llm-wiki

A Claude Code plugin for building and maintaining persistent, LLM-maintained knowledge bases in Obsidian format. One install, four skills, and any directory becomes a compounding wiki the LLM writes for you.

Based on Andrej Karpathy's LLM Wiki, extended with scaling patterns from agentmemory. Full acknowledgements in origin/CREDITS.md.

Status. Early, not yet on a public Claude Code marketplace. Install by cloning — see Install.

The idea in one line

Stop re-deriving knowledge with RAG every time; have the LLM compile it into a structured wiki that compounds over time. You curate sources and ask questions. The LLM does the rest — summarising, cross-referencing, flagging contradictions, keeping everything consistent.

Obsidian is the IDE. The LLM is the programmer. The wiki is the codebase.

For the long version, read origin/llm-wiki.md (the original pattern) and origin/llm-wiki-v2.md (scaling extensions).

What you get

Four skills, each triggered by natural language:

Skill	Say something like…	What it does
wiki-bootstrap	"set up a new knowledge base here" · "build me a second brain" · "start a research vault"	Inspects the directory, interviews you about use case and domain, scaffolds raw/ + wiki/, writes a tailored CLAUDE.md schema, seeds index.md / log.md / overview.md, configures .obsidian/. One-shot per vault.
wiki-ingest	"ingest this" · "add this article" · "file this paper" · "process this meeting"	Reads one source, writes a summary page in wiki/sources/, updates entity & concept pages, updates the index, appends to the log. Flags contradictions; does not silently overwrite existing claims.
wiki-query	"what does the wiki say about X?" · "query the wiki" · "compare A and B from the vault"	Searches the wiki, synthesises an answer with wikilink citations, picks the right output format (prose, table, timeline, mermaid). Offers to file substantive answers as wiki/analysis/ pages so explorations compound.
wiki-lint	"lint the wiki" · "health check" · "clean up the wiki"	Runs seven checks: frontmatter drift, broken wikilinks, orphans, missing cross-references, contradictions, index freshness, overview staleness. Auto-fixes safe issues, flags the rest for you to decide.

All four skills read the vault's CLAUDE.md at runtime, so the process is plugin-level but the schema is per-vault — each vault's conventions, entity types, and page shapes travel with it.

Capabilities

• Any domain. Personal (goals, journal, self-improvement), research (topic deep-dives), book companions, business/team wikis, course notes, competitive analysis, trip planning — bootstrap interviews you and tailors the schema.
• Obsidian-native output. Wikilinks, callouts, YAML frontmatter, mermaid diagrams. Configured so Graph View and Dataview work out of the box.
• Confidence scoring. Every page carries a 0.0–1.0 confidence value; skills weigh older/newer claims and surface contradictions rather than overwriting.
• Chronological log. Append-only wiki/log.md is grep-parseable: grep "^## \[" wiki/log.md gives you the full timeline.
• Safe defaults. Bootstrap never overwrites silently; ingest never modifies raw/; lint never resolves contradictions unilaterally; nothing commits to git unless you ask.

Prerequisites

• Claude Code (CLI, IDE extension, or desktop).
• Obsidian for browsing the vault (optional — everything is plain markdown).

Install

Clone the repo once, then register it as a local marketplace in Claude Code.

# 1. Clone anywhere on your machine
git clone https://github.com/bmentges/knowledge-base-template ~/code/llm-wiki
cd ~/code/llm-wiki

Then, inside Claude Code:

/plugin marketplace add ~/code/llm-wiki
/plugin install llm-wiki@llm-wiki

Verify the skills are loaded — start a fresh Claude Code session and run:

/plugin

You should see llm-wiki listed as installed, with four skills: wiki-bootstrap, wiki-ingest, wiki-query, wiki-lint.

Updating

cd ~/code/llm-wiki
git pull

Then in Claude Code: /plugin marketplace update llm-wiki.

Uninstall

/plugin uninstall llm-wiki@llm-wiki
/plugin marketplace remove llm-wiki

Fallback (if the plugin flow doesn't work yet)

The skills are plain markdown files and work as user-level skills too. Symlink them in:

mkdir -p ~/.claude/skills
for s in wiki-bootstrap wiki-ingest wiki-query wiki-lint; do
  ln -s ~/code/llm-wiki/skills/$s ~/.claude/skills/$s
done

Skills are then available in every Claude Code session, no plugin commands needed.

