raw-to-reel
When to use
Use this skill when the input is a pile of raw, unedited footage — phone clips, multi-camera event recordings, interview takes — and the goal is one or more finished, watchable compilation videos, built by driving ffmpeg directly rather than through Remotion or another code-video framework. If the project already involves React components, useCurrentFrame, or a Remotion composition, use frame-and-explain instead. This skill is about editing decisions and ffmpeg filter graphs on existing footage, not generating new motion graphics.
Provenance. This skill was distilled from a real multi-hour editing session producing several team-compilation videos from ~40 raw clips of a rocketry/STEM event, including the bugs hit and how they were actually root-caused — not written speculatively. Where a rule looks oddly specific, it's because it fixed a real defect.
The shape of the work
- Triage raw footage into named, identified pieces.
- Transcribe locally (Whisper, word-level timestamps) to find exact cut points.
- Trim each usable segment to its essential content.
- Chain segments together with crossfades via a small offset-calculator script.
- Overlay b-roll as context under dialogue, not as standalone transitions.
- Finish with logo, music bed, ducking, and a limiter.
- Verify per-stream integrity before calling anything done — this is not optional, see below.
- Eval the finished compilation against the house-style checklist below before considering it done — technical integrity (streams match, no clipping) is necessary but not sufficient; a compilation can pass every technical check and still have a doubled logo, a hard-cut join, or a sentence that trails off mid-thought.