LLM-Wiki — Concept Test

A hands-on implementation of the LLM-Wiki pattern described by Andrej Karpathy. Applied here to Alice's Adventures in Wonderland (Ch. I–III) as a proof of concept.

What is an LLM-Wiki?

An LLM-Wiki is a persistent, LLM-maintained knowledge base built from raw source documents. The key principle: the LLM does all the maintenance work. The human provides sources and asks questions.

Rather than re-deriving knowledge on every query, the LLM builds up a structured wiki of markdown pages and keeps them incrementally updated. The wiki becomes richer with every ingest and query operation.

Original concept: Andrej Karpathy's gist

---

Architecture

Three Layers

sources/    ← Raw documents (immutable, human-curated)
    ↓
wiki/       ← LLM-maintained markdown pages (never edit manually)
    ↓
SCHEMA.md   ← Rules and conventions governing the LLM's work

Key Files

File	Role
SCHEMA.md	The LLM's instruction manual: page types, frontmatter, naming, operations
index.md	Auto-maintained table of contents, organized by category
log.md	Append-only chronological log of all operations

Wiki Structure

wiki/
  concepts/   ← Abstract ideas, themes, theories
  entities/   ← Persons, places, works, organizations

Every wiki page begins with YAML frontmatter:

---
title: 
type: concept | entity
entity_type: person | place | work | organization
sources: []
tags: []
created: YYYY-MM-DD
updated: YYYY-MM-DD
related: []
---

---

Operations

Three operations are defined as local Claude Code slash commands (.claude/commands/):

/ingest <filename>

Process a new source document. The LLM reads the source, identifies concepts and entities, creates or updates 5–15 wiki pages, updates index.md, and appends to log.md.

/query <question>

Answer a question using only the wiki as the knowledge base. The LLM reads index.md, opens relevant pages, and synthesizes an answer with citations. If the answer reveals a new insight not yet in the wiki, a new page is created.

/lint

Health-check the wiki: orphaned pages, missing cross-references, incomplete frontmatter, contradictions, and broken index links. Writes a report to lint-report.md.

---

Current State (Alice in Wonderland, Ch. I–III)

Source

sources/alice-ch1-3.txt — Project Gutenberg plain text, chapters I–III (633 lines)

Wiki (14 pages)

Concepts (5)

• Caucus Race — satirical ruleless race; everyone wins
• Identity and Self — Alice's crisis of self as she transforms
• Innocence as Disruption — Alice's naivety causes unintended chaos (emerged from QUERY)
• Logic and Nonsense — Carroll's core technique
• Transformation and Size — Drink Me / Eat Me mechanics

Entities — Persons (6)

• Alice, Dinah, The Dodo, The Lory, The Mouse, The White Rabbit

Entities — Places (3)

• The Garden, Hall of Doors, Pool of Tears

Operation Log

[2026-04-21 16:45] INGEST sources/alice-ch1-3.txt — 9 entity pages, 4 concept pages
[2026-04-21 18:00] QUERY "Welchen Sinn hat Dinah?" — new concept: innocence-as-disruption

---

Key Observations

What works well

• The SCHEMA.md reliably guides the LLM to produce consistent, well-linked pages
• QUERY operations can discover concepts not extracted during INGEST
• Git history serves as a second audit trail alongside log.md
• Cross-references between pages emerge naturally

Known limitations

• Cold-start retrieval is slow: a fresh Claude session must read SCHEMA + index + pages before answering
• Mitigation without code: keep sessions open; enrich index.md with more metadata so fewer pages need to be opened
• At scale (100+ pages), a vector index would replace the LLM's manual index scan

Design decisions

• English throughout (better embedding model coverage for future retrieval)
• YAML frontmatter on all pages (machine-readable metadata)
• Slash commands are project-local (.claude/commands/) — nothing global
• Sources are tracked in git alongside the wiki (simplicity over size optimization)