Three extractors turn an already-watched video into something structured. All of them answer straight from the persistent index — deterministic, no extra model calls. The video must be watched first (see ).

watch-skill extract chapters <video_id-or-source>

Titled chapters with start/end timestamps, built from scene changes plus transcript topic shifts. Good for long tutorials, lectures, meetings — give the user the chapter list with timestamps, not a prose summary.

watch-skill extract bug-report <video_id-or-source>

Pinpoints where an error appears: timestamp, the exact frame, the on-screen error text as OCR read it, and the steps that led up to it (from the transcript/actions preceding the failure). Returns honestly when no error signal exists — report that as "no error found in the recording", not as an error.

Paste the output into the user's issue tracker format when they ask for a fileable report; the frame path is the attachment.

watch-skill extract hook <video_id-or-source> [--seconds 15]

Scores the opening seconds on four measured axes — attention trigger in the opening line, speech pacing, visual change rate, on-screen text — each with an actionable critique, plus a 0–100 score and verdict. Give the user the per-axis critiques; the number alone helps nobody.