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.

Every crv run produces a self-contained output directory designed for direct LLM consumption. The files are deliberately simple — plain JPEG images, plain UTF-8 text, and a short manifest — so that any model can ingest them without special parsing, tool calls, or pre-processing. The directory is intended to be re-run in place: running crv again on the same source and output path will overwrite the previous results.

​
Directory structure

crv-out/
├── source.mp4          # downloaded or copied source video
├── frames/
│   ├── frame_001.jpg
│   ├── frame_002.jpg
│   └── ...             # scene-aware, deduplicated key frames
├── transcript.txt      # full transcript (plain text)
├── MANIFEST.txt        # summary for the LLM
├── audio.m4a           # (optional, --keep-audio) full soundtrack
├── dropped/            # (optional, --report) rejected duplicate frames
└── report.html         # (optional, --report) dedup visualiser

​
MANIFEST.txt

MANIFEST.txt is the primary artifact to give to an LLM. It is a plain-text file that summarises the run, lists the frame and transcript paths, and — when --why is used — opens with a focused viewing intent that instructs the model to prioritise relevant content over a generic summary.

​
Example

viewing intent: find the pricing strategy
(reader: analyse the frames and transcript with this intent as the lens — surface what serves it first, skip what doesn't)

source: https://youtu.be/...
duration: 312s | frames: 42 (scene-change + density floor, deduped from 187 extracted)
frames dir: crv-out/frames
transcript: crv-out/transcript.txt (transcribed by whisper)
--- transcript ---
Hello, welcome to...

​
Field reference

FieldDescription
viewing intentPresent only when --why was passed. States the purpose of the analysis and is followed by a line instructing the LLM to use it as its analytical lens.
sourceThe original URL or file path passed to crv.
durationVideo length in whole seconds.
framesCount of kept frames after deduplication, with a parenthetical showing the extraction and dedup parameters.
frames dirAbsolute path to the frames/ subdirectory.
transcriptAbsolute path to transcript.txt, plus a note on how it was produced (existing subtitles or Whisper), or a human-readable explanation if no transcript is available.
audioPresent only when --keep-audio was passed. Absolute path to audio.m4a, or a note that the video has no audio track.
--- transcript ---Separator followed by the full plain-text transcript inline, so a model reading only MANIFEST.txt has the words without opening a second file.

​
frames/*.jpg

Key frames are stored as JPEG images inside the frames/ subdirectory. Each image is:
  • 640 px wide, with height scaled proportionally to preserve the original aspect ratio.
  • Named frame_NNN.jpg where NNN is a zero-padded integer starting at 001, in strict chronological order. The numbering is assigned after deduplication and capping, so there are no gaps.
Frames are selected by a single chronological ffmpeg pass that captures every scene change (controlled by --scene) plus a density floor of at least one frame per --fps-floor seconds, ensuring that both fast-cut reels and slow screencasts produce appropriate coverage. Near-duplicate frames are then removed using real RGB pixel-diff comparison against a sliding window (see --dedup-threshold and --dedup-window).

​
transcript.txt

transcript.txt is a plain UTF-8 text file containing the spoken content of the video, one subtitle line per text line. All timecodes, sequence numbers, WEBVTT headers, and inline styling tags (e.g. <v ->, <b>) are stripped, leaving only the words. Source priority (first match wins):
  1. Sidecar file — a .srt or .vtt file with the same base name as the local source video (e.g. lecture.srt alongside lecture.mp4).
  2. Embedded subtitle stream — a subtitle track muxed directly into the video file, extracted via ffmpeg and converted to plain text.
  3. Whisper transcription — audio is extracted to a temporary 16 kHz mono WAV, passed to the whisper CLI (requires pip install openai-whisper), and the result is saved as transcript.txt. The language is controlled by --lang.
If none of the above succeed (no subtitles, no Whisper install, or a video with no audio track), transcript.txt is not written and MANIFEST.txt explains the reason clearly.

​
audio.m4a

audio.m4a contains the full original soundtrack — music, speech, and sound effects — as a single M4A container file. It is only written when --keep-audio is passed. The encoding strategy prioritises quality and speed:
  • Lossless stream copy when the source audio codec is already AAC or ALAC (no quality loss, fast).
  • AAC re-encode at 192 kbps as a fallback for other codecs (e.g. Opus, Vorbis).
The audio file complements the transcript: the transcript has the words, but audio.m4a lets a model that can listen (Gemini, GPT-4o, and similar) hear the music, tone, pacing, and sound design — detail that a text transcript cannot convey.

​
report.html

report.html is a self-contained HTML visualiser for the deduplication pass. It is only written when --report is passed. Open the file in any browser (no server required) to see:
  • Every frame extracted before deduplication, displayed in a grid.
  • Each frame labelled with its name, pixel-diff percentage, and status:
    • Green border — kept after dedup.
    • Red border, faded — dropped as a near-duplicate.
    • Orange border, faded — removed by the --max-frames cap after dedup.
  • A summary line showing the threshold, window, and kept/total counts.
When --report is active, dropped frames are also moved into a dropped/ subdirectory (rather than deleted), so the HTML <img> tags that reference them resolve correctly. Use report.html to tune --dedup-threshold and --dedup-window for your specific content: if too many visually distinct frames are red, lower the threshold; if too many near-identical frames are green, raise it.
Handing the output to an LLM: drop every image from frames/ plus MANIFEST.txt into the model’s context window. The manifest is self-contained enough for a quick pass; adding the frame images gives the model the full visual context. If you also passed --keep-audio, include audio.m4a for models that accept audio input (Gemini, GPT-4o, etc.).

Build docs developers (and LLMs) love

Get started for freeTalk to us