The project README has been overhauled to be more approachable for new users, moving from dense technical details to a clear, quick-start guide organized by effort level. Advanced technical documentation and implementation details have been moved into a collapsible section, improving overall readability without sacrificing depth.

Added new effort level controls (quick, normal, deep, max) to the vu teardown, overview, and budget tools. Users can now balance analysis depth with resource usage, with deep offering part-level breakdowns and max providing frame-by-frame granularity. This change provides greater flexibility in how thoroughly the system processes video content.

We've added a comprehensive deep teardown pipeline to the video understanding skill that decomposes videos into detailed technical components, including native editing cuts, motion tracking via GSAP eases, typographic synchronization, and audio analysis. This system runs six parallel analysis lenses, leveraging updated Swift-based vision tools and a dedicated NumPy/OpenCV/Pillow environment, to provide a structured digest and style guide for AI processing. This granular data extraction allows for much deeper programmatic insights into how complex videos are structured and produced.

Updated the vu video understanding tool to share its voice-detection model (ggml-silero-v5.1.2.bin) with reelcut's watch tool. By checking for the model in reelcut's existing directory structure before attempting a download, we reduce redundant storage usage and improve integration between these related tools.

Added a new 'vu.py' toolset that enables Claude to interact with video files using a Think-Act-Observe loop. By combining native Python processing with ffmpeg, whisper.cpp, and macOS Vision, the agent can now perform tasks like searching transcripts, OCR, frame-grid analysis, and motion mapping. This update includes a comprehensive skill guide and predefined recipes for tasks ranging from reel critiques to complex video-to-step breakdowns.