A Japanese video engineer handed FFmpeg to Claude Code: 64 days to 4,600 stars, by "measure first, cut, then check your own output"

A Japanese live-streaming video engineer with one public repo and 6,500 followers on X built ffmpeg-skill in Sep 3, letting Claude Code, Cursor and Codex edit video locally. By Nov 6 it had 4,594 stars and 369 forks. The features aren't new: several FFmpeg MCPs existed a year earlier and all stalled at a few hundred stars. This piece looks at where its twentyfold extra stars came from, and whether you can do the same for your project.

One Japanese tweet, a string of lines you can say out loud

2026 September 4, a little after 1pm Tokyo time, the X account @kebura_P posted a Japanese tweet: "I built a Skill that lets AI agents actually use FFmpeg properly." It skipped the architecture and listed five sentences first:

"Turn this video into 9:16" "Cut out the parts with no sound" "Add subtitles" "Convert iPhone HDR to SDR" "Export in a format suited to X"

Only then did it say it's more than a list of commands: it probes the source, avoids re-encoding when it can, checks the result after processing, then checks it against the delivery spec. The last two lines: it runs locally so the video never gets uploaded to the cloud, and it's for "people who want AI to edit video but find FFmpeg a hassle."

By Nov 6 the post had 78k views, 1,470 likes and 1,813 bookmarks. More bookmarks than likes means people saved it as a tool to come back to. The author is a video engineer at SEVENTHWELL, a Japanese company doing event livestreams, and had 6,494 followers on X when he posted. That same day Chinese Weibo and Sina users were relaying it, a Korean site reposted it 4 September and again 5 days later, and in Sep 7 someone wrote a long Japanese note.com post calling it a "godsend tool." The Korean outlet aifficial wrote in Sep 9: 6 days to 781 stars.

What it actually does

In one line: it splits FFmpeg into 42 Python scripts with fixed parameters, plus a manual an agent can read, so the agent stops hand-assembling ffmpeg commands.

The Why section at the top of the README was quoted almost word for word by nearly every write-up that followed:

An agent that "knows FFmpeg" still guesses: it assumes a frame rate, picks a codec the container cannot hold, re-encodes a file that only needed a stream copy, and reports "done" without opening the result.

An agent that knows FFmpeg is still guessing: guessing the frame rate, picking a codec the container can't hold, re-encoding a file that only needed stream copy, and declaring "done" without opening the output. Its fix is one fixed workflow: probe to measure real duration, frame rate, resolution, color and channels, go lossless where possible, probe again after, check against the target platform's spec, and if the picture changed, produce a contact-sheet thumbnail so the agent can look at it. The README example is one sentence, "keep the 0:45–3:10 and 5:00–6:30 parts of interview.mp4 and cut it to exactly 60 seconds for Reels". The agent runs six scripts in order and replies "final.mp4:59.98 seconds, 1080×1920,30 fps, AAC stereo".

A few engineering details back up the "reliable" claim: each script's MCP parameter definitions are auto-generated from the same argparse parser, and CI checks that they don't drift apart; it doesn't accept shell strings or filtergraphs from the caller; the MCP exposes only 12 core tools by default to save context; and CI runs against four FFmpeg versions, 6.1, 7.1, 8 and 9. Install is one line, npx ffmpeg-skill, Python standard library only, runs wherever ffmpeg exists, no account and no API key.

The stars came in waves

The stars didn't come from one day. They stacked up across three or four waves of reposts, with frequent releases keeping it visible in between.

Tom Dörr's post was one line: "Add 42 video editing tools to Claude Code, Cursor, and MCP clients with npx ffmpeg-skill." He has 220k followers and focuses on reposting new GitHub tools. The post got 103k views, 2,141 likes and 3,028 bookmarks, carrying it from the Japanese circles and skill directories into the mainstream English developer audience. On npm the package saw roughly 30k downloads in nearly 20 days, and its download peak lines up with the star peak. But npx counts a download every time it runs, and 157 versions inflate the number too, so treat it only as a direction.

The waves connect partly because the author ships absurdly fast: 64 commits, 488 PRs and 345 releases in 157 days. PR authors include claude, cursoragent and openhands-agent, so he is letting agents iterate on it. Sites like newreleases and devotd auto-generate a release page for every version, and more than a dozen skill directories such as typingmind, skillsmp, claudepluginhub and mcpmarket list it automatically, each one effectively a fresh write-up.

