edl-edit

The edit is a JSON file

Don't hand-author a one-off filter_complex for a multi-cut edit and throw it away. Write the edit as edit.json — a list of clips with src, start, end, and a written rationale — and let one command execute it:

uv run video-agent edl edit.json -o out.mp4

Why: the edit becomes text you can read, diff, and re-render. A revision is "change two numbers and re-run," not "reconstruct the filtergraph." The rationale field forces you to write down why each cut was made (which take won, why the others lost, why the cut point sits where it does) — better decisions and an inspectable trail for the user.

Schema

{
  "fps": 60, "width": 1920, "height": 1080,        // optional (these are the defaults)
  "grade": "luts/warm.png",                         // optional LUT for the whole cut (.cube or HALD .png)
  "audio_fix": "loudnorm=I=-14:TP=-1.5:LRA=11",     // optional filter chain on the speech bus
  "seam_fade": 0.015,                               // optional; click-killing fade at each cut
  "music": { "src": "inputs/bed.mp3", "gain_db": -18, "duck": true,
             "start": 0.0, "fade_in": 0.5, "fade_out": 2.0 },   // optional music bed
  "clips": [
    { "src": "takeA.mp4", "start": 1.89, "end": 60.81,
      "first_words": "Hey everyone",                 // doc only — the cut's first words
      "candidate_takes": ["A001","A004"],            // doc only — what you considered
      "rationale": "A004 cleanest complete take: zero ums, clean ending; A001 had a 5.8s dead pause" },
    { "src": "takeB.mp4", "start": 12.0, "end": 20.0,
      "audio_lead": 0.4,                             // J-cut: sound arrives 0.4s before picture
      "punch": [1.0, 1.06],                          // slow push over the clip (static shot)
      "rationale": "answer starts under the tail of the question so the cut disappears" },
    { "src": "talk.mp4", "start": 70.0, "end": 78.0,
      "vsrc": "roomcam.mp4", "vstart": 161.3, "vend": 169.3,
      "rationale": "audio stays on the mic'd talk; cut the PICTURE to the wide cam while the slide is static" }
  ]
}

start/end are seconds (floats) on the source timeline. Cuts are frame-accurate (trim filter, not -ss seeking); the renderer quantizes every boundary onto the frame grid so times written by snap/tighten can't drift picture against sound.
first_words, candidate_takes, rationale are documentation only — the renderer ignores them; humans read them. Always fill rationale.
Every segment is normalized to width×height (scale-to-fit + pad) at fps, so clips of different resolutions/fps concat cleanly (e.g. a 4K take next to a 720p one).
audio_lead makes the cut a split edit (+ = J-cut, − = L-cut). This is the main thing that stops an edit feeling choppy — see the cutting-rhythm skill for how much to use. Video and audio are concatenated as two independent chains, so a lead never desyncs.
punch is a slow zoom push (1.06 or [1.0, 1.06]) — life for an otherwise static shot.
music mixes the bed in the same render pass (looped to cover, ducked under speech by default), so a montage doesn't need a second re-encode. audio_fix runs on the speech before the mix, so ducking triggers on normalized dialogue.
seam_fade (default 15 ms) is applied at every internal audio boundary; it's inaudible as a fade but removes the click a butt-splice makes. Set 0 to disable.

Multicam cutaways (`vsrc`)

A clip can take its audio from src but its picture from a different camera via vsrc/vstart/vend. The audio timeline stays continuous (one mic); only the video switches — a clean camera cut, no audio seam. This is how you express "screen recording with cutaways to the room cam" (see the editor skill's multi-camera section for finding the per-cutaway sync offset by audio cross-correlation).

Get cut points from the transcript, not by scrubbing

Build the EDL from text: transcribe src.mp4 --words (word timestamps) for what's said and speech-segments src.mp4 for frame-accurate silence edges. Grep the transcript for the line you want and write its start/end into the EDL. Never eyeball a timeline.

Then let snap place the cuts exactly, instead of nudging numbers by hand:

uv run video-agent snap edit.json -o edit.snapped.json --to silence   # onto silence edges
uv run video-agent snap edit.json -o edit.snapped.json \
    --to beats --ref inputs/music.mp3 --tolerance 0.4                 # onto the music grid

It prints every move it made and leaves alone any cut with no candidate inside --tolerance.

Check the pacing before you render

uv run video-agent edl edit.json -o /dev/null --report --dry-run

Prints each shot's length as a bar plus warnings for monotone shot lengths, missing split edits, and a too-short final shot — the three things that make a correct cut list watch badly. It costs nothing and it catches problems that are invisible in the JSON. Act on it via the cutting-rhythm skill. --draft renders a fast 480p version for your own verification.

Verify by re-transcribing the OUTPUT

The strongest check that the cut is right is to transcribe what you actually rendered and compare to intent:

uv run video-agent edl edit.json -o out.mp4
uv run video-agent transcribe out.mp4 --clean -o check.txt   # read it: right words, no fillers, nothing dropped

If a take was supposed to be filler-free, grep the re-transcript for um/uh. If a cut landed wrong, the output transcript will show a clipped or repeated word that a frame-check misses. Fix the offending clip's numbers in the EDL and re-run. (Internal verification only — per the no-partial-previews rule, show the user the finished video, not the checks.)

Gotchas

grade / audio_fix are raw ffmpeg applied to the assembled cut — audio_fix runs on the concatenated audio (good place for loudnorm, acompressor; see audio-edit), grade is a LUT path, either .cube or a HALD .png from grade gen-lut (see color-grade).
One re-encode. The whole EDL renders in a single filter_complex concat pass (h264_videotoolbox, aac 48k) — don't post-process with stream-copy concat afterward. That now includes the music bed and the grade, so there's no reason to add a second pass.
This LGPL ffmpeg lacks eq/drawtext — audio_fix and grade must use available filters (loudnorm/acompressor/curves/colorbalance/lut3d/haldclut), not eq.
A vsrc cutaway must match the length of the audio it covers. A cutaway swaps the picture only; if vend-vstart ≠ end-start the picture comes apart from the sound for the rest of the edit. The renderer rejects a mismatch rather than rendering it.
audio_lead needs source material to reach into — a J-cut borrows audio from before the clip's in-point, so a clip starting near 0 in its source can't take a large lead. The renderer errors instead of silently shortening. It's ignored on clip 0 (nothing precedes it).