ffmpeg-skill (kajisho5/ffmpeg-skill)Partly tested🌏 World

ffmpeg-skill wraps local FFmpeg as 42 tools, so Claude Code / Codex measure duration, loudness and aspect ratio after an edit and hand in the check results with the video

ffmpeg-skill (by kajisho5, MIT, ~3k★ on GitHub), recommended in Chinese by @ZhongRichard, is a local editing skill for Claude Code, Codex and Cursor: silence removal, cut/join, 16:9→9:16, captions/karaoke, loudness normalisation and TikTok / Reels / YouTube / X export templates—42 tools that also run as an MCP server. The point is a fixed loop—probe the footage → edit → check against the platform spec → make a contact sheet for a vision model when the picture changed—so the agent can't just say "done". On Linux we ran probe → silence removal → 9:16 crop → Reels export → check → loudness fix → re-check → contact sheet on synthetic footage straight from the repo; we did not drive it from a live agent session or test transcription.

Visit site ffmpeg-skill (kajisho5/ffmpeg-skill) ↗
Pricing
Free
Region
Global
Where it's from
Overseas
Added
Oct 27, 2026
Tested on
Oct 27, 2026
Version we used
ffmpeg-skill v2.17.2 (main @ 1c37ad1, 2026-10-26); tested with FFmpeg 7.1.5 / Python 3.13 on Debian; ~3,047★ as of 2026-10-27
Categories
VideoAutomation & Agents
Original author
kajisho5 / ffmpeg-skill ↗
Our test
Partly tested
Cost / time
The skill is free (MIT) and runs fully local—nothing uploaded. You still need your Claude Code / Codex / Cursor plan or API. Auto-transcription needs a local speech engine (Parakeet or a whisper); otherwise pass an existing transcript.
Source
X @ZhongRichard(中文推荐) ↗GitHub kajisho5/ffmpeg-skill ↗npm ffmpeg-skill ↗

The task

Hand repetitive talking-head and screen-recording chores (dead air, 16:9→9:16, loudness, captions, per-platform export) to a coding agent—and make it prove the output meets spec instead of just saying "done".

Steps

  1. Install: `npx ffmpeg-skill` (Claude Code by default, into `~/.claude/skills/ffmpeg-skill`; `--cursor`, `--codex` or `--all` for others). You need `ffmpeg`, `ffprobe` and `python3`—no API key.
  2. Health check: `npx ffmpeg-skill doctor`. Overall ok does not mean every tool works (e.g. a Homebrew ffmpeg without the subtitles filter can't caption); check per-tool `usable` in `doctor --json`.
  3. Ask in plain words and demand proof: which ranges to keep, 60 s vertical for Reels, and the check result plus a contact sheet when done.
  4. The agent should go probe.py → cut / silence / fit → export.py --preset → check.py --platform → look.py; or one `render.py clip.mp4 --template tiktok --cues cues.txt` for a platform template.
  5. Read the check table it returns: duration, aspect, resolution, fps, codec, loudness (LUFS), true peak, each PASS/FAIL. On a FAIL, have it apply the suggested fix (e.g. `loudness.py -I -14`) and re-check.
  6. Watch the result yourself: a PASS on parameters isn't a PASS on content—typos in captions or buttons hidden under platform UI need human eyes (as repliers to the source post point out).
  7. For MCP: `mcp/server.py` lists 12 core tools by default; set `FFMPEG_SKILL_MCP_FULL=1` to list all 42.

Prompt

Use ffmpeg-skill on talk.mp4: remove pauses longer than 0.6 s, reframe to 9:16, normalise to -14 LUFS and export for Reels. When done, give me the check.py --platform reels result plus a contact sheet; fix any FAIL before handing it over.

What the author says

@ZhongRichard: 42 tools for cut/join, silence removal, animated captions, aspect changes and loudness, exporting to YouTube, Reels, X templates; tool schemas are generated from the Python argparse parsers and reused for MCP so docs and real flags can't drift; the loop is probe → process → check output → contact sheet for a vision-capable agent when the picture changed. The README adds author-run numbers: 92/92 verification steps on a 10-file real-device corpus, 0 missed gaps across 20 silence cases, and 72/72 routing/honest-refusal on 24 prompts × 3 agent runs.

Our test

We did not drive it from a Claude Code / Codex session and did not test transcription or animated captions. Shallow clone (package.json v2.17.2, same as npm latest), scripts run directly on Debian's FFmpeg 7.1.5: `doctor` reported 73 capabilities available, 0 required and 2 optional missing (parakeet, whisper), and warned about a known CVE in that FFmpeg version. On a 10 s 1920×1080 synthetic clip (silent from 4–6 s) we ran probe → silence → fit --aspect 9:16 --fit crop → export --preset reels → check --platform reels → loudness -I -14 → check again → look. With FFMPEG_SKILL_MCP_FULL=1 the MCP server's tools/list returned 42 tools.

What we found when we tried it:Silence removal cut 10 s to ~8.3 s (1.7 s removed, default margins kept). fit gave 608×1080; after export it was 1080×1920, constant 30 fps, h264/aac. The first check had 1 FAIL out of 13: loudness -32.4 LUFS (target -14±2), with the fix command printed. After loudness it read -14.0 LUFS, true peak -8.6 dBTP; re-check showed 0 FAIL and 1 WARN (no subtitle track). look produced a 12-frame contact sheet. So "measure the output, fail loudly and say how to fix" holds at the script level.

How to get in

GitHub and npm are public; the README is English. The recommendation post is in Chinese on X. Runs fully local with no geo limits.

Who it suits

Creators and small teams shipping talking-heads, tutorial recordings and vertical cut-downs every week who already use Claude Code / Codex and want to offload FFmpeg grunt work without trusting a bare "done".

Pitfalls

  • All-PASS only means the specs are met, not that the edit is good or the captions are right—watch it once yourself.
  • Silence removal is volume-threshold based; with music or loud ambience, tune `--threshold` or it will under- or over-cut.
  • In our run the post-silence intermediate was flagged as variable frame rate (VFR?); only after the reels export was it constant 30 fps—don't ship intermediates.
  • `doctor` warns about known CVEs in your FFmpeg build (it flagged our 7.1.5); upgrade before processing untrusted files.
  • The install is a copy and does not auto-update; re-run `npx ffmpeg-skill` to refresh.
  • We did not test Chinese captions or Chinese filler-word removal; the README's filler lists cover de, en, es, fr, it, ja, pt.

More in this category