Automatic Video Subject Analysis and Keyframe Selection

Problem

Short video clips often require manual scrubbing before an engineer can answer simple questions: who is the subject, where does the action occur, and which frame is worth using as a representative thumbnail. Uniform frame sampling misses important motion and pushes many near-duplicate frames into a vision model, wasting tokens.

How It Works

This skill automates a focused workflow: extract keyframes from a video and use a vision model to produce a readable report.

• Reads duration, resolution, codec, and bitrate metadata with ffmpeg
• Extracts keyframes using scripts/extract_keyframes.sh and I-frame detection
• Outputs JPEG images at 640px width for downstream visual analysis
• Analyzes the extracted frames and returns a text report plus 3 representative screenshots

Limits and Caveats

It fits short-video subject identification, action summarization, and thumbnail candidate selection. It is not a replacement for fine-grained temporal tracking in very long videos or continuous complex motion. The token estimates in the reference material depend on video length, frame count, resolution, and model behavior.