Watching Videos Agent Skill

Watching videos

You don't have a video input; this skill gives you one. Anything a user

hands you — a YouTube link, a TikTok, a lecture, a meeting recording, a

screen capture, an .mp4 on disk — goes through the same pipeline: frames

(scene-aware, deduplicated), OCR, transcript (captions first, local

Whisper offline fallback), all persisted into one index.

Before watching: check the index

watch-skill list

If the video was already analyzed — this session or any earlier one —

do NOT watch it again. Ask it directly (see the asking-with-evidence

skill):

watch-skill ask <video_id> "<question>"

One video

watch-skill watch "<url-or-path>" [--start T --end T] [--max-frames N] [--transcript-only]

• Works on any yt-dlp-supported site (1800+), direct media URLs, HLS/DASH

manifests (--duration 60 bounds live streams), and local files.

• Video over ~10 minutes and the user cares about one part → use

--start/--end for dense sampling of that window.

• User only needs what was said → --transcript-only (fastest, often no

video download at all).

The report prints Indexed: <video_id>, frames with t=MM:SS

timestamps, OCR text, and the transcript. Read every frame path listed —

in a single message, parallel Read calls — then answer from frames + OCR

+ transcript, citing timestamps.

Many videos (playlist, channel, folder)

watch-skill batch "<playlist-url-or-folder>" [--limit N]

Everything lands in the same index; one broken video never stops the

rest. Afterwards a single watch-skill search "<phrase>" spans the whole

batch.

First run on a machine

If any command fails with a dependency error, run watch-skill doctor

once — it installs missing ffmpeg/yt-dlp itself. No API key is required

for any of this; transcription is local by default and the video file

never leaves the machine.