LLM Knowledge Base

LLM Knowledge Base

Build a personal knowledge base based on Andrej Karpathy's LLM Wiki pattern. Keep raw sources unchanged. Incrementally maintain a linked Markdown wiki. Save useful analyses as durable outputs.

Original pattern: https://gist.github.com/karpathy/442a6bf555914893e9891c11519de94f

Confirm the setup

Before writing files, establish:

1. The target directory. Default to the current working directory only when that is clearly the intended project.
2. The topics or domains the knowledge base should cover.
3. Whether existing files should be included as raw sources.
4. The host agent. Use AGENTS.md for Codex and CLAUDE.md for Claude Code. For another agent, use its documented project-instruction filename.

Ask only for missing information. Do not infer personal domains from unrelated context.

Inspect the target before editing. If files or directories already exist, preserve them and explain any conflicts. Do not overwrite an existing agent instruction file without the user's approval.

Create the structure

Create only the directories that are useful for the user's sources:

raw/
  articles/
  papers/
  notes/
  images/
wiki/
  _templates/
  index.md
  log.md
outputs/
AGENTS.md or CLAUDE.md

Use .gitkeep files when the user wants empty directories tracked by Git.

Create the agent instructions

Read assets/agent-guidelines-template.md. Adapt it to the chosen domains and instruction filename, then write it to the target directory.

Keep paths project-relative so the knowledge base remains portable. Preserve any existing project instructions unless the user explicitly approves a replacement. If an instruction file already exists, propose a scoped addition instead of silently replacing it.

Create the core files

Create wiki/index.md:

---
type: index
updated: YYYY-MM-DD
---

# Wiki Index

## By Domain

### Domain name

## Recently Added

Create wiki/log.md:

---
type: log
---

# Operations Log

## [YYYY-MM-DD] init | Knowledge base created

- Created the raw, wiki, and outputs layers.
- Added the agent operating instructions.
- Based on Andrej Karpathy's LLM Wiki pattern.

Create wiki/_templates/entity.md:

---
type:
domain:
created: YYYY-MM-DD
updated: YYYY-MM-DD
sources: []
tags: []
---

# Title

## Summary

## Key Details

## Related

## Sources

Use real dates when creating files. Use the user's chosen domains in the index.

Handle existing content safely

If the user wants existing content included, copy it into the appropriate raw/ directory by default. Preserve the originals. Move or delete source files only with explicit approval.

Do not ingest secrets, credentials, private keys, or material the user is not allowed to copy. Flag likely sensitive files before copying them.

After copying approved sources, process them one at a time unless the user requests a batch:

1. Read the source fully.
2. Summarize the important claims and provenance.
3. Create or update relevant wiki pages.
4. Add cross-references.
5. Update wiki/index.md.
6. Append an entry to wiki/log.md.

Prefer updating a relevant page over creating a near-duplicate.

Explain the operating loop

Give the user three simple commands or prompts:

• Ingest: Process a named file in raw/. Update the wiki, index, and log.
• Query: Answer from the wiki and cite the relevant wiki pages and raw sources.
• Lint: Find contradictions, stale claims, missing citations, duplicate topics, orphan pages, and broken links.

Recommend a lint pass after a meaningful batch of ingests or when the wiki starts to feel inconsistent. Do not claim that monthly is the right schedule for every knowledge base.

Maintain provenance

• Treat raw/ as immutable unless the user explicitly asks to correct a source file.
• Cite raw sources from derived wiki pages.
• Distinguish source claims from agent inference.
• Record contradictions instead of silently choosing one claim.
• Include dates on time-sensitive claims.
• Never present an agent-generated synthesis as a primary source.

Adapt to the user's tools

The pattern works best with an agent that has persistent filesystem access.

• For Codex, use AGENTS.md.
• For Claude Code, use CLAUDE.md.
• For Obsidian, keep [[wiki-links]] and optional YAML frontmatter.
• For editors without wiki-link support, use standard Markdown links if the user prefers them.
• For chat products that only support file uploads, explain that updates and file organization will be manual between sessions.

Do not require Obsidian, Dataview, a vector database, or external command-line tools. Offer them only when they solve a stated need.

Finish with verification

Before reporting completion:

1. Confirm the expected directories and core files exist.
2. Confirm the instruction filename matches the selected agent.
3. Check that no existing files were overwritten unexpectedly.
4. Check that every generated wiki page cites at least one source or is clearly labeled as an inference.
5. Summarize what was created and give one concrete first-ingest prompt.