STOP. READ THIS ENTIRE SKILL.MD BEFORE CALLING ANY ENDPOINT.

Watch Skill

Watch any video (URL or local file) and get:

1. Scene-change frames — one JPEG per detected shot via ffmpeg select=gt(scene,…)
2. Transcript — native captions first, then Whisper via scillm
3. SRT emotion/scene analysis — for local files with subtitles
4. Scene element table — timecode, transcript text, marker image, movie segment, and sound note per sampled scene
5. HTML inspection table — thumbnails, visual description status, playable video/audio controls where local paths are available
6. Structured report — frames manifest + transcript + pacing metrics

Planned diarization is a separate audio evidence lane, not a visual watcher and not a character-identification model. See "Diarization Contract" below.

Quick Start

cd ${HOME}/workspace/experiments/agent-skills/skills/watch