Video keyframes

Extracts the frames that show something new — not one frame every N seconds, and not every frame. Produces frames/*.jpg and timeline.json, a token-lean digest you read before opening any image.

When to use

The user wants what was on the screen: "pull the slides out of this", "what's on screen at 4:20", "screenshot the commands he runs", "read the diagram".

When not to: they want the spoken words (video-transcribe), or a written note (video-notes — it calls this skill itself).

Speech-only video is not worth framing. For a podcast, interview or talking-head, frames cost effort and teach nothing. Check the genre first.

Input is a local file