Open-Source ReelBench Automates Video Shot Analysis With Claude Code
A set of Claude Code and Codex skills that automate shot-by-shot film analysis, using ffmpeg for hard numbers and the model only for judgment calls.
Takeaways
- reelbench-skills bundles Claude Code and Codex skills for automated shot-by-shot video analysis, over 1k stars in a month
- video-shots breaks finished films into tables of duration, shot size, camera move and frame description
- ffmpeg measures all cuts and timings so the model never hallucinates numbers, only labels semantics
- 14 quality gates audit every call; demo shows 53 shots, 3.83s average, all gates green
- video-sync composites shot data alongside footage; video-scrub strips GPS and telemetry metadata from AI clips
- No npm deps, no API keys, runs on existing session quota: github.com/eternityspring/reelbench-skills
Analyzing a finished video shot by shot usually means scrubbing the timeline, recording cut points, and labeling camera work by hand. The open-source reelbench-skills project automates that workflow through Claude Code or Codex, producing structured data and an interactive report. The repository currently has 8 stars.
The project divides the work between ffmpeg and a multimodal model. ffmpeg measures shot boundaries and durations, while the model labels shot size, category, camera movement, and visible content. Fourteen validation gates check the resulting analysis before the report is generated.
Three jobs, one pipeline
| Skill | Purpose | Output |
|---|---|---|
video-shots |
Breaks a finished video into shots and labels each segment. | Timing, shot size, category, camera movement, frame description, keyframes, and validation results. |
video-sync |
Composites the analysis beside the source footage. | A synchronized review video whose shot data updates at each cut. |
video-scrub |
Removes metadata from generated clips before publication. | Picture and sound without fields such as GPS data, device identifiers, software tags, and telemetry. |
video-sync uses orientation-specific layouts: landscape inputs stack the footage and data vertically, while portrait inputs place them side by side. A single CSS file controls the presentation. The release materials describe video-scrub as a companion utility, so its current location and availability should be checked against the repository tree.