What problem does it solve? Customer walkthrough recordings (Loom, Meet, Zoom captures) are often the only ground truth for a system you cannot access directly, but manually scrubbing hours of video to document screens, fields, and spoken business rules is slow and error-prone. This Skill turns a recording into a buildable spec by joining extracted UI frames with a timestamped transcript. ## Core Features & Use Cases - Frame Extraction with Crop & Scene Detection: Probe video geometry, crop out webcam overlays, and extract distinct UI frames using ffmpeg scene detection tuned for form-heavy screens (threshold 0.02). - Two-Tier Local Transcription: Transcribe the full recording locally with Whisper (base.en), then re-run targeted segments with larger models to verify domain terminology. - Master Index Generation: Join frames and transcript into a timecoded Markdown index for writing findings, vocabulary files, and question lists. - Use Case: You receive a 90-minute Zoom recording of a customer demoing an insurance portal. Run probe, frames, transcript, and index to produce a findings document with navigation paths, field inventories, branch points, and traps — ready to feed a live capture session for selectors. ## Quick Start Ask the agent to mine the walkthrough video at a given path by running the probe, frames, transcript, and index commands with a run id, then review the generated index under .local/recordings.
