LLM-Wiki — Workshop Introduction

The three patterns

Side-by-side · the mechanism

Time & space complexity

Variables: N = chunks (RAG) or sources/pages (wiki); L = avg chunk/page length; d = embedding dimension (~1536); K = top-K retrieved or pages read; T = triples; h = hops; LLM = forward-pass cost.

When to use which

Real-world hybrids

• Graph-RAG. A knowledge graph replaces the vector index. Retrieval is graph traversal rather than vector similarity. Wins when relationships matter more than text similarity.
• Wiki + RAG. Wiki for stable, repeatedly-queried concepts; RAG for the long tail of fresh material you haven't synthesized yet.
• KG + Wiki. KG holds the typed backbone; wiki holds prose explanations, open questions, contradictions.

The four operations

The wiki has a small, fixed vocabulary of things you ask the agent to do. Each is a verb that operates on the same folder.

The schema is the contract

CLAUDE.md (or AGENTS.md) is the file the agent reads first. It defines what counts as a thing, what a page looks like, and how the operations work. Here's the concept-page template from the starter:

# <Concept Name>

> <single-sentence summary that doubles as a tagline>

## What it is
<2–6 sentence explanation, plain English first>

## Why it matters
<the practical or theoretical consequence>

## Key ideas
- bullet
- bullet

## Related
- [[concepts/<other>]]
- [[entities/<paper-or-person>]]

## Sources
- [[sources/<file>]] — <one-line note>

## Continue reading
- **<short reader-facing label>** → [[<target-page>]]
- **<short reader-facing label>** → [[<target-page>]]

There are five page templates total — concept, entity, synthesis, comparison, question — each with its own shape. The schema also declares style rules (kebab-case filenames, ISO dates, no claim without a source), forbidden behaviors (don't edit sources, don't invent citations), and the four operations.

Editing the schema is the loop

After your first ingest, you'll notice things you'd have done differently — a page split where you'd have merged, the tone is off, a template doesn't fit your domain. Don't fix the wiki by hand. Edit CLAUDE.md instead, ask the agent to re-ingest, and the next pass produces what you wanted.

The schema is the contract; the wiki is its output. Iterating on the schema is how a generic template becomes your wiki.

What a page actually looks like

This is what an agent renders, page by page — clean markdown with a blockquote tagline, structured sections, cross-links, citations, and a navigational footer.

Every page follows this shape. The agent writes them; you read them; the cross-links form the graph you see in Obsidian.

The pattern's sweet spot

LLM-Wikis win in a specific regime: a corpus you actually care about (not too big), questions that recur (not one-off), and an audience that includes humans. Below are the use cases I've seen pay off, grouped by who they serve.

Personal

A topic you're learning, mastering, or thinking through over weeks or months.

Team / enterprise

The pattern as a context layer for an AI-augmented data or engineering team.

Scaling up — what if I have multiple topics?

The LLM-Wiki pattern doesn't care how many wikis you have. The choice is one of relationship: are your wikis independent, or are they sub-topics of a larger theme?

How sub-topic wikis look on disk

The most common case — one vault, multiple sub-wikis, with a hand-curated README.md at the top:

LLM Wiki/                       ← Obsidian vault root
├── .obsidian/                  ← vault config (colors, plugins)
├── README.md                   ← hand-curated hub: "what's here"
├── Ontology/                   ← sub-wiki 1
│   ├── CLAUDE.md
│   ├── sources/
│   └── wiki/
├── RAG/                        ← sub-wiki 2
│   ├── CLAUDE.md
│   ├── sources/
│   └── wiki/
└── Distributed-Systems/        ← sub-wiki 3
    ├── CLAUDE.md
    ├── sources/
    └── wiki/

Three ways to cross-link inside this structure:

• Basename match: [[retrieval-augmented-generation]] — Obsidian finds it anywhere in the vault.
• Explicit path: [[../RAG/wiki/concepts/retrieval-augmented-generation]] — unambiguous, robust against renames.
• Hub navigation: the top-level README.md lists each sub-wiki with a one-line description.

When NOT to use it

The wiki isn't always the right tool. Where it loses:

A worked example

Concretely: what does an ML team's "experiment log" wiki look like after a few months of use?

sources/
├── 2026-04-15-baseline-finetune.md
├── 2026-04-22-lr-sweep.md
├── 2026-05-03-data-cleaning.md
├── 2026-05-10-attention-ablation.md
├── 2026-05-18-eval-redesign.md
└── 2026-05-24-instruction-tuning.md

wiki/
├── concepts/
│   ├── learning-rate-tuning.md
│   ├── data-quality-effects.md
│   └── attention-head-pruning.md
├── entities/
│   └── datasets/training-set-v2.md
├── syntheses/
│   └── what-actually-moves-the-needle.md
├── comparisons/
│   └── full-ft-vs-lora.md
└── questions/
    └── why-did-attention-ablation-help.md

Instead of six disconnected experiment writeups in Notion that nobody re-reads, the team gets a navigable wiki where "what we've learned about X" is a single page with citations back to the underlying experiments. New hires read the wiki; experienced team members keep adding sources; the synthesis is always current.

You can start in five minutes

A working LLM-Wiki needs three things: a markdown editor (for reading), an LLM agent (for writing), and the template repo (for the schema). All three are free and take a couple of minutes each.

