Index videos once to cut repeated AI analysis costs

Running a full video analysis every time someone asks a new question about the same file wastes work. A better setup treats the video as a reusable knowledge source instead of temporary input. The pipeline first pulls out the transcript, OCR text, scene breaks, and key frames, then stores each observation with its timestamp.

It creates embeddings and builds a local index. Later questions use hybrid retrieval to find only the relevant evidence. Only that evidence is sent back to the language model, instead of sending the whole video through the multimodal pipeline again.

The main shift is making the model reason from selected evidence rather than re-understand the entire video each time. The approach is packaged as an open-source project called Watch Skill, with access through MCP, a command-line interface, and a REST API.