skills/llm-wiki/SKILL.md

name	description	allowed-tools
llm-wiki	Karpathy-pattern LLM knowledge base. Builds and maintains a persistent wiki of structured, interlinked markdown files using Claude Code — fundamentally different from RAG. Instead of re-deriving answers from raw documents at query time, the LLM incrementally builds the wiki: reading sources, extracting concepts, updating entity pages, and flagging contradictions. Uses Obsidian as the visual frontend. Trigger when user says: - "set up my wiki" / "create a knowledge base" / "build a second brain" - "set up an LLM wiki about [topic]" - "ingest this article / URL / transcript / video" - "add this to my wiki" - "query my wiki" / "what does my wiki say about" - "lint the wiki" / "health check my wiki" - "connect my notes on [topic]" - "Karpathy wiki" / "LLM wiki"	Bash(firecrawl *) Bash(mkdir *) Bash(git *) Bash(grep *) Bash(find *) Write Edit Read Glob

LLM Wiki — Karpathy Pattern

A persistent, compounding knowledge base maintained entirely by Claude Code. Obsidian is the IDE. Claude is the programmer. The wiki is the codebase.

---

Core Concept

RAG re-derives knowledge on every query. The LLM has to find and piece together fragments each time. Nothing accumulates.

This wiki pre-compiles knowledge. When a source is ingested, Claude reads it, extracts key information, and integrates it into the existing wiki — updating entity pages, noting contradictions, adding cross-references. Ask a question later and the synthesis is already there.

The wiki is a persistent, compounding artifact. It gets richer with every source ingested and every question asked.

---

Three-Layer Architecture

raw/            ← immutable source documents (user controls, LLM reads only)
wiki/           ← LLM-generated markdown (LLM owns, user reads)
CLAUDE.md       ← schema (tells Claude how to behave in this wiki)

---

Operations

Read the phase files for full protocol:

Operation	Trigger	Phase File
Setup	"set up a wiki about X"	phases/setup.md
Ingest	"ingest this", "add this article"	phases/ingest.md
Query	"what does my wiki say about X"	phases/query.md
Lint	"lint the wiki", "health check"	phases/lint.md

---

Setup (Quick Reference)

When asked to set up a new wiki:

# 1. Create vault structure
mkdir -p ~/wikis/<topic>/{raw/assets,wiki/{sources,concepts,entities},tools}
cd ~/wikis/<topic>
git init
printf ".obsidian/workspace.json\n.DS_Store\n.firecrawl/\n" >> .gitignore

Then write these files (use templates):

• CLAUDE.md — from templates/CLAUDE.md.template
• wiki/index.md — from templates/index.md.template
• wiki/log.md — from templates/log.md.template
• wiki/overview.md — from templates/overview.md.template

Then open the vault:

obsidian ~/wikis/<topic>

---

Ingest (Quick Reference)

When asked to ingest a source:

# If URL: scrape it first
firecrawl scrape "<url>" -o raw/<slug>.md

# If file already in raw/: proceed directly

Then:

1. Read wiki/index.md to find related existing pages
2. Write wiki/sources/<slug>.md — structured summary
3. Update up to 15 related concept/entity pages
4. Update wiki/index.md — add new entry
5. Append to wiki/log.md

---

Query (Quick Reference)

When asked a question about wiki content:

1. Read wiki/index.md — identify relevant pages by scanning summaries
2. Read 3-7 most relevant pages (parallel Read calls)
3. Synthesize answer with [[WikiLink]] citations
4. Offer to file answer as wiki/queries/<slug>.md

---

Log Format

Always use this prefix for log entries (enables grep parsing):

## [YYYY-MM-DD] ingest | Source Title
## [YYYY-MM-DD] query | Question asked
## [YYYY-MM-DD] lint | Health check pass N

Quick log inspection:

grep "^## \[" wiki/log.md | tail -10
grep "ingest" wiki/log.md | wc -l

---

Search (when wiki grows large)

# Simple grep-based search
bash tools/search.sh "transformer architecture"

# Or use qmd for semantic search (install if needed)
# npm install -g qmd
# qmd search "transformer architecture" wiki/