video-edit

Read a video as text, decide the cut as data, render it with a tested script. Never improvise raw ffmpeg for cutting; use the scripts below. scripts/... means this skill's folder + scripts/...: run them by full path from the user's working directory (don't cd into the skill folder), e.g. uv run /path/to/skills/video-edit/scripts/probe.py talk.mp4.

Rules

• Never modify, move or overwrite a source file. Everything goes to edits/<name>/ next to it.
• Always render --preview first and tell the user the path; render the final only after they agree (or they said to go straight to final).
• Print the exact output paths you produced.
• Long jobs: transcription runs at a few times realtime on CPU (small) and 15x+ on a laptop GPU (large-v3-turbo). For a 1 h file on CPU, warn the user, suggest --model base, or run it in the background.

Workflow

1. Probe - duration, fps, resolution, VFR/rotation warnings: uv run scripts/probe.py INPUT If it warns about variable frame rate (phone footage), convert to CFR first (see references/edl.md) and edit the converted file.
2. Transcribe - writes edits/<name>/transcript.words.json and transcript.txt: uv run scripts/transcribe.py INPUT [--language en] [--model small|large-v3-turbo] [--device cpu] With an NVIDIA GPU it fetches CUDA libraries through uv once and uses large-v3-turbo. --api groq / --api openai use GROQ_API_KEY / OPENAI_API_KEY instead (audio is uploaded). Use --prompt "Names, Jargon" for unusual spellings.
3. Read transcript.txt. One line per segment: [s0012 00:01:03.420-00:01:07.900] text. Segment ids are stable; refer to them in the EDL.
4. Write an EDL (edits/<name>/<edit-name>.edl.json). Minimal forms:

   {"sources": {"main": "../../talk.mp4"}, "transcripts": {"main": "transcript.words.json"},
    "drop": [{"segs": "s0004-s0006"}, {"start": "00:02:10.5", "end": 135.2, "reason": "off topic"}]}

   {"sources": {"main": "../../talk.mp4"}, "transcripts": {"main": "transcript.words.json"},
    "clips": [{"segs": "s0030-s0034"}, {"segs": "s0002"}, {"start": 61.0, "end": 64.5}]}

   clips = what to keep, in output order (may reorder, may use several sources). drop = what to remove from the clips (or from whole sources when there are no clips). Full schema, options (pad, fade, loudnorm, output size) and multi-source joins: references/edl.md.
5. Preview: uv run scripts/render.py EDL --preview -> <edit-name>.preview.mp4 (480p, fast).
6. Final: uv run scripts/render.py EDL -> <edit-name>.mp4 plus <edit-name>.words.json and <edit-name>.txt (the transcript remapped to the new timeline; use them for captions or shorts of the edit without re-transcribing).

What render guarantees

Frame-accurate cuts (piece boundaries snap to the source frame grid), a 15 ms audio fade at every cut, audio and video of identical length per piece (A/V sync never drifts), mixed resolutions/frame rates normalized to the first source (or output.width/height/fps), sources without audio get silence, optional two-pass EBU R128 loudness normalization.

Related skills

• video-cleanup: automatic silence / filler / retake removal -> an EDL for this renderer.
• video-shorts: vertical 9:16 clips with face tracking and captions.
• video-captions: SRT/VTT/styled burned-in captions and translation.
• video-ffmpeg: tested one-off ffmpeg recipes (compress, GIF, speed, music ducking...).

Troubleshooting

• ffmpeg not found: install it (winget install Gyan.FFmpeg / brew install ffmpeg / apt install ffmpeg).
• CUDA errors fall back to CPU automatically; --device cpu skips the GPU entirely.
• Bad word timings on music-heavy audio: try --no-vad or a larger --model.