rea-video Tears Apart Any Video Shot by Shot With 99.6% Accuracy

Takeaways

rea-video is an Apache-2.0 agent skill that reverse-engineers short videos shot by shot.
Classifies each shot as live action, AI-generated, digital human, 3D, code animation or screen recording.
Exports SRT/ASS subtitles, caption styling, cut timing, BPM and voiceover type per clip.
Reports 99%+ accuracy on original files, 88.6% on re-encoded uploads without a vision model.
Runs inside Claude Code or Codex via a single install instruction; needs ffmpeg and Python 3.10+.
Pairs with capsule-cinema for automated AI-shot regeneration onto the original timeline.

Reverse-engineering a short video often requires frame-by-frame inspection to separate live footage, AI-generated scenes, screen captures, and motion graphics built with HTML and CSS. rea-video turns that work into a structured, shot-level analysis.

Released under the Apache 2.0 license, the project integrates with coding agents such as Claude Code and Codex. It identifies how each shot was produced, explains the evidence behind its classification, and proposes a practical recreation plan. The repository has 24 GitHub stars, while retaining a narrow scope: its primary output is an analysis report rather than a finished remake.

A forensic pass, shot by shot

Given a local video file, rea-video detects shot boundaries, separates voice, music, and sound effects, then collects evidence from each segment. Video checks cover frame-rate patterns, image noise, whole-frame zooms and pans, embedded AI-generated content metadata, hands, text, and physical consistency. Audio analysis estimates tempo, musical key, and whether vocals are sung or spoken.

The workflow gives stronger forensic signals, such as metadata and frame-rate signatures, more weight than visual impressions. Conflicting signals trigger another visual inspection by the agent.

The resulting report.html includes:

Per-shot screenshots and timestamps
A production-method classification for every shot
Evidence and reasoning behind each verdict
Reverse-engineered generation prompts where applicable
Camera, lighting, color, and sound notes
A recommended production route for rebuilding the sequence

Six routes to a rebuild

Many AI-content detectors assign one authenticity score to an entire clip. rea-video instead classifies each shot and maps it to a production workflow.