LLM Research Wiki

A personal, LLM-maintained knowledge base for academic research, built on plain markdown files and operated via Claude Code.

Origin

This project implements Andrej Karpathy's LLM Wiki pattern (April 2026) — the idea that instead of using RAG to re-derive knowledge from raw documents on every query, an LLM should incrementally build and maintain a persistent, structured wiki that compounds over time.

The adaptation was developed by Paulo de Assis in conversation with Claude Code (for implementation, schema authoring, and the full wiki build).

1. Karpathy's post — the core pattern: raw sources → LLM-maintained wiki → structured queries
2. Design exploration — adapting the pattern to academic research: domain-specific page types, ingest/query/lint workflows, folder conventions
3. Implementation via Claude Code — building the full system: CLAUDE.md schema with six page templates and three workflows, folder structure, first ingest producing 38 wiki pages in a single pass

Template vs. live wiki

This repository is the template — the unspecialised seed. The CLAUDE.md here ships with placeholder Domain Context ("Customize this section") and an empty raw/. Fork it and make it yours.

My own working wiki — the one I actually ingest into daily — is a separate version specialised to my research (Continental philosophy of science, new music studies, posthuman music). On top of this template it adds a populated Domain Context (Deleuze, Simondon, Bachelard, DeLanda, Ferneyhough, Rheinberger…), pre-named project subfolders for active research and writing projects, and a few editorial conventions specific to my workflow. The schema, workflows, and page templates are the same as what you find here.

A note on related systems

Two adjacent systems in my own setup share this repository's DNA but serve different ends, and they sometimes get confused with the research wiki:

• MetamusicX wiki — a Quartz site built from research-team meeting transcripts. Same Karpathy pattern, but the raw layer is spoken conversation and the synthesis layer is atomic entities (composers, philosophers, concepts, threads) extracted by a dedicated agent.
• Alluvium — a journal-to-atomic-notes pipeline organised around PARA rather than concepts/authors. Same idea (immutable raw, curated atoms, append-only log) applied to daily voice and text journals.

These are mentioned only so forkers understand the scope of this template: it is the academic-research variant. Meetings, journals, and other input types live in separate systems with their own schemas.

What it does

Every time a source is ingested, Claude reads it, extracts key claims and quotes, and writes or updates markdown pages across the wiki — concept pages, author pages, debate pages, synthesis pages, and source notes. The raw documents are never modified. The wiki is the living layer that accumulates knowledge across all ingested sources.

Over time, the wiki becomes smarter than your memory about connections across your reading.

Architecture

Three layers:

Workflows

The operational logic lives in CLAUDE.md. Claude reads it at the start of every session and acts as a dedicated research intelligence agent. Three core workflows:

INGEST

Drop a source into raw/, then say "ingest [filename]". Claude reads it, discusses key takeaways, writes a source note, and updates every concept/author/debate/project page touched by the source. A single ingest typically creates or updates 10–15 wiki pages. Everything is logged in log.md.

QUERY

Ask any research question. Claude checks index.md first — identifying the relevant cluster, then reading the synthesis page for that cluster if one exists, then following the related: field on each page to navigate to adjacent concepts. This three-step cascade (cluster → synthesis → related) dramatically reduces the number of pages read per query as the wiki grows.

LINT

Say "lint" to audit the wiki for duplicates, contradictions, orphan pages, stale content, concepts mentioned but lacking pages, and thin source support. Results are presented as a prioritized issues list. Nothing is auto-fixed.

Folder structure

raw/
  articles/   books/   chapters/   notes/   annotations/
  transcripts/   images/   _staging/
wiki/
  concepts/   authors/   debates/   themes/   methods/
  syntheses/   source-notes/   projects/
outputs/
  essays/   slides/   handouts/   tables/
conversations/
archive/

Page types

Six page templates are defined in CLAUDE.md, each with YAML frontmatter:

Navigation design (v2)

After ~10 sources and 50+ pages, a flat alphabetical index becomes slow to navigate. The system uses a three-layer navigation cascade to keep query cost constant as the wiki grows:

1. Concept clusters

index.md organizes concepts into thematic clusters (4–6 per domain) instead of one flat alphabetical list. Each cluster has a one-sentence description, its core pages, a pointer to any synthesis page, and the key authors. A query identifies the relevant cluster(s) first.

2. Synthesis pages (wiki/syntheses/)

Synthesis pages are the "inner grooves" of the wiki: pre-digested argumentative overviews across a cluster of related pages. When a synthesis page exists for a cluster, a query reads it first — one page instead of six. Syntheses are created when a cluster has enough source support to warrant a standing overview.

3. related: YAML field

Every concept and author page carries a related: frontmatter field listing 3–5 of its closest neighbors (by filename stem). After reading one page, the agent uses related: to navigate to the next most relevant pages without re-scanning the index.

---
title: "Apparatus / Dispositif"
type: concept
tags: [apparatus, Foucault, Barad, Stiegler]
related: [archive-foucault, distribution-of-the-sensible, program-industries, agency, assemblage]
created: 2026-04-06
updated: 2026-04-10
---

The cascade: cluster → synthesis → related fields → individual pages. Each step narrows the reading set. At 100+ pages, this keeps query cost roughly constant.

Getting started

1. Clone this repo
2. Open the folder in Claude Code (terminal or VS Code)
3. Edit CLAUDE.md to add your own domain context — your research areas, key thinkers, core concepts
4. Write a short research map in your own words (2–3 pages) and save it in raw/notes/
5. Say "ingest raw/notes/your-research-map.md"
6. Watch the wiki populate with concept pages, author pages, and debate pages from your first source
7. Add real sources one at a time. Supervise the first 5–10 ingests closely.

Conventions

• Filenames: lowercase-kebab-case (simondon-transduction.md)
• YAML frontmatter: Every page starts with title, type, tags, related (concepts/authors), created, updated
• Wiki links: Relative markdown links ([Transduction](../concepts/transduction.md))
• Log everything: Every ingest and change is appended to log.md
• raw/ is immutable: Source files are never modified after ingest
• The wiki answers questions, not the raw sources: Build the wiki so it contains what you need

Scaling advice

• Start with your own research map as the first ingest — it seeds the wiki with YOUR conceptual framework
• Ingest sources one at a time for the first 10–15 sources — supervise quality
• Don't try to ingest your entire library — only ingest what matters to your active projects
• Use raw/_staging/ as a holding pen for candidates — review weekly, promote or remove
• Run lint every 10–15 ingests to keep the wiki healthy
• Create your first synthesis page when a cluster has 4+ sources and keeps coming up in queries
• At ~50 sources the wiki becomes a genuine research tool; at ~100 it's indispensable

Requirements

• Claude Code (terminal or desktop app)
• The CLAUDE.md file in the root folder — this is what makes Claude a wiki agent
• No database, no embeddings, no plugins — just markdown files and folders

Credits

• Pattern: Andrej Karpathy, "LLM Wiki" (April 2026)
• Adaptation and implementation: Paulo de Assis
• Schema authoring and build: Claude Code (Anthropic)

License

MIT