watch-videolisted
Install: claude install-skill rasarafi/settled-frames
# Watch Video
Claude can't take video directly. This turns a video into timestamped frames plus a local Whisper
transcript aligned to each frame, has Haiku agents caption the frames in parallel, and lets the main
model reason over the captions instead of hundreds of images.
**Why 2 fps + settled frames:** tools that pick frames by scene change skip forms that fill in after
a spinner, and drop small popups, dropdowns and tooltips. This samples twice a second and keeps the
last frame of every stretch where the screen held still, so it keeps what the screen settled on.
Pairing each frame with its speech also catches homophones ("Murray" said, "Marie" on screen) and
said-vs-done gaps.
**Paths:** `<skill-dir>` below is this skill's base directory (shown when the skill loads).
`<venv>` is `~/.local/share/watch-video-venv` unless the user set it up elsewhere.
## Steps
1. **Prepare** (local, nothing uploaded). Write output to a scratch or temp folder, never into a repo:
```bash
<venv>/bin/python -I <skill-dir>/scripts/prepare.py "<file or URL>" --out <scratch>/watch-<slug> \
[--start MM:SS --end MM:SS] [--lang en] [--ignore x0,y0,x1,y1]
```
It prints raw/kept frame counts, speech segments, batch count and timings. `--help` lists the
other flags (`--threshold`, `--fine`, `--hold`, `--fps`, `--batch`, `--whisper-model`, `--no-audio`).
If the venv is missing: `python3 -m venv <venv> && <venv>/bin/pip install faster-whisper pillow numpy yt-dlp`.
**Check th