Personal LLM Wiki Setup: What Works After Six Months

TL;DR

A personal LLM wiki is a structured markdown knowledge base where an LLM writes and updates pages from your raw notes — not a chat history dump or a folder of AI summaries.
Mine has 35 pages after six months, organized in three layers: raw input, curated wiki pages, and schema definitions. The schema layer is what separates it from "notes with AI."
The biggest failure mode isn't discipline — it's treating the wiki as a search engine instead of a thinking partner.

My Obsidian vault has a folder I open almost every day. It's not a to-do list or a journal. It's a folder called wiki/ that now has 35 pages written mostly by Claude, based on things I've learned, read, or thought through over the past six months.

I call it my personal LLM wiki. Here's what that actually means, and how to build one that doesn't collapse after two weeks.

What a Personal LLM Wiki Actually Is

Most people's first instinct is to think of it as "ChatGPT but you own the data." That framing is wrong in a useful way.

A personal LLM wiki is a structured knowledge base where:

You supply raw context — pasted articles, rough notes, voice transcripts
An LLM writes the wiki pages following a schema you define
You curate and correct the output before it becomes part of the permanent record

The key word is structured. Without a schema telling the LLM what fields to extract and how to format them, you end up with a folder of AI-generated summaries that don't link to each other. I ran that version for the first three weeks. It felt productive and taught me nothing I couldn't have found by searching.

The schema file is what makes a personal LLM wiki different from "notes with AI." I wrote about the Karpathy-origin version of this in my post on the karpathy llm wiki pattern. This post is the practical complement: here's how to build your own from scratch.

My Setup After Six Months

My personal LLM wiki lives in D:/projects/personal/knowledge/obsidian/wiki/ and has four directories:

raw/ — original sources (articles, transcripts, screenshots). Immutable, never edited. Around 200 files.
wiki/ — curated wiki pages. 35 pages. These are the ones I actually reread.
schema/ — 4 schema files defining page structures for different content types: people, concepts, tools, originals.
indexes/ — 2 index files I regenerate when I add new pages.

The LLM (Claude, usually via Claude Code) reads from raw/, writes new wiki pages following the relevant schema, and I review before saving to wiki/. I do this for maybe 20 minutes on weekday evenings when something was interesting enough to warrant a page.

35 pages after six months sounds small. The value isn't in the count.

Why My Previous Notes Systems All Failed Me

I've tried every note-taking setup over the past five years. Bear, Notion, Logseq, plain markdown, Obsidian with no structure. They all broke down the same way: I'd take notes in the context of what I was reading, file them somewhere organized, and never look at them again.

The problem is a mismatch between creation context and retrieval context. I write notes because a specific article triggered a thought. I want to retrieve them because a specific problem needs an answer — which is almost never the same context.

My personal LLM wiki fixes this because the schema forces the LLM to extract structured information that's useful across contexts. My concept/ schema has a related_concepts field that links each page to 3-5 related ideas. When I ask Claude to write a new wiki page, it populates those links based on what else exists in the wiki. Over time, the wiki becomes a web, not a pile.

That cross-linking behavior is the single biggest thing I didn't understand before building one.

The Three Things That Make It Actually Personal

Your schema reflects your actual mental model.

I have a schema category called originals/ for hot takes and frameworks that are mine. No one else's personal LLM wiki has an originals/ category because no one else has the same specific opinions I do. The schema is where your thinking style gets encoded into the system.

Your raw/ layer is irreplaceable.

I once deleted a raw article after I'd processed it into a wiki page. Six months later I found myself disagreeing with something I'd written and couldn't locate the source to check. The raw layer is your audit trail. Keep it immutable and it stays permanently useful.

You correct the LLM and save those corrections.

Every time I edit a Claude-written wiki page, I'm implicitly teaching the schema what works for me. I've added notes to my schema files like "don't use bullet lists for the tldr on person pages — use a single paragraph instead." That's my preference, not a general best practice. Over time, the wiki starts to sound like you.

How to Build Yours This Weekend

This takes 3-4 hours for a working version. Here's the sequence that actually works:

Step 1: Create the folder structure (~15 min)

vault/
  wiki/
    raw/
    wiki/
    schema/
    indexes/

Step 2: Write one schema file (~30 min)

Start with a concept schema. Mine looks roughly like this:

Fields: name, category, tldr (1 paragraph), how_i_use_it, related_concepts (3-5), sources
Format: markdown with H2 for each field
Tone: first-person, 200-400 words total
Do not: use bullet lists in tldr

Keep it simple. You'll revise after the first five pages.

Step 3: Pick three pieces of raw material (~10 min)

Put them in raw/. These should be things you've already read and had thoughts about — not articles you bookmarked intending to read someday.

Step 4: Ask an LLM to write the first pages (~45 min)

Prompt: "Read this raw article and write a wiki page following this schema. The page should reflect my first-person perspective as someone who [your context]."

Review each page. If something's off, fix it and note why in the schema file.

Step 5: Build a small index (~20 min)

A simple concept-index.md with one line per page — title, tldr, tags. This is what you'll grep when you can't remember where something lives.

After the weekend you'll have 3 pages, a working schema, and a clear sense of what you want to track. That's enough to build from.

What a Personal LLM Wiki Won't Fix

It won't make you read more or think harder. If you're not engaging with interesting material regularly, you won't have raw input to process. This is a tool for people who already consume a lot and lose most of it.

It also won't replace your thinking. I've had Claude write wiki pages that were technically accurate but missed what was actually interesting about the source. Catching those misses is where the real learning happens — not in the pages Claude gets right.

And it won't scale to 500 pages without infrastructure work. At 35 pages, plain markdown plus grep still works fine. Above roughly 200 pages you'll probably want a database layer. I'm not there yet and don't plan to be for a while.

FAQ

What LLM works best for a personal wiki?

I use Claude, mostly through Claude Code, because I'm already in that environment. Any capable LLM works as long as it reliably follows structured instructions. The schema file matters more than which model you use.

Does a personal LLM wiki replace Notion or Roam?

Not really — they solve different problems. Notion and Roam help you manage tasks and projects. A personal LLM wiki is specifically for building a long-term knowledge base from things you read and think about. I use both.

How long until it becomes genuinely useful?

Mine started feeling useful around the 15-page mark, maybe two months in. Before that it was practice. The value compounds — 30 pages that link to each other is worth more than 30 isolated pages.

Do I need Obsidian specifically?

No. Obsidian works well because of its graph view and fast local search, but the personal LLM wiki pattern works with any plaintext markdown setup. I've seen people run similar setups in VS Code with no plugins.

Jim Liu is an independent developer in Sydney. He runs openaitoolshub.org and has been building in public since 2024.