4r): a snapshot of files from Meta's personal agent Muse (internal codename Hatch) cloud VM, packaged as an unofficial archive. It contains 68 standalone SKILL.md files, 40 connector permission lists, 12 eval scenarios, plus runtime scripts and binaries. He added bilingual READMEs, a skill category table, a "read these first" list, and a Chinese analysis report. The first line of the README is his Bilibili link, the second is the Muse invite code ZLSD3V.

Takeaway: This repo took off mainly on timing plus the author's own channels. 2026-09-22 On HN, a post titled "I asked Muse to send me its filesystem and it sent 6.8GB" got 359 points. 09-24 The Verge followed up, and Meta's Nat Friedman publicly responded: "These aren't secrets, they're files on the user's VM, this is intended behavior." In other words, anyone can export these files, so takedown risk is low. win4r created the repo 09-25 (about 15 hours after the Verge piece) and posted on X under the title "🚨Big news! Muse AI-related files and Skills leaked!!!" That post got 85k views and 896 bookmarks, and within two days drove about 255 stars and 84 forks. At least 4 other English repos posted the same content in that period, and each got only single digits to 26 stars. The difference is that win4r repackaged the raw files as "Chinese-language, readable Agent workflow material" and had a ready distribution channel. From 11-16 to 18 it gained about 110 more stars; the cause is unconfirmed.

Data

  • GitHub repo (repo_day): 10-09 334 stars / 108 forks → 10-28 397 → 11-15 436 → 11-16 457 → 11-18 545 / 156 forks. From October to November it gained 1–6 per day; 11-16 gained 21,11-16 to 18 over two days, and 88. gain_7d 114, prev_7d 22.
  • Live GitHub API: 551 stars, 156 forks, watchers 1. 10 commits total: 09-25 on one day, 6 commits (creation, bilingual README, skill categories), 09-28 1 commits, 11-07 4 commits (all changing the Bilibili link into a large heading at the top of the README).
  • Fork timing (GitHub API forks list): 09-25 45 forks, 09-26 39 forks, then 0–5 per day; 11-15 5, 11-16 18, 11-17 4.
  • Third-party chart record: on koayxinchun4-cmd/github-trending-daily 09-27 it ranked #10, with 255 stars at the time.
  • Fork/star ratio of 29% is high. Many people came to save a full copy, which matches the pattern of leak-type repos. win4r's March Claude Code leak snapshot cc-notebook had 913 stars / 387 forks, and free-code had 197 stars / 252 forks, same shape.

Why it took off

1. Platform event: the Muse "filesystem export" controversy (confirmed)

  • Meta 2026-09-08 launched Muse. Every user gets a persistent cloud Linux VM.
  • 09-22, Peter James published "I asked Meta's Muse for its filesystem and it sent me 6.8GB" on mouse.dev. It hit 359 points on HN with 172 comments (hackernews table, id 49802871). On Mastodon, Jonny L. Saunders (GitHub sneakers-the-rat) said it was "extremely easy" to reproduce.
  • 09-24, The Verge published "Muse will apparently let you download its entire filesystem" (link). A Meta spokesperson said "like the laptop in front of you, of course you can see the files"; Nat Friedman replied to The Verge: "those aren't secrets you dummies those are the files on the user's VM, this is intended behavior" (X; the Verge quote page shows about 195k views). Over a dozen outlets followed, including TokenPost, The Brief, and explainx.
  • This matters for win4r in two ways: first, attention has already been ignited by major media; second, Meta officially calls it "expected behavior," so the archive is unlikely to get hit with a DMCA takedown. The github/dmca repo 2026-09 through November has no takedown notices related to Meta/Muse. But similar repos like mikezio/muse-runtime and rallat/muse-skills have now 404, reason unknown.

