| name | channel-to-kb-ytdlp |
| description | Build an OKF (Open Knowledge Format) knowledge base from any YouTube channel. Uses Google's OKF standard so any AI can read the output with zero setup. Free, no API key. Uses yt-dlp (most reliable). Invoke with /channel-to-kb-ytdlp @ChannelHandle |
| argument-hint | <@ChannelHandle or channel-URL> |
Build an OKF Knowledge Base from a YouTube Channel (yt-dlp)
Turn any YouTube channel into a synthesized, cross-linked knowledge base in Open Knowledge Format (OKF). OKF is Google's open standard for LLM-readable knowledge bases: plain markdown + YAML frontmatter, navigated by index and relative links, no database or embeddings required. Any AI can read the output with zero setup.
Uses yt-dlp for both channel enumeration and subtitle download. Free, no API key required.
Trade-offs: most reliable method (yt-dlp is actively maintained and handles YouTube changes fast), captures publish dates automatically. Recommended for production use.
Before you start
Read the OKF contract this bundle obeys:
- Read
SCHEMA.md in this repo (the maintainer contract: page types, frontmatter schemas, linking rules)
- Skim the OKF SPEC for the hard rules (every page needs
type in frontmatter, index.md is reserved, cross-links are relative markdown paths)
Step 1: Fetch transcripts
uv run .claude/skills/channel-to-kb-ytdlp/scripts/fetch_transcripts.py $ARGUMENTS --output-dir ./raw
This enumerates all videos, downloads English subtitles as json3, converts them to timestamped OKF raw-transcript pages. Output: raw/<slug>.md files (each with type: raw-transcript frontmatter, OKF-conformant) + raw/manifest.json.
Flags:
--limit N to cap the number of videos (start here for testing)--delay 3.0 to increase the pause between requests if you hit 429 rate limits
Wait for the script to complete before proceeding.
Step 2: Build the OKF knowledge base
Read .claude/references/pipeline-guide.md for the full process. It produces an OKF-conformant bundle with the structure defined in SCHEMA.md. The stages:
- Extract - read each raw transcript, extract concepts/entities/quotes as JSON to
scripts/extractions/
- Canonicalize - merge all extractions into a frozen taxonomy (
scripts/manifest.json + scripts/taxonomy.json)
- Write - write OKF concept/entity/source pages from the manifest (each with proper
type, title, description frontmatter per SCHEMA.md)
- Index - build
index.md files for each directory (OKF's navigation layer)
- Validate - run
python lint.py to enforce OKF conformance, link integrity, and index coverage
Process in batches per the guide. For channels under ~30 videos, this fits in one session. For larger channels, save your extraction JSONs and resume across sessions.
Step 3: Validate OKF conformance
python lint.py
This enforces the OKF contract: every .md has type frontmatter (E1), every relative link resolves (E2), every page appears in its directory's index.md (E3), sources/raw parity (E4). Fix all errors. The knowledge base is ready when lint passes clean.
