video · git:20260930.ec54481 · 2026-09-30 · sha256 ba22edfb269c34ac
video git:20260930.ec54481A
Immutable. This exact content never changes and is served at /api/v1/blob/ba22edfb269c34ac.
No tool to install first
claude plugin marketplace add https://markdownregistry.com/mp/a/art_qz3dmvdjwga2373z/ba22edfb269c34ac4410a47134273ad6d2698aabef1018d7df3b8f6c9253a40a.json
Then /plugin install video@markdownregistry-qz3dmvdjwga2373z in Claude Code. It installs this exact version, the one graded A here, pinned to the commit it was graded at, and leaves it there.
--- name: video description: Read a video file (.mp4/.mov/.webm) by extracting frames with ffmpeg and summarizing what it shows, e.g. a QA bug recording. Use when given a video path; not for still images or live app testing. --- - Require `ffmpeg`; if missing, report it and stop. Install only with user approval, preferring a no-sudo static build. - Work in a temp dir (`$CLAUDE_JOB_DIR/tmp` when set, else `mktemp -d`); never write beside the video. - Get the duration with `ffprobe`. Extract frames at 1 fps with `-vf "fps=1,scale='min(1280,iw)':-2"` (never upscale); use `-ss`/`-t` when the user gives a time range, and higher fps only for fast transitions. - Read up to 15 frames directly. For more, delegate to one subagent when available, passing the frame dir, the user's bug description, the video duration, and the frame-to-second rule below with an instruction to report each moment as a second from that rule (never invent times beyond the video duration), and use only its summary; otherwise sample fewer frames. - Extract and transcribe audio only when the user wants speech and `whisper` is available; otherwise state that audio was not analyzed. - Report only what frames show, with the second of each observation (second = frame number − 1 at 1 fps, plus any `-ss` offset), and state that motion, timing between frames, and sound are not covered. Do not infer a root cause without code evidence.