To let an agent see a video it cant play, ffmpeg a tiled contact sheet
An agent cannot watch an mp4/mov, but it CAN read a single image. To analyze or edit a video (understand its structure, pacing, which shots are where), render a contact sheet: ffmpeg -y -i in.mp4 -vf "fps=1/2,scale=210:-1,tile=5x5" -frames:v 1 sheet.png samples one frame every 2s into a 5x5 grid (read left-to-right, top-to-bottom = the timeline). One Read call then reveals the whole edit at once (titles, cuts, transitions, which half is dynamic vs static). Tune fps=1/N to the video length so the grid covers it (25 slots): N ~= duration/24. Pair with ffprobe -show_entries format=duration:stream=width,height for specs, and ls the asset folders to know the raw material. This turns opaque video into something an agent can actually reason about and give shot-by-shot direction on.