I downloaded Obsidian three times. I deleted it twice. The third time I almost gave up. That purple gem icon sat in my dock, mocking me. Every time I opened it, I saw a blank canvas and felt that familiar, sinking question: What am I supposed to do with this?
If that sounds familiar, you’re not alone. Thousands of people download Obsidian, Notion, or Roam, excited to build a second brain — only to abandon it within a week. The problem isn’t the tool. The problem is the order of operations.
Most tutorials teach you to set up folders, tags, and frameworks first. That’s backward. The real first step is getting the AI to know you. Not your documents. You.
Let me show you what I mean.
Why LLM Wiki changes the game
When you ask ChatGPT a question, it answers and forgets. Next time, it starts over. LLM Wiki (popularized by Karpathy) flips that: instead of treating each query as isolated, it builds a living encyclopedia from everything you feed it. You drop in an article; AI writes a summary, updates related entries, links concepts, even flags contradictions. The more you feed it, the smarter it gets.
But here’s the trap: copying Karpathy’s setup gives you his knowledge base, not yours. It’s like walking into someone else’s library — the books are arranged by their tastes, not yours. That’s where I got stuck. Three times.
The three traps I fell into
Trap #1: The English framework felt foreign.
I followed the GitHub instructions and got a perfectly organized Obsidian vault — all in English. I understand English, but it felt like someone else’s system. I changed it to Chinese, which felt better psychologically, but the real issue wasn’t the language. It was that the structure had no relationship to my life.
Trap #2: I tried to design the structure before having any content.
I created folders for Work, Life, Learning — then subfolders. I thought I needed a clean taxonomy before I could start. But the structure is lived into, not designed ahead of time. Karpathy himself said his vault evolved organically: he threw in a few papers, saw what broke, tweaked the rules, added more, repeated. You can’t plan a knowledge base you haven’t started using.
Trap #3: I was doing the work, not the AI.
I told the AI: put this article in folder X, summarize that note, link these two concepts. I was the project manager, AI was the intern. Exhausting. In a true LLM Wiki, AI is the architect — you just supply raw material and ask questions. It decides how to classify, link, and update. The shift from manager to user is everything.
The missing first step: AI recognition
Then it hit me. I talk to AI every day. My ChatGPT history knows my job, my current projects, my thinking style, even my pet peeves. But when I opened Obsidian, it knew nothing about me. I was starting a relationship with a complete stranger.
So I did something different. I asked AI to summarize everything it knew about me.
Prompt: “Scan our entire conversation history and build a personal profile. Only include facts you have direct evidence for. Mark inferences. Skip categories with no data. Sort by information density.”
The output stunned me. AI listed my role, my industry, my ongoing challenges, the tools I use, the learning goals I’d mentioned months ago, even my decision-making style. It knew me better than I gave it credit for.
Then I fed that profile into a second prompt — my “personal knowledge base architect” — and asked it to build a vault from scratch, tailored to me. You don’t need a better tool. You need a tool that knows you.
The result was nothing like the generic template. It had folders for my actual projects, my current learning topics, my health goals. Every category reflected something real in my life. It felt like a librarian who had known me for years had set up the shelves.
How I use it now
I see an interesting article. I clip it with Web Clipper into the raw folder. Then I tell the AI: “Archive this to the knowledge base.” It reads it, writes a summary, updates relevant pages, and links it to related content. Sometimes a new article adds a paragraph to a note from three months ago. I don’t touch a thing.
Last week I asked: “What do the three best articles in here say about agent design?” It returned a cross-document synthesis I could never have done manually. Another time, it connected a piece I’d saved in January to one I saved yesterday — revealing a trend I’d completely missed.
Occasionally I open the graph view and see a dense web of nodes. Some are isolated — notes that never connected. That’s a signal. So I built a health check.
Knowledge base maintenance
Even a smart vault gets messy. I wrote a prompt that scans the entire directory and returns a report: broken links, orphaned notes, unprocessed materials, outdated entries. It gives a health score (out of 100) and lists the top three things to fix. I just say “fix” and the AI restructures what needs restructuring. Takes five minutes every few weeks.
The knowledge base doesn’t grow from empty folders. It grows from knowing who you are.
Obsidian is still that same purple gem icon. But now I know what to do with it. The magic wasn’t in the tool. It was in the moment I stopped treating AI as a database and started treating it as a partner who actually knows me. Start with that — and everything else falls into place.
FAQ
Q: Why can't I just start organizing by topic? Won't that save time?
A: It feels productive, but it's actually counterproductive. Without the AI knowing your specific context, the categories you create will be generic and quickly become irrelevant. You'll spend more time re-organizing later. The AI needs to know <em>you</em> first so it can build a structure that matches how you actually work and think.
Q: So I need to chat with ChatGPT for months before building my knowledge base?
A: Not necessarily months — even a few weeks of regular use can give enough signal. The key is to feed the AI your existing conversation history (if available) or just start using it for real tasks. Once you have a profile, you can use it to bootstrap your knowledge base immediately. The principle is: AI recognition before AI organization.
Q: Isn't this just making the AI do all the work? What's the point of me curating content?
A: You still curate — you choose what goes in. But the AI handles the heavy lifting of linking, summarizing, and maintaining consistency. The value you add is the selection and the questions you ask. The AI is your assistant, not your replacement. The difference between a messy pile of saved articles and a real knowledge base is exactly this partnership.