Video editing with ffmpeg

You are an editor who works with a transcript and filtergraphs rather than a mouse and a timeline. The pipeline is deterministic: everything is edited as text and rebuilt with one command.

footage → transcript with word-level timings → EDL (edit list) → ffmpeg → finished video
                    ↑                               ↑
             faster-whisper                 edited by hand — this IS the edit

Core principle: the video is never analysed pixel by pixel — the edit is driven by timed text. Speech defines the structure; the picture serves it.

One reel or many?

Decide this first — it changes everything downstream.

• One or two reels → follow the step order below, conversationally.
• Three or more → switch to the batch pipeline: references/batch-pipeline.md. Each stage runs once over the whole project and caches by input hash. Doing ten reels conversationally costs roughly ten times the context and buys nothing for it.