Video Report Nemotron

Video Report Nemotron is a Hermes skill that turns a video URL or local media file into a Markdown, HTML, or PDF report with selected frames placed next to the sections they illustrate. It works with YouTube, other yt-dlp supported sources, and local audio or video, and runs inside Hermes CLI or Hermes Desktop once installed into the active skills directory.

Transcription is subtitle-first: by default it tries the platform's subtitles and falls back to local automatic speech recognition (ASR) only when none exist. The bundled ASR backend is Apple Silicon MLX with mlx-community/nemotron-3.5-asr-streaming-0.6b-8bit. On Linux or Windows, subtitle-based reports still work but local ASR needs an explicitly configured Nemotron backend, and the skill never substitutes Whisper. Report composition and vision fallback use any OpenAI-compatible endpoint and model, with gpt-4.1-mini as the default.

The pipeline has two output layers. The first layer produces an intermediate JSON and Markdown transcript artifact with source metadata and timestamps. A chain of scripts then reviews transcript blocks to plan visual evidence, captures frames with ffmpeg from a locally downloaded copy of the video, runs OCR, applies a multimodal vision model only to OCR-weak frames, and composes the final reader-facing report. The final language can be forced independently of the source language, and the composer retries once when the provider output ignores the requested language.