Why this one and not the ones from a year earlier

Agent-driven FFmpeg tools from earlier that year: video-creator/ffmpeg-mcp, egoist/ffmpeg-mcp and others, all 2025 in the first half, 89–216 stars, mostly stalled; KyaniteLabs/kinocut has nearly the same positioning, "a guarded video editing MCP", at 248 stars. The difference comes down to three things:

1. It's a skill, not just an MCP. Autumn 2026 was exactly when the Agent Skills format was taking off and "auto-cut Shorts with Claude Code" tutorials were everywhere. Every "which video skill should I install" roundup needs an FFmpeg layer for processing existing footage, and this one had the most literal name, so it got picked up by Zhihu's "180 video Skills" list and ai211's "four open-source video skills".
2. It sells "checks its own work", not "more tools". "Agents are guessing" hits anyone who has watched an agent run ffmpeg, and "no uploads, no keys" is the second hook.
3. The author had an account that could pop, and launched in his native language. 6,500 followers isn't many, but in the Japanese Claude Code scene it's a tight, relevant audience. The post skipped the architecture and led with five sentences you could say straight to an agent.

Missing any one of those three, it probably wouldn't have reached 4,600. The author also launched a dozen-plus companion libraries (subtitles, color grading, QA, remotion-ffmpeg and so on), each with only 0–18 stars. The ecosystem isn't feeding this repo stars; the repo is driving traffic to the author.

Who's competing for this spot

The big ones are all about generating video; the small ones never took off. The narrow spot of "hand FFmpeg to an agent reliably" was still open in September, and it took it with the most literal name. I didn't find anyone copying it; the same-named n0an/ffmpeg-skill was created earlier and has 15 stars.

Are the stars real?

One thing is a bit off: the author's GitHub profile has 0 followers, but the repo has only 7 watchers. Still, the fork-to-star ratio is about 8% (the median for similarly aged, similarly sized new repos is around 11%, not low), and every wave of star growth lines up with a specific tweet or batch of articles. The judgment is organic growth with low odds of star-buying, though it can't be fully ruled out: I didn't get the stargazer breakdown this time, and account quality wasn't sampled.

If it were you

Verdict: B. The playbook is worth copying, but the FFmpeg spot is already taken, and "xx-skill" is no longer empty ground: 120 days produced 876 new skill-named repos, and 45 of them have passed a thousand stars (GitHub data, 11-06). What you're now betting on is whether the tool you pick has a content wave pushing it, and whether you have a community that can make your first post take off.

Who it suits: someone who is a heavy user of an old command-line tool and knows where agents get it wrong; someone with a few thousand precise followers in a language or vertical niche; someone who can tolerate near-daily releases for two months.

Who it doesn't suit: anyone wanting to build yet another video-editing skill (the FFmpeg layer is taken, and the generation layer has HeyGen, Remotion and Browser Use); anyone hoping to win on tool count (42 tools isn't why this one took off); anyone planning to post once and not maintain it.

How to do it: Pick an old tool that agents all claim to support but keep misusing, wrap it in a script with fixed parameters, lock in one flow of "measure → do → check the result", ship a machine-readable spec, and make it installable in one line with npx or pipx. No key, no cloud. Name it "toolname-skill". 11-06 Look, these have no solid skill version yet: pandoc (only a 586-star MCP, nearly two years old), yt-dlp (MCP at 300 stars, skill at 214 stars), ImageMagick, GDAL, SoX (nothing serious in the libraries). Blender, KiCad and FreeCAD already have MCPs with thousands to tens of thousands of stars, so skip those.

How to get users: launch in your native language, don't lead with architecture, list five sentences you can say directly to an agent, and add "no uploads, no keys". Once the first post takes off, write the README's Why section so someone can copy the whole paragraph; then keep shipping a few versions a week so release aggregators and skill directories keep re-listing you; and once you hit a thousand or two stars, reach out to English accounts like Tom Dörr that specialize in new tools.

On timing: the "Claude Code video editing" wave is already cooling (weekly stars fell from 1,833 to 715). Which tool the next wave lands on is hard to predict. Watch right now for "do xx with Claude Code" tutorials and roundups, and when a category starts appearing densely, ship the matching skill within two or three weeks.

What it means if you misread it: ① If your first post gets under a few hundred bookmarks and the repo stalls under 300 stars in a week, the audience or topic is wrong. Switch tools, don't force it. ② If three or more similar skills show up for the same tool within a month, the niche is gone. ③ If Anthropic or the tool's maintainers ship their own skill (like Remotion did), indie versions of the same thing are basically done.

Data and sources: GitHub repo and README (11-06). Daily star counts come from GitHub readings (from 10-17 onward) and Trendshift. Author's first post at x.com/kebura_P/status/2095736887034212809, Tom Dörr's repost at x.com/tom_doerr/status/2115405924646985924; views, likes and bookmarks are from 11-06 readings. aifficial.net coverage 9-09. npm downloads from api.npmjs.org. Competitor and similar-skill star counts are GitHub readings as of 11-06. Where the text says "judgment" or "unclear", it's inference.