watch-video gives Claude eyes and ears for any video — local files, public URLs (YouTube, Loom, Vimeo, TikTok, ~1,500 sites via yt-dlp), or Jira attachments fetched directly via API token. One command produces a complete evidence bundle: timestamped JPEG frames, a Whisper transcript (local faster-whisper or hosted Groq / OpenAI), optional OCR for on-screen text, and three report formats — report.md, self-contained report.html with base64-embedded frames, and corporate-friendly report.docx. Key features: - LLM-driven highlights: pass a prompt ("summarize the rate decision and inflation outlook", "identify the bug and when it occurs") and Claude picks the N most relevant moments with frame + reason + verbatim quote. - Smart dedup with perceptual hash + transcript-aware protection: typically drops 40-60% of redundant frames on screen recordings without losing narrated moments. - OCR for screen-recording UIs: Tesseract with 2x upscale + auto-invert + PSM 6, output to grep-able ocr.txt. - Per-step content-hash cache: re-runs are nearly free. - Opt-in Jira posting with a confirmation gate BEFORE any attachment upload — declining leaves the ticket untouched, no orphan state. - Batch mode for sprint retros: process a JQL query's worth of bug tickets in one command.