What You’ll Learn

James has 684 YouTube videos. If you asked him what’s in any of them, he couldn’t tell you. That’s not a storage problem — it’s a memory problem. And AI tools like Claude have the same problem: every new session starts from zero.

In this video, James walks through how he solved this using an idea from Andrej Karpathy (one of the people behind ChatGPT). The concept is called an LLM Wiki — a structured second brain built from your own content, readable by Claude in seconds.

By the end, you’ll understand what an LLM Wiki is, why it matters for educators and coaches building an agent-powered teaching business, and what you can start doing with 10 videos (not 684).

The Problem: Claude Starts From Zero Every Session

Every time you open a new conversation with Claude, it remembers nothing from the last one. You have to re-explain your context, your past work, your audience, your frameworks — every time.

James had this exact problem at scale. He had 684 YouTube videos, research documents, transcripts, and Zoom recordings — but none of it was accessible to Claude as a unified reference. Claude had to search through individual files or videos each time, which defeated the point of having an AI assistant.

“Every session I don’t have to reexplain anything. It’s not 684 videos and 684 files. It is one single place of truth.” — James

Most educators hit this wall much earlier. If you’ve been running Zoom calls, creating YouTube videos, writing content, or hosting community discussions for any length of time — you have more valuable thinking archived than you realise. The problem is that Claude can’t use it without a way to find it fast.

What Is an LLM Wiki?

An LLM Wiki is a structured collection of markdown files that Claude can read like an employee handbook. It’s not a search engine and it’s not a database. Think of it as a curated, compressed summary of your knowledge — organized in a way that Claude can navigate quickly.

Karpathy’s idea was simple: instead of pointing Claude at 684 raw video files, you have Claude process all of that raw material into a structured wiki first. Claude then reads the wiki, not the raw files.

Here’s the flow in plain English:

• Raw material — Your YouTube transcripts, Zoom recordings, research docs, spreadsheets, coaching call notes
• Claude processes it — Reads everything, extracts summaries, identifies concepts, maps relationships
• Wiki is written — Organized markdown files built around a schema you define
• Claude reads the wiki — Now every session starts with full context, not zero

💡 In Plain English: Imagine hiring a librarian to read every book in your library, write a one-page summary of each, and organize all of them by topic. You’d still have access to the original books — but you could also just ask the librarian and get an instant answer. That’s what an LLM Wiki does for your content.

The Schema: What Gets Extracted

A wiki is only as useful as how it’s organized. Karpathy introduced the idea of a schema — a set of categories that defines what Claude extracts and stores from your raw content.

James set up a schema specifically for his YouTube videos. His categories include:

• Comparisons — Videos where he compared tools, approaches, or choices
• Concepts — The frameworks and mental models he’s taught over time
• Entities — People, companies, and tools he’s mentioned
• Techniques — Specific methods or tactics he’s demonstrated
• Tools — Every AI or software tool he’s covered
• Videos — Summaries of individual videos, linked to the original

You define the schema based on what matters to your teaching business. If you’re a business coach, your schema might include frameworks, client outcomes, and common objections. If you teach creative writing, it might include story structures, writing exercises, and student questions.

Once the schema is in place, Claude uses it to organize every piece of content you feed into the wiki — automatically.

✓ Check Your Work: Before building a wiki, ask yourself: “What would I want Claude to know about all my content?” Your answer becomes your schema categories.

What Obsidian Adds (Optional)

James uses a free tool called Obsidian to visualize the wiki. Obsidian shows all your content as an interconnected graph — you can see how topics link to each other, which concepts appear most often, and how different videos relate.

You don’t need Obsidian to get value from an LLM Wiki. But if you’re a visual learner or you want to see patterns across your content library, it’s worth looking at.

James describes it this way: looking at the graph view of his 180+ ingested videos, he could see deep connections between topics like AI agents, course creation, YouTube strategy, and community-led learning — connections he wouldn’t have noticed by scanning individual files.