Prerequisites

The five steps

git clone https://github.com/hong-chu/llm-wiki-starter.git my-wiki
cd my-wiki

What you'll have at the end

A folder you own, full of markdown pages an agent wrote for you, cross-linked into a navigable graph you can browse in Obsidian. You add sources over time; the wiki compounds. The schema co-evolves with your taste.

Specifically, after a few sessions:

• A working CLAUDE.md tuned to your domain
• 10–50 wiki pages covering concepts, entities, syntheses, comparisons, and open questions
• A coloured knowledge graph in Obsidian that's readable at a glance
• A dated change log proving every claim traces back to a source
• Portable, vendor-free, and entirely yours

Don't want to clone?

If you'd rather build a wiki from scratch without cloning, open Claude Code in an empty folder and paste this prompt:

Scaffold an LLM-wiki for the topic <YOUR TOPIC>, following the
pattern in Andrej Karpathy's gist
(https://gist.github.com/karpathy/442a6bf555914893e9891c11519de94f).
Create:

- CLAUDE.md (the schema, with page templates for concepts/entities/
  syntheses/comparisons/questions and ingest/query/lint workflows)
- An empty sources/ directory I'll fill in
- An empty wiki/ tree: concepts/, entities/, syntheses/, comparisons/,
  questions/, plus index.md, overview.md, log.md

Use kebab-case.md filenames, [[wiki-links]] for cross-references, and
end every page with a ## Sources section.

The agent will create the structure from scratch. The result should look like the directory tree from Tab 01.

Best practices

Disciplines that separate "neat trick" from "actual living wiki." These come from running the pattern repeatedly and watching what fails.

Pick the right topic

The single biggest predictor of wiki health is topic narrowness. Most "my wiki isn't working" stories trace back to a topic that's too broad.

Each good topic is a narrow slice of the corresponding bad one. Same row → same domain, different scope.

Ingest discipline

The first ingest sets the tone for everything that follows. Treat it like a calibration step, not the finish line.

Page hygiene — what every page should have

This list is exactly what lint checks. Internalize it and your wiki passes lint by default.

Schema is the loop

The single most-important discipline. Master this and the wiki maintains itself.

Workflow when a page disappoints:

1. Identify the rule that produced the issue (e.g., "the agent split this concept across two pages when it should be one")
2. Update CLAUDE.md to fix the rule
3. Ask the agent to re-ingest the relevant sources
4. Verify the new output matches your taste

The agent doesn't change CLAUDE.md on its own. That part stays human-initiated — the contract is yours.

Source rules

Lint cadence

Run lint weekly or after each major ingest. Two flavors of things it should catch:

Anti-patterns to avoid

The six most common ways people break their own wiki:

A deeper lens: ontology & VSM

Optional — for the curious. Two frameworks from outside the LLM world (philosophy and cybernetics) that sharpen how you think about what an LLM-Wiki is and what it's missing.

Ontology — five lenses on what's real

Ontology is the branch of philosophy that asks "what exists?". Five canonical answers have been proposed over the last 2,500 years. They aren't steps in a hierarchy — they're rival frames, each catching what the others miss.

The river metaphor: water (substance), flow (process), the network of banks and rain (relational), the pattern of differences (informational), the river as it shows up to someone on the bank (phenomenological). All five descriptions are true; none is sufficient alone.

VSM — six functions a system needs to stay viable

Stafford Beer (1926–2002) was a British cybernetician. He asked a different question: "what does any system — a cell, a team, a nation — need to stay alive in its environment?" His answer is the Viable System Model: six interacting functions. Drop any one and the system dies in a specific, recognizable way.

Applying VSM to LLM-Wiki

Run LLM-Wiki through the framework. What does it cover, and what does it miss?

What's missing — and how to add it

LLM-Wiki gives you S1–S3* out of the box. To build a fuller system, two layers need to be added.

The autonomy ladder — L0 to L4

Removing VSM layers from the top down produces a hierarchy of autonomy. A practical framing for evaluating any AI infra project, not just wikis:

Most "AI + notes" stacks today sit at L2–L3. LLM-Wiki out of the box is roughly L3: S1–S3* present, but no S4 watchers and an implicit S5. Adding watchers and an explicit identity.md lifts you to L4.

The composite architecture

Putting the lenses together: a complete Personal OS isn't LLM-Wiki alone. It's a composition of four pieces, each picked at its strongest.

• LLM-Wiki at the core — synthesized knowledge that compounds (covers S1, S2, S3, S3*)
• An agent runtime around it (OpenClaw-style, with heartbeats and webhooks) — adds S4
• An explicit identity.md on top — makes S5 first-class
• The human in the loop — the phenomenological lens: taste, refusal, weekly review

Where to dig deeper

If this lens resonates, the worked example is the Ontology vault this page was generated from. It contains real pages on:

• concepts/ontology — the five lenses, deepened
• concepts/viable-system-model — Beer's framework with the kitchen analogy and the autonomy ladder
• concepts/triad — the human-AI-fossil architecture
• concepts/personal-os — the composite stack
• syntheses/personal-os-composite-architecture — the full proposal
• questions/what-belongs-in-identity-md — the open thread on S5

Read those pages, then come back to your own wiki and ask: what's my S5? What's my S4? Which lens am I underusing? Those answers sharpen the pattern from "neat trick" to "real architecture."