video-frames-for-agent

name	video-frames-for-agent
description	Extract sparse PNG frames from operator video on Windows using scripts/video-frame-samples.ps1 and ffprobe; use L1-L3 taxonomy in docs/ops/VIDEO_FRAME_EXTRACT_FOR_AGENT.md.

Video frames for agent context (Windows)

When to use

• The operator shared an MP4 (or other ffmpeg-readable media) and the assistant needs visual context (terminal, Maestro, UI).
• The operator can give approximate seconds (“interesting at ~3s”).

Ritual (token-aware)

1. Read docs/ops/VIDEO_FRAME_EXTRACT_FOR_AGENT.md for the L0-L3 ladder if unsure.
2. From repo root, prefer .\scripts\video-frame-samples.ps1:
   • -ProbeOnly first when duration / “is this worth sampling?” matters.
   • -Timestamp / -TimestampsCsv at operator hints.
   • -MaxWidth 960 or 1280 when full resolution is unnecessary.
3. read_file on each printed PNG path (vision path).
4. If ffmpeg missing: report the script’s one-line install hint (winget install ffmpeg); do not pretend video was watched.

Anti-patterns

• Claiming the model “watched” a video without ffprobe / frame extraction.
• Dense fps=1 over long clips without operator approval (L3 cost).

Pointers

• scripts/video-frame-samples.ps1 (-Help)
• docs/ops/VIDEO_FRAME_EXTRACT_FOR_AGENT.md
• Filename discovery elsewhere: scripts/es-find.ps1