llm-real-video (Free)

Your LLM reads subtitles, not videos. This free skill fixes that.

What it does

Give it a YouTube URL or a local video file. It extracts:

• Scene-aware keyframes — only the frames that actually changed (scene-change detection + dedup, no ML model downloads)
• Transcript — from subtitles when available, or local Whisper
• Contact sheets — pack a 58-second clip into 3 images instead of 58 frames

Everything is written as plain local files (frames + JSON manifest + markdown), so any LLM can read them. Nothing is uploaded anywhere.

Proven open source

This is the same claude-real-video that has 2k+ GitHub stars and hit the Hacker News front page. MIT licensed — the source is public.

Requirements

• Python 3.10+, ffmpeg
• pip install claude-real-video (the skill runs this CLI)

Want deeper analysis?

llm-real-video Pro adds creator-analysis modules: editing rhythm, camera language, and a second-by-second sound/voice/emotion/gesture timeline. Find it on my publisher page.

New in 0.10.5: Apple Silicon GPU transcription

pip install "claude-real-video[mlx]" runs Whisper on the Mac GPU: a 21-minute talk went from 381 s to 62 s on an M4, transcript identical. A voice-activity gate still runs first, so music or silence yields "no speech" instead of invented captions. Other machines are unaffected.

Also since 0.7.9

• URL videos reuse the platform's own captions when they exist (0.10.2), so most YouTube runs skip speech recognition entirely.
• --from / --to analyse only part of a long recording; --frame-width keeps small on-screen text readable (0.10.0).
• --speakers labels every transcript line ([SPEAKER_00], [SPEAKER_01], ...) with a fully local 45 MB diarization model (0.7.9). Install with: pip install "claude-real-video[speakers]"