| name | media-editor |
| description | Transcript-driven video post-production — set up the toolchain, transcribe with whisper.cpp, and cut/speed/export with ffmpeg. Load the matching reference for detailed commands. |
Media editing
Reusable procedures for transcript-driven video editing. The end-to-end
workflow is: inspect → set up tools → transcribe → plan from the transcript →
cut/speed → export → re-transcribe → verify.
Load the reference for the step you are on:
- Toolchain setup — install and verify ffmpeg, ffprobe, whisper.cpp, and a
whisper model (Homebrew, Nix, or apt). See
references/setup-tools.md.
- Transcription — extract 16 kHz mono audio and produce a timestamped
SRT/JSON transcript with whisper.cpp. See
references/audio-transcribe.md.
- Video editing & export — inspect with ffprobe; cut, speed-adjust
(setpts/atempo), concatenate, and export a publication-ready H.264 MP4. See
references/video-editing.md.
Core discipline (applies to every step)
- Never overwrite source recordings — write derived files (WAV, SRT, clips,
finals) alongside them or into
output/, keeping the source basename.
- Transcription and long renders take minutes — run whisper-cli and long
ffmpeg jobs as background tasks and check results on completion.
- The transcript is the source of truth for edit decisions — cite timestamps
when proposing cuts, and show the edit plan (segments, actions, estimated
output length) before rendering the final export.
- Prefer stream copy (
-c copy) for lossless cuts when keyframe alignment
allows; re-encode only when a frame-accurate cut or a filter requires it.
