watch-videolisted

# watch-video — let Claude actually watch a video Most tools only read a transcript and miss everything visual. This skill gives Claude the **frames + the words, aligned on one timeline**, so it can reason about cuts, on-screen text, demos, and pacing — not just the audio. ## How to use it When the user gives a video URL: 1. Run the helper (it downloads, samples frames, and transcribes): ``` python scripts/watch.py "<VIDEO_URL>" --out watch_out ``` Useful flags: `--max-frames 60` (cap frames), `--no-whisper` (captions only, fully free/offline), `--every 0` (auto-pace by duration). 2. It writes: - `watch_out/frames/f000.jpg …` — sampled frames (named by index) - `watch_out/transcript.txt` — the spoken words - `watch_out/index.json` — `[{frame, t, file}]` so each frame maps to its timestamp 3. **Read the frames as images** (use the Read tool on the PNG/JPGs) alongside `transcript.txt`, lining each frame up with what's being said at its `t`. Then answer the user's request (summary, breakdown, hook analysis, study notes, a fresh script, etc.) grounded in BOTH the visuals and the audio. ## Analyze — the full breakdown Once `watch.py` has run, read the frames + `transcript.txt` aligned by `index.json`, then write a structured breakdown. Ground every point in a specific frame `t` and/or transcript line: - **Hook (first ~3s)** — exactly what's on screen and said in the opening, and *why* it stops the scroll. - **Format** — the style (talking-head, facele