reelbench-skills breaks video down into a shot table

reelbench-skills breaks a finished video into a per-shot table of duration, shot size, category, camera movement, framing and pacing, one row per cut. The repo reached 249 stars in a day. That pace suggests developers want cut-level data without asking a vision model to guess at numbers it can't measure precisely.

The repo includes two skills for Claude Code and Codex. video-shots produces that shot table. video-sync renders the same video beside the table so it scrolls and highlights in step with the cuts during playback.

Cut points and durations come from ffmpeg. The model only judges what ffmpeg can't: whether a shot reads as a close-up or a wide, or whether the pacing feels rushed. 15 quality gates reconcile the two before a shot table ships. The only dependencies are Node 18+ and ffmpeg, so the measurable half needs no API call at all.

Anyone building a skill on a vision model can copy the split. Give a deterministic tool anything with an exact answer, like cut timing, and save the model for calls that need judgment.

The repo doesn't say how often ffmpeg and the model disagree on a shot boundary. It also doesn't say how the 15 gates get updated when a new edge case breaks one. Until that shows up in public, the shot tables are unverified past what the project's own examples show.