Description
FFmpeg can do anything, provided you remember the commands. Turning a video vertical, cutting out the silent stretches, burning in subtitles — each one is a string of filter arguments you look up again every time, and half the time the first attempt is wrong.
ffmpeg-skill takes that memory burden away. It is an agent skill installed for Claude Code, Cursor or Codex: say "make this 9:16", "cut the silent parts", "add subtitles", and it composes the commands, runs FFmpeg and checks the result. Everything happens locally — nothing uploaded, no API key.
What deserves attention is how seriously it takes being trustworthy. Before acting it probes the machine — whether a GPU encoder exists, whether a font substitute is needed, whether this particular tool can actually run here. Every tool states plainly whether it re-encodes audio. After exporting it verifies the file really succeeded before reporting "done", and leaves no half-written junk behind on failure. The design even spells out which judgements the AI should not make on its own — anything subjective like "does this look good".
It currently ships 28 tools covering cutting, joining, silence removal, aspect-ratio fitting, pixel-exact cropping, rotation and flipping, green-screen keying and overlays, reverse playback, stabilisation, image-sequence assembly, solid-colour and gradient backgrounds, Ken Burns style photo moves, subtitles, karaoke, HDR with SDR and LUT conversion, audio clean-up, drift sync and batch rendering.
Plain-language instructions: Describe what you want and the skill translates it into FFmpeg commands and runs them, with no arguments to memorise.
Checks the environment first: Before executing it verifies the tool can actually run here, detecting GPU encoders and font substitutions, so jobs do not fail halfway through.
Lossless where possible: It avoids re-encoding when it can, and when colour conversion forces a re-encode it preserves subtitle and additional audio tracks as far as possible. Each tool declares whether it touches the audio.
Track selection: For files with several audio tracks — dubs, alternate languages — nearly every tool lets you pick which one to use, and subtitle track languages can be read out too.
Two subtitle routes: Burn them into the picture, or attach them in a way that keeps quality intact and stays switchable later; timings accept hh:mm:ss:ff timecode as well.
Verified exports: It confirms the file genuinely wrote before reporting completion, leaves no partial files on failure, and works out the cause and whether a retry makes sense.
Low-resolution previews: A downscaled version can be exported specifically for AI analysis, sparing the model a large file.
Runs locally: Everything executes on your own machine with no cloud dependency, and an MCP interface is included for other agent frameworks.
ffmpeg-skill takes that memory burden away. It is an agent skill installed for Claude Code, Cursor or Codex: say "make this 9:16", "cut the silent parts", "add subtitles", and it composes the commands, runs FFmpeg and checks the result. Everything happens locally — nothing uploaded, no API key.
What deserves attention is how seriously it takes being trustworthy. Before acting it probes the machine — whether a GPU encoder exists, whether a font substitute is needed, whether this particular tool can actually run here. Every tool states plainly whether it re-encodes audio. After exporting it verifies the file really succeeded before reporting "done", and leaves no half-written junk behind on failure. The design even spells out which judgements the AI should not make on its own — anything subjective like "does this look good".
It currently ships 28 tools covering cutting, joining, silence removal, aspect-ratio fitting, pixel-exact cropping, rotation and flipping, green-screen keying and overlays, reverse playback, stabilisation, image-sequence assembly, solid-colour and gradient backgrounds, Ken Burns style photo moves, subtitles, karaoke, HDR with SDR and LUT conversion, audio clean-up, drift sync and batch rendering.
Features
Plain-language instructions: Describe what you want and the skill translates it into FFmpeg commands and runs them, with no arguments to memorise.
Checks the environment first: Before executing it verifies the tool can actually run here, detecting GPU encoders and font substitutions, so jobs do not fail halfway through.
Lossless where possible: It avoids re-encoding when it can, and when colour conversion forces a re-encode it preserves subtitle and additional audio tracks as far as possible. Each tool declares whether it touches the audio.
Track selection: For files with several audio tracks — dubs, alternate languages — nearly every tool lets you pick which one to use, and subtitle track languages can be read out too.
Two subtitle routes: Burn them into the picture, or attach them in a way that keeps quality intact and stays switchable later; timings accept hh:mm:ss:ff timecode as well.
Verified exports: It confirms the file genuinely wrote before reporting completion, leaves no partial files on failure, and works out the cause and whether a retry makes sense.
Low-resolution previews: A downscaled version can be exported specifically for AI analysis, sparing the model a large file.
Runs locally: Everything executes on your own machine with no cloud dependency, and an MCP interface is included for other agent frameworks.