2. The author's own distribution (confirmed)

  • 09-25 08:22 UTC, @AISuperDomain (5,867 followers) posted under the title "🚀Big news! Muse AI-related files and Skills leaked!!!", with the repo link in the body and the invite code in the first reply. fxtwitter data: 84,860 views, 489 likes, 82 reposts, 41 replies, 8 quotes, 896 bookmarks (X). Bookmarks are nearly 2× likes, so readers treated it as reference material to come back to.
  • The author's footprint: YouTube @AIsuperdomain 78.6k subscribers, Bilibili 110k fans, X 5.9k, plus a blog at aivi.fyi and CSDN. His 24 repos total over 10k stars, and his consistent approach is riding hot agent projects: the OpenClaw family (ClawTeam-OpenClaw 1,476, openclaw-a2a-gateway 551, lossless-claw-enhanced 534), the Claude Code leak snapshot (cc-notebook 913), and video-director 491.
  • 10-05 He published a hands-on video of Muse's free video generation (50k YouTube views) and posted the invite code again in an X thread; 11-07 he turned the top of the README into a large Bilibili heading. The repo's job is to send traffic back to the video channel and earn Muse referral rewards (new Muse users can redeem a friend's invite code within 48 hours for token credits, see home/hatch/docs/referrals.md in the repo).
  • Secondary spread: on Threads, @yrzheee 09-26 wrote "Someone released 68 Muse Skills, is it a Lobster repackage?" (797 views), discussing Muse's relationship to OpenClaw. skillsllm.com, qai.io, TypingMind skills, gitdiagram, githubtree, ProSkillsMD (issue #9440 candidate list), majiayu000/claude-skill-registry and other skills indexes auto-listed it, and GitHub code search returns 4 results referencing "win147r/MuseAI-Skills".

3. Packaging: turning raw files into "Agent workflow material" (confirmed as the differentiator; whether it was the cause of the traction is inference)

  • The raw export was a 6.8GB full disk image. win4r kept only the useful parts (about 1.55GB after dedupe) and added: skills sorted into 10 categories as a table, each with a one-line purpose plus links to the permission lists and eval scenarios; a "read these first" list of 8 items (wide-research, skill-creator, artifacts/testing, goals, forget, etc.), each with a sentence on "the design worth studying"; and a several-thousand-character Chinese PROJECT_ANALYSIS plus a SHA256 checksum list.
  • The pitch is "see how a big company writes Skills: when they trigger, how they route, how authorization and failures are handled, and how delivery is checked," which hits Chinese developers currently writing Claude Code / Codex Skills. That audience is the author's core base.
  • He also explicitly states "unofficial archive, unverified source, not a deployable package," which lowers the complaint risk.

4. The 11-16 to 18 second wave (unconfirmed)

  • gained about 110 stars in two days. 11-16 of the 18 forks came from real accounts, old and new alike: Indonesia (jadilahbaik, pocongtobat008, mfaris-rch, izzdiin), Vietnam (hankbui, thuongtin), China (EastLee 北京, QingYan1012, czq90182), and the US (valgaze, registered 2012 years ago, 94 followers). These accounts all had their own normal activity over the same period, with no sign of inflated numbers.
  • Two things line up in timing, but neither has direct evidence:
    • Around 11-14, a wave of Muse bulk-registration and invite-code referral repos hit the Chinese and Southeast Asian crowd: duwqijlk/muse-protocol was created on 11-14, got 1,600 stars and 630 forks in 4 days; xyzww888/muse-JP-Reg appeared around the same time. Forker ahtungko 11-15 forked muse-protocol, then 11-16 forked this repo. People searching "Muse invite code" may land on this README, which has an invite code pinned at the top.
    • Meta Bug Bounty 11-16 is running a live hacking event in Tokyo, opening the Muse runtime cell to researchers for testing (Meta blog). This repo already has scripts for opt/hatch/runtime-cell.
  • x-radar, Jina, Exa and Brave found no large account linking this repo between 11-15 and 17.

