analyze-videolisted

# analyze-video Self-contained pipeline that takes one or more video sources, downloads or resolves them locally, extracts frames, uses captions or Whisper for transcripts when available, tiles frames into contact sheets for cheap visual review, selects representative frames, and produces a polished Word document with timestamped analysis. ## Safety and privacy boundary This skill runs `yt-dlp`, `ffmpeg`, and `ffprobe` locally. Source video files, frames, contact sheets, manifests, and the final `.docx` stay on the user's machine. Extracted audio is sent to Groq or OpenAI only when no native captions are available and a Whisper key is configured. Do not try to bypass platform bot detection or access controls. If a site blocks unauthenticated downloads, the safe fallback is explicit user authorization: ask whether the user wants to use their own browser session via `--cookies-from-browser <browser>` or a cookies file via `--cookies <path>`. Do not spoof watch sessions, forge tokens, automate hidden playback to trick a site, or use unrelated hosting/services as an evasion layer. ## Token strategy Do not read every frame. The pipeline emits per-chunk contact sheets and a lightweight manifest so you can preview the video at low cost: 1. Read `manifest_lite.json` first. It omits transcript text AND per-frame arrays but includes chunk metadata, timestamps, contact-sheet paths, `docx_image_dimensions`, `suggested_docx_name`, `transcript_path`, quick-mode flags, and the full m