cole-medin-knowledge-base

type: overview title: "Cole Medin AI Knowledge Base" description: "An Open Knowledge Format (OKF) wiki mined from Cole Medin's entire long-form YouTube catalog, built to be dropped next to any project as agent-ready reference." tags: [readme, overview] updated: 2026-07-21

Cole Medin - AI Knowledge Base

A synthesized, densely cross-linked knowledge base mined from Cole Medin's entire long-form YouTube catalog - agentic coding, AI engineering, RAG, harnesses, memory systems, and more. It is an Open Knowledge Format (OKF v0.1) bundle and a Karpathy-style LLM wiki: plain markdown + YAML frontmatter, navigated by index and relative links, no database and no embeddings required.

The point: drop this folder next to your project and any coding agent can answer questions from Cole's videos with citations back to the exact video and timestamp - zero integration, no RAG stack, no API keys.

What's inside

• concepts/ - the ideas, techniques, patterns, and mental models (the PIV loop, context engineering, agentic search, and so on), each synthesizing every video that discusses it.
• entities/ - the tools, people, and organizations (Claude Code, Archon, MCP, Anthropic, …).
• sources/ - one summary page per video, with provenance to the raw transcript.
• raw/ - the immutable, timestamped transcripts (the source of truth).
• index.md - start here (after SCHEMA.md). The compiled wiki is the search index.

Use it with an agent (copy-paste)

Point your coding agent at this bundle with a prompt like:

…

Build your own (from any YouTube channel)

Three Claude Code skills are included that replicate the full pipeline - just point them at a channel. Each fetches transcripts using a different method, then walks you through the extract-canonicalize-write process described below.

# Example: build a KB from 3blue1brown's channel using yt-dlp
/channel-to-kb-ytdlp @3blue1brown

Each skill fetches all transcripts as raw/*.md files, then guides Claude through the same pipeline used to build this bundle: extract concepts from each transcript, canonicalize (merge duplicates into single canonical pages), write cross-linked concept/entity/source pages, and validate with lint.py. The full pipeline reference is at .claude/references/pipeline-guide.md.

For small channels (under ~30 videos), the whole build fits in one Claude Code session. For larger channels, the skill batches the work and you can resume across sessions.

How it was built (reproducible)

1. Source. Full transcripts for every long-form video are pulled from the source database and written as immutable raw/.md files (timestamped), plus raw/manifest.json.
2. Extract. Each transcript is mined for the concepts and entities it teaches, with timestamped quotes.
3. Canonicalize. All candidates are deduplicated into one taxonomy so each durable idea is a single page synthesizing every video that covers it (synthesis, not one-page-per-video).
4. Write & link. Concept/entity/source pages are written and cross-linked with typed relationship headings.
5. Validate. python scripts/build_indexes.py regenerates the indexes; python lint.py gates conformance, link integrity, index coverage, and orphans.

For the full story - the two-pass agent architecture, the tooling, and the validation results (citation integrity, adversarial QA, recall) - see docs/MAKING-OF.md and tools/.

See SCHEMA.md for the full contract and docs/ingestion-workflow.md for the step-by-step. Coverage and growth model are in roadmap.md.

License / provenance

Knowledge is synthesized from publicly available YouTube videos by Cole Medin; each page cites its sources. Raw transcripts are included under raw/ for provenance and auditability.