slides

Slides

Static images (JPG/PNG/static WebP) can be read natively. This skill is only for video and animated images — anything that has multiple frames across time. It uses peepshow (ffmpeg under the hood) to extract a sequence of relevant still frames so they can be viewed as a timeline.

Input

The user may provide the video (or animated image — GIF, APNG, animated WebP) in any of these forms:

• A local file path (absolute, relative, or a network mount like /Volumes/share/...)
• An http:// or https:// URL
• A data:video/...;base64,... or data:image/(gif|apng|webp|png);base64,... URI pasted into the prompt
• The literal - to read bytes from stdin (you would pipe via Bash)

If $ARGUMENTS contains the video reference, use it directly. Otherwise ask the user to share one of the above.

Steps

1. Run the CLI with the Bash tool. JSON is the most reliable output for parsing:

   PEEPSHOW_CLIENT=claude-code PEEPSHOW_SESSION="${CLAUDE_SESSION_ID:-}" peepshow "$ARGUMENTS" --emit json
   

   The PEEPSHOW_CLIENT + PEEPSHOW_SESSION env vars tag the run in the manifest and the access log so a shared peepshow serve instance can attribute every run + HTTP call back to the right Claude Code session. Both are optional — peepshow runs fine without them — but setting them costs nothing.

   (If the path contains spaces, quote it. Use --emit paths if you prefer reading the human-readable list.)

2. Parse the output. In JSON mode, frames[].path is the ordered list of absolute paths. The video object gives you container, codec, resolution, fps, duration, and file size — useful context for the user's question without extra prompting. The video.tags object carries container-level metadata embedded in the file (title, artist, album_artist, director, producer, publisher, copyright, genre, description, creation_time, show, episode_id, season_number, etc.) — use it to ground your answer in what the video says it is before describing what you see. The extraction object tells you which strategy (scene vs fps) was used, how many frames were pruned, how many were dropped by the perceptual-hash dedup pass (framesDeduped + dedupDistance), and a coarse motion signal across the kept frames (motionSignalAvg numeric + motionSignalLevel low/medium/high). Use the motion signal to colour your narration (e.g. "rapid action segment" vs "near-static timelapse"); combine with framesDeduped == 0 + motionSignalLevel == "high" to recognise a high-information clip where the LLM should pay close attention to every frame.