edl-edit

原文涵盖的主题

• The edit is a JSON file
• Schema
• Multicam cutaways (vsrc)
• Get cut points from the transcript, not by scrubbing
• Check the pacing before you render
• Verify by re-transcribing the OUTPUT
• Gotchas

Full Production Studio note: this skill drives the toolkit in agamm/video-agent. Clone that repo and run your agent inside it for the scripts/paths below to work; the guidance itself is usable anywhere. Don't hand-author a one-off filtercomplex 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: 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. 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. firstwords, candidatetakes, 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

这里只展示节选，请以完整原文和当前要求为准。