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.

The crv command (a short alias for claude-real-video) is the primary way to use this tool from a terminal. You can also invoke it as python -m claude_real_video — both forms accept identical arguments.

Synopsis

crv <source> [options]

Arguments

source
string
required
Video URL (YouTube, Instagram, TikTok, and any site supported by yt-dlp) or a local file path (e.g. lecture.mp4). This is the only positional (required) argument.

Options

-o, --out
string
default:"crv-out"
Output directory where all files are written. Created automatically if it does not exist. Re-running overwrites the existing contents.
--scene
float
default:"0.30"
Scene-change sensitivity on a 0–1 scale. Lower values mean more frames are captured on each cut; higher values require a more dramatic visual change before a new frame is recorded. Adjust downward for fast-cut videos and upward to reduce redundant frames in slow-paced content.
--fps-floor
float
default:"1.0"
Guarantee at least one frame is captured every N seconds, regardless of scene-change detection. Acts as a density floor so that long static scenes still produce at least one representative frame per interval.
--max-frames
integer
default:"150"
Hard cap on the total number of frames after deduplication. When the surviving frame count exceeds this limit, frames are thinned uniformly so that coverage remains spread across the full video timeline.
--lang
string
default:"auto"
Whisper language code for audio transcription. Accepts any code the Whisper model understands, such as en, zh, fr, ja, or auto (let Whisper detect the language automatically). Has no effect when --no-transcribe is set or when existing subtitles are found.
--cookies
string
Path to a Netscape-format cookie file for login-gated video sources. Pass cookies that belong to your own authorised account only — never commit credential files to a repository.
--no-transcribe
flag
Skip audio transcription entirely. No transcript.txt will be produced, and the --- transcript --- section of MANIFEST.txt will be empty. Useful for silent videos or when you only need the frames.
--dedup-threshold
float
default:"8"
Percentage of pixels (in a downscaled RGB thumbnail) that must differ from every frame in the dedup window for a new frame to be kept. Higher values are more permissive and keep fewer frames; lower values are stricter and retain more visual variation. A value of 8 means a frame is kept only if at least 8 % of its pixels changed colour compared to each of the last N kept frames.
--dedup-window
integer
default:"4"
Size of the sliding window used during deduplication. Each candidate frame is compared against the last N kept frames, not just the immediately preceding one. This prevents an A-B-A editing pattern from re-introducing a shot the model has already seen after a cutaway. Set to 1 for classic consecutive-only comparison.
--report
flag
Write a report.html deduplication visualiser and move all dropped frames into a dropped/ subdirectory instead of deleting them. Open report.html in any browser to inspect every keep/drop decision and its pixel-diff percentage — useful for tuning --dedup-threshold and --dedup-window.
--why
string
A short description of your viewing intent, written verbatim into MANIFEST.txt. For example: --why "find the pricing strategy". When present, the manifest instructs the LLM to analyse with this intent as its lens — surfacing relevant moments first rather than producing a generic summary.
--kb
DIR
Path to a knowledge-base directory (e.g. your Obsidian vault or notes folder). When provided, crv copies the analysis as a dated Markdown note (YYYY-MM-DD-<slug>.md) into that directory so the result persists alongside your other notes instead of living only in crv-out.
--keep-audio
flag
Also save the full original soundtrack as audio.m4a in the output directory. The audio file contains music, speech, and sound effects — everything the transcript does not capture. Intended for audio-capable models such as Gemini and GPT-4o.

Examples

crv "https://www.youtube.com/watch?v=dQw4w9WgXcQ"

Exit codes

CodeMeaning
0Success — all output files have been written to the output directory.
1Error — a descriptive message is printed to stderr. Common causes include a failed download (private video, missing yt-dlp), a missing ffmpeg installation, or an invalid argument.

Build docs developers (and LLMs) love