Video To Skill

video-to-skill

Convert any video into a queryable SKILL.md — local, PII-redactable, runs on laptop.

video.mp4 ─→ [faster-whisper + PySceneDetect + Tesseract] ─→ SKILL.md
                                                              chapters/
                                                              cheatsheet.md
                                                              vocab.md

What it does

• Transcribe audio via faster-whisper (local, no cloud)
• Sample frames with ffmpeg, optional OCR with Tesseract
• Extract vocabulary, topics, chapters via pluggable local LLM (Ollama / LM Studio)
• Render a SKILL.md mirroring the book-to-skill format (~4 KB)
• Redact phone numbers, emails, IDs, explicit names — by default

When to use

• Turn lecture recordings into reusable lesson notes
• Index webinars, talks, how-to videos for docs/agent memory
• Convert training videos into SKILL.md for downstream agents
• Run on air-gapped/laptop machines without cloud APIs

Quickstart

git clone https://github.com/awesome-edu/video-to-skill.git
cd video-to-skill
pip install faster-whisper PySceneDetect

python scripts/video_to_skill.py path/to/video.mp4 \
    --name my-skill \
    --redact-name Alex --redact-name Liam

Output:

output/my-skill/
├── SKILL.md          # main skill (4-6 KB)
├── cheatsheet.md     # 1-page quick reference
├── vocab.md          # domain vocabulary
├── chapters/         # scene-by-scene chunks
└── raw_transcript.json

Type-aware extraction (optional)

Pass --type for schema tuned to content kind:

• --type lecture — chapter by intro/body/conclusion
• --type cartoon — character dialogues + plot
• --type meeting — speaker diarization + agenda
• --type howto — step-by-step commands
• --type documentary — timeline + subjects

Design Principles

1. Pure extraction by default — no LLM classification unless you ask
2. Type-aware is opt-in — pass --type for schema tuning
3. PII redaction is on by default — phone/email/ID/names scrubbed at source
4. Local-first — ffmpeg + faster-whisper + Tesseract + optional local LLM
5. Zero magic deps — Python 3.10+ and ffmpeg is the minimum

Cross-skill compatibility

Output SKILL.md layout mirrors book-to-skill (one-line-per-fact, 4-6 KB). Downstream tools that load book-to-skill work without changes.

Requirements

• Python 3.10+
• ffmpeg on PATH
• faster-whisper (pip install faster-whisper)
• Optional: Tesseract for --ocr
• Optional: Ollama/LM Studio on localhost:11434 for type-aware

License

MIT © 奥深 Awesome Education

Acknowledgments

Built by 奥深 Awesome Education (Alex + Athena + Mia + Cleo + Bria + Vera). First public version: 03:10, 2026-07-16.