Skip to main content

Documentation Index

Fetch the complete documentation index at: https://mintlify.com/HUANGCHIHHUNGLeo/claude-real-video/llms.txt

Use this file to discover all available pages before exploring further.

This page summarises the main changes introduced in each version of claude-real-video. For the very latest updates and patch releases, see the GitHub releases page.

​
0.3.0

​
New flags

  • --why — Write your viewing intent directly into MANIFEST.txt. The model reads this up front and analyses the frames and transcript through that lens rather than producing a generic summary. 
    crv "https://youtu.be/..." --why "find the pricing strategy"
  • --kb — Save the analysis as a dated markdown note (e.g. 2025-01-15-video-slug.md) into a folder you specify — an Obsidian vault, a plain notes directory, anywhere. The note lives on after crv-out is overwritten. 
    crv "https://youtu.be/..." --kb ~/notes
  • --keep-audio — Preserve the full original soundtrack as audio.m4a alongside the frames and transcript. The audio is copied losslessly where possible, and re-encoded to AAC otherwise. This lets audio-capable models (Gemini, GPT-4o, …) actually hear the video — not just read the words.
  • --report — Generate report.html in the output directory, showing every extracted frame — kept, dropped as a duplicate, or removed by the --max-frames cap — along with its pixel-diff percentage. Open in a browser to tune --dedup-threshold and --scene visually.

​
Subtitle priority chain

claude-real-video now checks for existing subtitles before falling back to Whisper, in this order:
  1. A sidecar .srt or .vtt file sitting next to a local source file.
  2. An embedded subtitle stream inside the video container.
  3. Whisper transcription of the audio track (unchanged fallback).
Using captions that already exist is both faster and more accurate than re-transcribing.

​
Sliding-window deduplication (--dedup-window, default 4)

Previously, deduplication only compared each frame against its immediate predecessor. Version 0.3.0 compares against the last N kept frames (default 4). This catches A-B-A cuts — a shot the model has already seen won’t be sent again just because a different frame appeared in between. Set --dedup-window 1 to restore the old consecutive-only behaviour.

​
Honest transcript notes

MANIFEST.txt now distinguishes clearly between two “no transcript” situations that were previously conflated:
  • Silent video (no audio track): (none — this video has no subtitles and no audio track)
  • Whisper not installed: (none — no existing subtitles; install whisper to transcribe: pip install openai-whisper)
The current release is 0.3.0. Install or upgrade via pip: pypi.org/project/claude-real-video.

Build docs developers (and LLMs) love

Get started for freeTalk to us