Watch Video

watch-video

Claude cannot play video, but it can read images. This turns a recording into a contact sheet plus individual frames, and transcribes the audio when there is any.

Public share links need no auth — most screen-recording tools (Zight, CloudApp, Loom, Droplr, ...) produce links that are viewable without logging in. Fetch them directly; don't ask the user for screenshots as a first move.

Run it

"${CLAUDE_PLUGIN_ROOT}/scripts/video-digest.sh" <url|file.mp4> [--ticket <id>] [--title <text>]

${CLAUDE_PLUGIN_ROOT} is the plugin's install directory — use it rather than a relative path. The script sits at the plugin root, not inside this skill's own folder, so a relative ./scripts/... resolves to the wrong place and fails with "no such file or directory". Outside a plugin install (a plain clone), ./scripts/video-digest.sh from the repo root is the equivalent.

Pass --ticket and --title when known — they name the output directory, which is how the user navigates a persistent cache later. By default output goes to the OS temp directory; pass --output <dir> for a persistent one.

Useful flags: --max-frames N, --no-dedupe, --transcript never, --force, --sample-fps N (raises the pre-dedupe frame pool — combine with --no-dedupe when a transient frame keeps getting dropped). Full list and current config: --help.

Read it

1. contact-sheet.jpg first — one image, the whole clip, timestamps burned in. Usually enough on its own.
2. frames/frame_NNN.jpg only when the sheet leaves something unresolved.
3. transcript.txt when present. It is grouped by frame: each block gives the words spoken while that frame was on screen, so speech and screen state read together. A — means nobody spoke over that frame. The unbroken transcript follows at the end for a quick "what is this clip about".

The first and last frames are always included — the opening state and the outcome. "Last" is the last frame that survived dedupe, not the end of the file, so it lands on real content instead of a black screen a recorder can leave at the end.

Timestamps on the sheet are approximate (frame ordinal × interval, not true PTS) — fine for navigating, not for quoting exact times.

Cell size adapts to the source, so tablet and phone recordings in either orientation both stay legible. The frames in frames/ are always full resolution — open one when the sheet's downscale hides a detail.

What to watch out for

• The default frame count is a starting point, not a limit. When the sheet doesn't settle the question, re-run with a higher --max-frames and say what was ambiguous. Do this before handing anything back to the user.
• A sequence where a count changes (items deleted, created, or reordered) needs dense re-extraction before any conclusion goes into a report. The default budget is tuned for "what is this clip about", not "what exactly happened, in what order". A real triage session read a sparse sheet as "a deleted item reappeared" — re-extracted at a much higher --max-frames, the same clip showed something else entirely (a second, different item was deleted; a transient loading state the sparse sheet had skipped explained the miscount). The wrong conclusion had already been written into a report before the density was raised. Treat any story involving a count as provisional until checked at higher density, not just when the sheet already looks ambiguous.
• Dedupe can drop the exact frame that explains a discrepancy. It keeps visually distinct frames, but a brief loading/transient state can look similar enough to the frame before or after it to get dropped. If a count or an order doesn't add up, re-run with --no-dedupe in addition to a higher --max-frames.
• A missing transcript.txt usually means silence, not failure. Screen recorders often write an empty audio track when the mic is off; the tool measures loudness and skips transcription below a threshold. The summary line says which happened.
• The recording date is not necessarily today, or the date it was shared. meta.json carries recorded and recorded_source — check them before assuming a recording reflects the current state of anything.
• archived/ (inside a persistent output dir) is the user's. A recording moved there is resolved; the tool refuses to re-process it without --force. Don't move things in or out on the user's behalf.
• source.url is a double-clickable shortcut to the original, for when frames aren't enough. It exists only for a URL input — a local file has no url to record.
• Frames answer what happened on screen. They do not prove why — confirm the mechanism elsewhere before reporting a cause.
• End the answer with where the output landed, one markdown link per line. The script's last lines print the output directory and the paths to contact-sheet.jpg, frames/, transcript.txt, source.url and meta.json — pass those through as [label](file:///abs/path) so the host can render them clickable; a bare path string is inert in most chat UIs. One link per line, not ·-joined on one line — several markdown links crammed together render as a cluttered run-on, each expanding inline. A summary with no paths leaves the user unable to open the sheet, check a frame you described, or reach the original recording, and they have to ask. A short closing list is enough:

Dependencies

ffmpeg (with libfreetype for timestamps), ffprobe and python3 are required and checked up front. curl is needed only for a URL input, whisper only when the audio carries speech, yt-dlp only for sites the built-in scraper can't resolve — all three degrade gracefully without failing the run. See "${CLAUDE_PLUGIN_ROOT}/scripts/video-digest.sh" doctor for what's currently available.