video-ingest

What problem does it solve?

Manually extracting video content is inconsistent, wastes compute on unnecessary vision engines for transcript-only tasks, and lacks validation for AI-generated visual summaries that may contain ungrounded claims.

Core Features & Use Cases

• Capability-routed ingestion: Automatically selects the optimal extraction method (native multimodal for long videos, Claude vision over ffmpeg-extracted frames for short/medium videos, yt-dlp for transcript-only tasks) based on your task needs, available engines, and video length to avoid wasted compute.
• Governance cross-check: Validates all multimodal visual summaries against the video transcript or a second frame pass to catch false or ungrounded claims before relaying them as fact.
• Edge case handling: Supports auth-gated videos via your logged-in browser session and clearly surfaces specific failure reasons (no captions, unsupported streams, unavailable engines) instead of returning empty results.
• Use Case: Need a summary of a 1-hour product demo with on-screen UI walkthroughs? It uses a native multimodal engine to capture both spoken commentary and visual details. Only need the transcript of a 5-minute team standup? It pulls captions directly via yt-dlp without spending a vision engine.

Quick Start

Use the video-ingest skill to get a full summary of both the spoken content and on-screen visuals from the YouTube tutorial at https://youtube.com/watch?v=example456, including key demo steps and talking points.