video-digest

Turn a YouTube URL or local video into one digest.md — the full timestamped transcript with every real on-screen visual (slide, screenshot, UI, diagram, whiteboard) captured inline: OCR'd, described, de-duplicated, with talking-head/filler frames filtered out.

Runs fully local against any OpenAI-compatible vision endpoint — no cloud keys. Fits a 16 GB GPU (or runs on CPU with ~24 GB RAM).

Example: an excerpt of the digest for a coding video (rendered):

[10:45] and also just to verify like it can [10:46] actually do the upload second I wanted

🖼 Visual @ 10:48 — screenshot

On-screen text: Arduino Board Setup /home/ladyada/Arduino/libraries/Adafruit_TestBed/examples/I2C_Scan … Now let's compile and upload the I2C_Scan example …

A terminal window displaying Arduino CLI commands for compiling, uploading, and checking serial output for an I2C_Scan example, with a Bash command prompt at the bottom asking for user input.

_{Frame from “Fully automating Arduino development” by Adafruit Industries, licensed CC BY.}

What you need

1. ffmpeg and yt-dlp on PATH.
2. A local vision model on an OpenAI-compatible endpoint. Easiest is LM Studio: install → search & download minicpm-v-45 (a ~6 GB vision model) → Developer → Start Server. That's the whole backend. (Ollama or llama.cpp's llama-server work too — anything that serves /v1/chat/completions with a vision model.)

Install

# 1. get the skill (until it's on a plugin marketplace)
git clone https://github.com/ollo12-prog/video-digest ~/video-digest
ln -s ~/video-digest ~/.claude/skills/video-digest    # or copy the folder there

# 2. one-time preflight — checks ffmpeg, yt-dlp, and your endpoint
python3 ~/video-digest/scripts/setup.py

setup.py prints the exact fix for anything missing and lists the model ids your endpoint serves (set CLASSIFY_MODEL / DESCRIBE_MODEL if yours differ from the minicpm-v-45-q4-32k-vision default).

Use

In Claude Code (or any skill host):

/video-digest https://www.youtube.com/watch?v=…

Or directly:

python3 scripts/digest.py "https://www.youtube.com/watch?v=…" out/
# → out/digest.md

# just the transcript: captions only, no video download, no vision model needed
python3 scripts/digest.py "https://youtu.be/…" out/ --transcript-only

# only part of a video
python3 scripts/digest.py "https://youtu.be/…" out/ --start 12:30 --end 18:00

Config (all optional, sensible defaults) is env-based — VLM_ENDPOINT, CLASSIFY_MODEL, DESCRIBE_MODEL, VD_RES; see SKILL.md.

How it works (short)

• Captions first, Whisper only as fallback. Most YouTube has captions (free, timestamped); caption-less videos use local whisper.cpp (optional, see notes).
• Interval keyframes, not scene cuts. It samples on a time grid and keeps a frame when it's visually new and persists — so it captures gradual change (a whiteboard drawn over 15 minutes) that scene-cut detection misses entirely, while dropping transient flickers and re-shows.
• One small vision model does both jobs (keep/drop + OCR/describe), so it fits 16 GB and never swaps. minicpm-v-45 was picked because it's the only model tested that reliably separates verbatim OCR from a summary on dense terminal/code frames (larger models garbled or hallucinated there).
• Talking-head filler is filtered — a presenter on camera is not a "visual."

Caption-less videos fall back to local whisper.cpp; see docs/WHISPER.md.

Credits

Inspired by bradautomates/claude-video (the /watch skill) — its clean skill packaging and "captions-first, Whisper- fallback" approach shaped this project. This tool goes a different direction on the video track: local vision-model OCR/description of de-duplicated keyframes instead of raw frame sampling.

License

MIT — see LICENSE.