Complete Guide
The LLM Knowledge Base System
By MrPrompts
Turn raw articles, notes, and research into a structured wiki that an AI can read, query, and build on. This guide covers every step, from an empty folder to a system that gets smarter each time you use it.
How it works in one paragraph
You collect raw sources. An AI compiles them into a structured wiki. You query the wiki. Good answers get filed back in. The knowledge compounds. Andrej Karpathy described this as one of the most underrated uses of LLMs. He does it with custom scripts. This guide shows you how to do it without writing any code.
How AI Knowledge Bases Work
01
Collect
Drop sources into a folder
02
Compile
AI writes structured articles
03
Query
Ask questions, get cited answers
04
Compound
Knowledge grows over time
The problem this solves
Every AI conversation starts from scratch. The AI does not remember what you discussed yesterday. The research you did last week? Gone. That article you read last month? It is in a browser tab you closed.
A knowledge base gives your AI persistent memory. You point it at a folder of structured articles, it reads them, and now every conversation builds on everything you have already collected. You stop repeating yourself. The AI stops giving you generic answers.
After a month, the wiki knows your topic better than any single conversation could. After three months, you have a reference nobody else has because it was built from your sources, for your questions.
Step 1: Set up your workspace
Two things. A place to store files and an AI that can read them.
For file storage, use whatever you already have open:
- Notion if you want collaboration. Create a new workspace.
- Google Docs if you want the simplest option. A folder in Drive works fine.
- A folder on your computer with text files. Nothing fancy needed.
- Obsidian if you want bidirectional linking. Steeper learning curve.
For the AI, anything that can read files:
- Claude handles long documents well. I use it for most wiki work.
- ChatGPT works. Upload files or paste content.
- Gemini, Perplexity, local models. All fine. The system is the same regardless of which AI you pick.
Create three folders:
Knowledge Base Folder Structure
Step 2: Write your CLAUDE.md
CLAUDE.md is the instruction file for the whole system. It tells the AI your topic, your folder structure, how articles should be formatted, and what rules to follow. The AI reads it first, every time. Without it, you get inconsistent output and have to repeat yourself constantly.
Create a file called CLAUDE.md in the root of your workspace and paste this template. Replace the brackets with your topic:
# [Your Topic] Knowledge Base # Version: 1.0 ## Purpose A structured knowledge base on [your topic], compiled from curated sources and maintained by AI. ## Structure - /Sources — Raw articles, notes, and reference material (NEVER edited by AI) - /Wiki — Compiled, interlinked wiki articles (generated and maintained by AI) - /Queries — Questions asked and answers generated (good answers get promoted to wiki) ## Article Format Every wiki article must follow this structure: 1. Title as H1 2. Summary (2-3 sentences that stand alone) 3. Body organized with H2 sections 4. Key Takeaways (3-5 bullet points) 5. Sources section (which source files informed this) 6. Related section (links to connected articles) ## Taxonomy Categories: [define 4-6 top-level categories] Tags: [define common tags] Naming: Title-Case-With-Hyphens.md ## Rules 1. NEVER modify files in /Sources 2. Always cite which source(s) informed each article 3. Link related concepts across articles 4. One concept per page 5. Every claim must trace back to a source 6. Add "Last updated" date to every article 7. New articles must link to at least 2 existing articles 8. Keep summaries self-contained
The CLAUDE.md deep dive explains every section if you want the details. The Masterclass has schemas for specific use cases.
Step 3: Add your first sources
Drop 3 to 5 source documents into the Sources folder. One file per source. Paste the full text, not just a link. Name each file with a date and short title: 2026-04-12-Karpathy-Knowledge-Bases.md
Sources can be articles you read, papers, your own meeting notes, documentation, interview transcripts, market research. Anything relevant.
Five good sources beat fifty mediocre ones. Start with the best things you have read on your topic. You will add more over time.
Step 4: Compile your first wiki articles
Open your AI. Give it the CLAUDE.md file first. Then give it your source files. Then use this prompt:
Read the CLAUDE.md schema. Then read all the source files. 1. Identify the 5-8 most important concepts across all sources. 2. For each concept, write a wiki article following the format specified in CLAUDE.md. 3. Link related articles to each other. 4. Create an index file listing all articles by category. Output each article as a separate section so I can save them individually.
You will get structured wiki articles with summaries, body sections, takeaways, citations, and cross-links. Review them for anything that looks off. Save each one into the Wiki folder.
That is a working knowledge base. About 20 minutes from start to here. Everything after this point makes it better.
Step 5: Start querying it
Now instead of asking your AI vague questions, you ask it questions against the wiki. Give it the CLAUDE.md and your wiki articles, then ask what you actually want to know:
Read the CLAUDE.md schema and all wiki articles. Based on the wiki, what are the three main approaches to [topic]? Which approach has the strongest evidence? What are the trade-offs between them? Cite which wiki articles support each point.
