Support video input (screen recordings) - native frame extraction for UI bug reports

Problem

Claude Code cannot read video files. For a non-developer using Claude Code to build an application, a screen recording is the most natural way to report a UI bug — a 20-second capture shows a dropdown reverting to its previous value, or a button that appears dead, far better than several screenshots and a paragraph of prose. Today, attaching an .mp4 yields nothing usable: Read supports images, PDFs and notebooks only.

Current workaround

Claude can shell out to ffmpeg and do it itself:

ffmpeg -i rec.mp4 -vf "fps=1/2,scale=1400:-1" f_%02d.jpg                     # frames every 2s
ffmpeg -i rec.mp4 -vf "fps=1/3,scale=640:-1,tile=3x3" -frames:v 1 sheet.jpg  # contact sheet

This works — it located a real bug for us today (a dropdown that silently reverted the user's selection). But it requires ffmpeg to be installed, and the model has to think of doing it. Most users will simply conclude "Claude can't see my video" and give up.

Also tested: an animated GIF is accepted by Read, but only the first frame is visible, and GIF palette quantization badly distorts dark-theme UIs (a dark dashboard came back tinted green). So GIF is not a usable workaround either.

Request, in order of preference:

1. Native video attachments in Read — sample N frames, ideally with timestamps.
2. A first-party plugin/skill that performs the extraction automatically when a video path is passed.
3. At minimum, document the ffmpeg recipe so the model reliably reaches for it.

Context

Windows 11, Claude Code desktop. Recordings come from the built-in Windows Game Bar (.mp4) — which is what an ordinary user has to hand, with no extra software installed.