KishoreVetha/watch-video-skill ? reverse-engineered prompt

Build me a simple video watching skill for my coding agent that can take a YouTube link, a direct video URL, or a local video file and turn it into timestamped frames the agent can inspect. I want it to work without any transcript API or paid service, just using local tools, and if captions are available it should grab those too. Please make it smart enough to sample more densely for short videos and keep a reasonable frame budget for longer ones, and burn the timestamp onto each image so the frames are easy to understand later.

If many images are awkward, also add an option to bundle them into contact sheets. I’d like it to output a simple manifest with the frames, timestamps, and captions path, and make it easy to use from Codex or Claude Code. If you need to check current install or usage details online, do that.