10 Ways to Use Your Wiki

Once your wiki is connected to Claude, here’s what becomes possible:

1. Content gap finder — Ask Claude what topics you’ve never covered. Stop guessing what to create next.
2. YouTube series builder — Ask Claude to map all your videos on a theme and group them by teaching angle. James did this live in the video for “AI agents” and got 6 angles, each with a list of existing videos.
3. Instant research before filming — Check what you’ve already said about a topic before recording. No more accidentally repeating yourself or contradicting past content.
4. Student use cases — Your students can query your wiki through an agent. They get answers grounded in your thinking, not a general web search.
5. Cross-wiki intelligence — Connect multiple wikis (James has three: YouTube videos, AI operating system, and business outputs). When they’re linked, Claude can find insights that span your entire knowledge base.
6. Contradiction detector — Find places where your thinking has evolved or shifted. Update old content or create follow-up videos.
7. Tutorial draft generator — Turn any question from a student into a tutorial draft, grounded in what you’ve already taught.
8. Tool and platform tracker — See which tools you’ve covered (Gemini, ChatGPT, Claude) and which you haven’t. Spot gaps in your coverage before your audience does.
9. Lesson planner — Build a course from scratch by asking Claude to map your best content into a curriculum. The wiki tells Claude what you’ve already covered at each level.
10. Business knowledge base — Connect your AI operating system to the wiki. Every output your business produces — emails, community posts, session notes — feeds back in and makes the wiki smarter.

The Competitive Moat Argument

This is the part James says most people aren’t getting yet.

A competitor can build the same LLM Wiki setup. They can copy your tools, your templates, and your schema. That takes a weekend.

What they cannot copy is your content library.

“They can’t copy 684 videos of me thinking. And if you’ve got two months, four months, six months of captured data — that cannot be replicated easily.” — James

Every educator, coach, and consultant who has been creating content already has a head start. Your Zoom recordings, your coaching call transcripts, your YouTube videos, your community discussions — all of it is raw material that a competitor would need months or years to replicate.

And once it’s in a wiki connected to Claude, it compounds. Every new video, every new session, every community discussion adds to the library automatically. The wiki gets smarter every week without you doing anything extra.

This is what separates an agent-powered teaching business from someone who just uses AI tools. Your knowledge base is the moat.

You Don’t Need 684 Videos to Start

James is quick to point this out. You can start with 10 videos — or 10 Zoom call recordings, or 10 coaching session notes. The wiki starts small and grows with you.

The important thing is to start capturing your digital exhaust now. Every time you teach, record it. Every time you coach, transcribe it. Every time you host a community session, save the notes. The raw material is worth nothing if it disappears — and it’s worth a lot once it’s in the wiki.

“It works with any content. Video calls, transcripts, spreadsheets, documents — it gets converted. Coaching calls, community discussions — it compresses, it learns, it gets better.” — James

Key Takeaways

• An LLM Wiki turns your accumulated content into a single, Claude-readable second brain — no RAG database needed.
• The schema defines what relationships Claude extracts: concepts, entities, tools, techniques, and individual video summaries.
• For educators, the wiki serves two audiences: your AI agents internally, and your students through agent access.
• Your content library is your competitive moat. Competitors can copy the tools — not the 684 videos of your thinking.
• You can start with 10 pieces of content. The wiki grows automatically as you create more.

Your Next Step

James set this up in a weekend by giving Claude the Karpathy setup guide and asking it to build the structure. He’s documented the process and all the prompts in the TrainingSites campus.

If you want to build your own LLM Wiki, the process is inside the campus at trainingsites.io. James has also linked resources in the video description so you can download copies of the prompts and use cases he walked through.

The starting point is simple: whatever content you have right now — YouTube videos, Zoom recordings, coaching call notes — that’s your raw material. Start capturing it if you aren’t already. Then connect it to Claude through a wiki, and you’ll have a second brain that remembers everything you’ve ever taught.