video-analysis

This skill provides video understanding for any model by choosing automatically between two modes based on file size. For files under a configurable threshold (default 20MB) it passes the video natively to a model that supports video input (default Gemini 3.1 Flash Lite) so the model sees every frame and hears the audio. For larger files it falls back to extraction: ffmpeg pulls keyframes (fixed interval for short clips, scene-change detection for long ones) and Whisper transcribes the audio, returning frame image paths plus transcript text for the current chat model to reason over. This solves the problem of analyzing, describing, summarizing, or transcribing video across models with differing native-video support and context limits.

The skill exposes a small Python API (analyze_video, get_config, get_video_info) with careful invocation guidance — it warns against dotted imports of the hyphenated directory name and against exec-loading analyze.py because of __file__ usage — and documents result keys for both native and extraction modes. It includes a model benchmark comparing cost, latency, and accuracy across several Gemini and Qwen models, and a layered configuration system where a workspace config file overrides factory defaults and survives skill updates. For web URLs it advises trying a companion web-crawler skill's transcript path before downloading.

Target users are agents and developers who need to understand local video files — describing content, summarizing clips, or transcribing recordings. It processes local media through standard tools (ffmpeg, Whisper) and model APIs with no destructive, network-exfiltration, or credential-harvesting behavior, so it is benign.