I Built a YouTube Knowledge Base Because I Kept Forgetting Everything I Watched

YouTube is the world's largest learning platform. Millions of hours of expert tutorials, conference talks, and deep dives — all free. But here's the problem: YouTube is designed for entertainment, not retention.

You watch a brilliant 45-minute system design video. Two weeks later, you need that exact information, and you can't remember which video it was, what the creator said, or even what to search for. The knowledge is gone.

I kept running into this. I'd watch 10+ hours of tutorials on a topic, take scattered notes, and still lose track of the insights buried across dozens of videos. YouTube search only finds new videos — it can't search within what you've already watched.

So I built Skip.

What Skip Does

Skip turns your YouTube videos into a searchable knowledge base. Save a video, and Skip extracts the full transcript, generates AI summaries, creates semantic embeddings, and makes everything searchable by meaning — not just keywords.

The key insight: search across your entire library at once. Not one video at a time. Your whole collection.

Here's how it works:

1. Save videos — Chrome extension, paste a URL, or bulk-import an entire channel
2. AI processes everything — transcripts, summaries, and semantic embeddings generated in ~90 seconds
3. Search by meaning — ask "how does connection pooling work?" and find the answer even if the creator said "database connection reuse"
4. Chat with your library — ask questions and get answers citing exact timestamps across all your videos
5. Discover creators — describe what you want to learn, and AI finds expert YouTubers the algorithm misses

Why Not Just Use NotebookLM?

Fair question. NotebookLM is great for uploading a few documents and having a conversation. But:

• NotebookLM doesn't do YouTube at scale. You can add YouTube links one by one, but there's no concept of a growing video library, bulk imports, or channel imports.
• No semantic search. NotebookLM is a chat interface. Skip gives you both chat AND search, with results ranked by relevance.
• No creator discovery. Skip can find expert YouTubers you've never heard of based on what you want to learn.
• MCP integration. Skip has an MCP server — you can query your video knowledge base from Cursor, Claude Desktop, or any MCP-compatible tool. Your video knowledge, inside your coding workflow.

For a deeper comparison, see our Skip vs NotebookLM breakdown.

The Tech Stack

• Frontend: Next.js 14, Tailwind, deployed on Railway
• Backend: FastAPI (Python), Celery workers for async processing
• Database: Supabase (PostgreSQL + pgvector for embeddings)
• AI: OpenAI for embeddings and chat, with RAG (retrieval-augmented generation)
• Transcripts: youtube-transcript-api with fallback strategies
• Search: pgvector cosine similarity over 1536-dim embeddings

For a deep dive into the architecture, check out our technical walkthrough.

Real Example

I imported 30 videos from a Kubernetes tutorial channel. Later, I asked Skip: "How do I set resource limits for a pod?"

Skip found the relevant 2-minute segment across a 3-hour playlist, cited the exact timestamp (1:23:45), and gave me a concise answer with a direct link to that moment. No rewatching. No scrubbing.

Try It

Skip is live at getskip.dev. Free tier: 50 videos/month, 100 messages, full search and chat.

Pro ($10/mo) gets you 500 videos, 2000 messages, MCP integration, and API access.

If you've ever lost track of something you learned on YouTube, Skip was built for you.