Use

1. Create a new wiki

In the directory you want the vault to live:

Claude, set up a new knowledge base here for my research on distributed systems.

wiki-bootstrap will interview you (use case, entity types, source types, scale), show the plan, then scaffold the vault. If the directory isn't empty, it walks you through placement options — bootstrap in-place, use a subdirectory, or pick a different directory. It never overwrites existing files silently.

2. Open in Obsidian

Obsidian → Open folder as vault → select the directory

3. Daily workflow

# Drop an article into raw/, then:
Claude, ingest raw/2026-04-17-paxos-explained.md

# Later, ask:
Claude, what does the wiki say about consensus algorithms?

# Periodically:
Claude, lint the wiki.

What a vault looks like

my-knowledge-base/
|
|-- raw/                        # Your source documents (immutable)
|   |-- assets/                 # Images, PDFs, attachments
|   |-- 2026-04-17-paxos-explained.md
|   +-- meeting-notes-2026-04.md
|
|-- wiki/                       # LLM-generated pages (the knowledge base)
|   |-- index.md                # Catalog of all pages with summaries
|   |-- log.md                  # Chronological record of operations
|   |-- overview.md             # High-level synthesis
|   |-- entities/               # People, projects, tools, orgs
|   |-- concepts/               # Ideas, patterns, theories
|   |-- sources/                # One summary page per ingested source
|   +-- analysis/               # Filed query results, comparisons, deep dives
|
|-- CLAUDE.md                   # Schema: rules, conventions, workflows
+-- .obsidian/                  # Obsidian vault config

How the pieces connect

graph TD
    YOU([You]) -->|drop sources| RAW[raw/]
    YOU -->|ask questions| CLAUDE([Claude + skills])
    CLAUDE -->|reads| RAW
    CLAUDE -->|reads & writes| WIKI[wiki/]
    CLAUDE -->|follows rules from| SCHEMA[CLAUDE.md]
    YOU -->|browses| OBSIDIAN([Obsidian])
    OBSIDIAN -->|renders| WIKI

    subgraph Three_Layers
        RAW
        WIKI
        SCHEMA
    end

    subgraph Skills
        BOOTSTRAP[wiki-bootstrap]
        INGEST[wiki-ingest]
        QUERY[wiki-query]
        LINT[wiki-lint]
    end

    CLAUDE --> BOOTSTRAP
    CLAUDE --> INGEST
    CLAUDE --> QUERY
    CLAUDE --> LINT

Loading

Plugin layout

.claude-plugin/
  plugin.json                   Plugin manifest
  marketplace.json              Single-plugin marketplace (for local install)
skills/
  wiki-bootstrap/               SKILL.md + templates/
  wiki-ingest/                  SKILL.md
  wiki-query/                   SKILL.md
  wiki-lint/                    SKILL.md
origin/                         Pattern docs this plugin is based on
  llm-wiki.md                   Karpathy's original idea file
  llm-wiki-v2.md                agentmemory scaling extensions
  obsidian-conventions.md       Formatting/frontmatter/wikilink rules
  CREDITS.md                    Acknowledgements
README.md
LICENSE

Scaling up

origin/llm-wiki-v2.md covers advanced patterns worth adopting as your wiki grows:

• Confidence scoring — already part of the default schema.
• Knowledge graph — typed entities and relationships beyond flat wikilinks.
• Hybrid search — BM25 + vector + graph traversal when index.md outgrows a single read.
• Automation hooks — auto-ingest on new files in raw/, scheduled lint. Not yet shipped in this plugin; tracked as future work.
• Consolidation tiers — working → episodic → semantic → procedural memory.
• Multi-agent / shared vaults — several contributors on one wiki.

Start minimal. Add layers when you feel the need.

Recommended Obsidian plugins

• Graph View (built-in) — visualize connections between pages.
• Dataview — query page frontmatter as structured data.
• Obsidian Web Clipper — browser extension to clip articles as markdown sources.
• Marp Slides — generate presentations from wiki content.

Roadmap

• Hooks: auto-ingest on new files in raw/, scheduled lint.
• Optional search backend (e.g. qmd) for vaults past a few hundred pages.
• Publish to a public Claude Code marketplace when one stabilises.

License

MIT

Credits

See origin/CREDITS.md.