video-insight-report

Provide one public video URL plus the questions or viewpoint that should guide the report. The skill acquires metadata and subtitles, extracts timestamped frames, and renders an offline HTML report with a pyramid summary, evidence cards, and a first-principles section.

When no usable subtitles exist and conclusions depend on speech, the skill falls back to local transcription using mlx-whisper on Apple Silicon. Cloud transcription is never used silently and requires explicit user authorization before any audio is uploaded.

No, the skill forbids passing Bilibili URLs directly to yt-dlp because platform protection can return 412 errors. It uses the bili CLI and opencli bilibili adapters instead, and a Bilibili report is not accepted until a real frame is verifiably extracted.

Yes, the rendered report.html is self-contained with embedded CSS and evidence images stored under the report directory. Validation checks offline resources, image hashes, and desktop, narrow-screen, and print behavior before completion.

A report is INCOMPLETE when a required condition is missing, such as real frames for a claim or a reliable transcript for speech-dependent conclusions. COMPLETE requires validated evidence, a first-principles section, passing manifest validation, and visual inspection.

It only supports single public YouTube or Bilibili videos, not playlists, private or member content, comments, or danmaku. Low-confidence transcript text cannot independently support core conclusions, and unresolved gaps are preserved rather than guessed.