The difference between this and a normal AI conversation is striking. The AI draws from your curated sources, organized by your schema, following your rules. It is synthesizing from your data, not guessing from its training data.
When you get an answer that teaches you something new, save it to the Queries folder. If it is good enough, promote it to a wiki article. The queries feed back into the system.
Step 6: Keep feeding it
When you read something relevant, drop it into Sources. One file, 30 seconds. Once a week, or whenever you have added a few new sources, run this:
Read the CLAUDE.md schema, all wiki articles, and the following new source files: [list them] 1. Do any new sources update or contradict existing wiki articles? If so, update those articles. 2. Do any new sources warrant entirely new wiki articles? If so, write them. 3. Update the index file. 4. Add links between new and existing articles.
Once a month, run a health check:
Read all wiki articles. Audit for: 1. Articles with no incoming links (orphans) 2. Claims without source citations 3. Contradictions between articles 4. Topics mentioned but never given their own article 5. Articles that are too short (under 300 words) 6. Outdated information (check dates) Report issues and suggest fixes.
A month in, you will have 15 to 25 interlinked articles. Three months, 40 to 60. At that point the wiki is the most complete reference on your topic that exists, because nobody else built one from your exact sources for your exact questions.
Step 7: Do things you could not do before
Once you have 20+ articles, some new things become possible.
You can ask the AI to compare what two different sources say about the same topic and tell you where they agree and disagree.
You can ask it to find gaps. "What questions can I not answer with the current wiki? What should I read next?"
You can ask for outputs. "Write a one-page summary of [topic] for my boss." Or "Create a slide outline for a 15-minute presentation." Or "Draft a blog post using only information from the wiki." The wiki becomes the source of truth for everything you produce on that topic.
If accuracy is critical (finance, legal, health), use a second AI model to check. Give it the wiki article and the cited sources. Ask: "Does this article accurately represent the source material?" When two different models agree, that is a strong signal. When they disagree, it is worth investigating.
People use this for all kinds of things
Why this works: what the research says
Andrej Karpathy wrote about using LLMs to build personal knowledge bases in early 2026. He said most of his AI token usage had shifted from writing code to managing knowledge stored as markdown files. He collects source documents, uses an LLM to compile a wiki, queries it for research, and files answers back in. He called it "an incredible new product waiting to be built."
But Karpathy is not the only one seeing this pattern. Google DeepMind's research on retrieval-augmented generation found that AI produces fewer hallucinations and more accurate answers when it references a curated knowledge base instead of relying on training data alone. The effect is significant enough that RAG has become the standard approach in production AI systems.
Anthropic's research on long-context performance tells a similar story. Structured reference documents consistently outperform ad-hoc instructions, sometimes by 40 to 60 percent on task quality. The CLAUDE.md schema applies this finding directly: instead of explaining your topic from scratch every session, you give the AI an organized reference it reads first.
The pattern across all of this research is the same: give AI organized, persistent context, and the output gets meaningfully better. The knowledge base system described in this guide is a practical implementation of that principle. Karpathy does it with custom scripts. Enterprise teams do it with vector databases. This guide shows you how to do it with a note-taking app and numbered prompts.
The MrPrompts Knowledge Base Starter Kit packages the full system: folder structure, CLAUDE.md schema, and 7 prompts that walk you through each step. No coding, no scripts, no special tools.
Get the starter kit
The complete folder structure, CLAUDE.md schema, and 7 numbered prompts. Customized to your topic. Works with any AI tool.
Get the Starter Kit (Free)Frequently asked questions
How long does it take to set up?
About 20 minutes to go from nothing to a working wiki with articles. Create the folders, write the CLAUDE.md, paste in a few sources, run the compile prompt. After that, plan on 15 to 20 minutes a week to add new sources and update the wiki.
Do I need Obsidian?
No. Karpathy uses it because he is a developer. Notion, Google Docs, Apple Notes, or a plain folder of text files all work. The knowledge architecture is what matters, not the tool it lives in.
How many sources before it gets useful?
Five to ten on a focused topic. You start seeing compound value when concept articles draw from multiple sources. At 20 to 30 sources the wiki starts answering questions you did not think to ask.
Does this replace RAG?
At small scale, it works like a personal RAG system without the infrastructure. Karpathy said he thought he needed fancy RAG but the LLM handled index files and document retrieval fine at the scale of a few hundred articles. For millions of documents, you would still want a proper pipeline.
What about hallucinations?
The CLAUDE.md rules require every claim to cite a source. If the AI writes something that is not in the sources, the citation rule catches it. For high-stakes work, use a second AI model to verify: give it the article and the cited sources and ask if they match.
Can a team share one wiki?
Yes. Use Notion or a shared drive so multiple people can add sources. The CLAUDE.md keeps output consistent regardless of who adds what. Sales teams, research groups, and leadership teams all do this.
Subscribe to the MrPrompts Newsletter
Join 5,000+ builders. One practical AI framework every week: prompt templates, workflow blueprints, and knowledge base strategies you can use the same day. Free.