Competitor (does the same thing: publishes Muse internal files, skills and prompts)

NameFormLaunchedScaleHow it differs
mouse.dev article (Peter James)Original discovery + blog2026-09-22HN 359 pointsThe source, which says "it exports"; not organized into a readable archive
The Meta/muse-agent directory in asgeirtj/system_prompts_leaksOpen source (big collection)10-04 added, 11-06 updatedMain repo 71,651 stars; Muse directory 900 files, including all skills, docs, SOUL/AGENTS.mdMore complete and kept up to date, but buried inside a large collection with no Chinese guide or categories
sneakers-the-rat/muse-skills(Jonny Saunders)Open source, auto-snapshot every hour09-2826 starMade by the original reproducer, can be tracked by version; no packaging
obra/muse-skills(Jesse Vincent)Open source09-2926 starEnglish, no explanation
jacobwell/muse-skillsOpen source09-25 same day9 starAlso a full zip plus 72 skills, with no distribution channel
mikezio/muse-runtime, rallat/muse-skillsOpen sourceEnd of SeptemberAlready 404Stored runtime by build number, now deleted
elder-plinius/CL4R1T4S, OrcaPromptVault, x1xhlol collectionsOpen-source prompt collectionApril to October54k / 343 / 145k starsOnly collects Muse Spark or Muse Code system prompts; no skills or runtime
aicodedecode/awesome-muse-skills, Anil-matcha/awesome-muse-connectors, jasonbitsmith/muse-skillsCommunity skills directoryLate September to October63 / 1,475 (reworked from an older repo) / 385 starsCollects community-written skills for Muse, not Muse's own internal files
Meta Help Center, "How Muse works with skills"Big-vendor documentation09-08OfficialCovers usage only, does not publish SKILL.md source

Copycats

  • siegezf/MuseAI-Skills (same name, a reworked 2025-year-old repo, 10-13 updated) plus siegezf.github.io ("Archive the Muse AI/Hatch agent runtime snapshot with 72 skill files"). It is a search-bait copy with 1 stars.
  • oldwinter/muse-skills: "Reading copy… restructured from win4r/MuseAI-Skills", 0 stars.
  • Same-name forks such as dqtx760, fiestayang, guchengod, brickhousestudios and lihouyong1919 are indexed by Brave, taking up space in search results.
  • chyroc/open-muse (124 stars, 09-29) directly uses this repo's skills as candidates for its own open-source Muse clone.

Visibility

It's circulating in small circles. The underlying event (Muse can export its entire file set) was widely discussed by The Verge, HN and dozens of other outlets, but this repo itself only spread within the Chinese AI community: the author's X post had 85k views, a Threads post had 797 views, and there are several skills indexes. No outlet wrote about it on its own. The content is already included in asgeirtj's collection with 70k stars, so "original Muse skill text" is no longer scarce.

Differentiation

The real differences are three: Chinese language; turning the raw export into a "textbook" with categories, a guide and an analysis report; and the author's own distribution channel with roughly 200k followers. An English repo with the same content on the same day got only 9–26 stars. asgeirtj's is more complete but didn't take off on its own, which suggests the content wasn't the cause. Packaging plus a channel, launched within 15 hours of The Verge's coverage, was. A few things don't count as differentiation: the title says "leak," but Meta has publicly called this expected behavior, and the repo can't actually run.

What others can borrow

  • When a big company's new product has an event where internal files can be exported or viewed, organize the raw material into a native-language guide right away (category table, reading priority, one design point per item), aimed at developers writing skills. The window is the 24 hours after media coverage.
  • First confirm the official stance (Meta says it's "expected behavior"), then archive. This keeps takedown risk low.
  • The repo is only a landing point for traffic: the top links to a video channel and a referral code, and the first comment on the X post links the referral code again. Revenue comes from the platform's invite rewards and video traffic, not from stars.
  • This approach needs a ready fan base. Similar repos without a channel (jacobwell, obra, sneakers-the-rat) show that content alone doesn't take off.

Uncertain

  • No direct evidence has been found for the source of the second wave between 11-16 and 18. The invite-code referral wave and Meta's Tokyo bug bounty event only line up in timing.
  • The stargazers API returns 404 or 403 for this machine's token, so the time of each star can't be retrieved. The first-wave distribution was inferred from fork times and third-party rankings.
  • I couldn't check whether the author made a Bilibili video about this repo: the Bilibili API was blocked by risk control, and none of the author's recent 30 YouTube videos mention it.
  • How many signups and rewards the invite code ZLSD3V actually brought in is unknown.
  • Why mikezio/muse-runtime and rallat/muse-skills disappeared is unknown (no matching notice appears in github/dmca).

This article was written by AI from public data and the data we collect daily. Numbers are sourced, and inferences are labeled.