watch

Purpose

Look at a rendered video and report what is actually on screen and in the audio. By default it watches the entire video and considers visuals, voiceover, music, and sound effects. Callers can narrow the scope (specific timestamp ranges, lower frame rate, disable audio tracks) when they want a cheaper or more focused pass.

This is the observation primitive that watch-and-refine calls before deciding what to fix. Other review and editing skills can call it directly.

Inputs

• video — path to a local video file. Required.
• ranges — optional list of [start, end] timestamps to watch. Accepts SS, MM:SS, or HH:MM:SS. Defaults to the whole video.
• fps — frame sampling rate. Defaults to auto by duration (≤30s → 1–2 fps, 30s–1min → ~1 fps, 1–3min → ~0.5 fps, 3–10min → ~0.25 fps). Hard cap 2 fps.
• max_frames — hard cap on total frames sampled across all ranges. Default 100.
• resolution — frame width in px. Default 512. Bump to 1024 only when on-screen text legibility matters.
• include_voice — bool, default true. Transcribe spoken VO/dialogue.
• include_music — bool, default true. Describe music presence, swells, drops, gain relative to VO.
• include_sfx — bool, default true. Note sound effects, foley, transition stingers.
• focus — optional free-text prompt describing what to pay attention to (e.g. "watch the end card", "judge cut timing on the beat drop").