vision-replaylisted

<essential_principles> **Video analysis pipeline:** video file -> deduplicate static frames -> ffmpeg frame extraction -> Claude vision analysis -> structured report. All frame extraction is handled by deterministic shell scripts. Workflows guide your analytical reasoning -- what to look for and how to structure findings. Scripts handle the mechanics. **Frame storage:** All frames go to `/tmp/claude-video-frames/<timestamp>/`. Always clean up after analysis using `cleanup.sh`. **Context window management:** Read at most 15-20 frames per batch. Use `batch-frames.sh` to organize large frame sets. Use `contact-sheet.sh` for quick overviews before detailed extraction. Works seamlessly with recordings from the record-browser skill — annotations drawn during recording appear as bright colored shapes in the extracted frames and are automatically prioritized during analysis. </essential_principles> <prerequisites> Before starting any analysis, verify: 1. ffmpeg is installed: `which ffmpeg` 2. The video file exists and is accessible 3. Set up working directories: ```bash SCRIPTS_DIR="${CLAUDE_PLUGIN_ROOT}/skills/vision-replay/scripts" WORK_DIR="/tmp/claude-video-frames/$(date +%s)" mkdir -p "$WORK_DIR" ``` </prerequisites> <universal_pipeline> **Run these steps for EVERY analysis before routing to a workflow:** **Step 1: Get video metadata** ```bash bash "$SCRIPTS_DIR/video-info.sh" "<video-path>" ``` Review the output: duration, fps, resolution. This informs fps selection.