Lemon Wiki (LLM Wiki)

A Lemon Wiki is a knowledge base built and maintained by LLM coding agents — not a static collection of pages someone typed, but a living system where AI agents research topics, ingest sources, compile synthesized articles, cross-reference them, and generate output artifacts. The name "Lemon Wiki" is the affectionate shorthand for "LLM Wiki."

This Buckyverse .wiki/ is a Lemon Wiki. Every article in it — the concept pages, the topic overviews, the reference registries, the playbooks — was compiled by an AI agent from raw sources using the llm-wiki plugin. The human decides what to research, what to ingest, and what to publish; the agent does the reading, synthesis, cross-referencing, and structural maintenance.

How It Works

A Lemon Wiki has a strict separation between raw material and compiled output, enforced by the tooling:

.wiki/
├── raw/              ← Immutable sources (articles, repos, notes, papers)
├── wiki/             ← Compiled articles (concepts, topics, references)
│   ├── concepts/     ← Core ideas and definitions
│   ├── topics/       ← Books, people, organizations, works
│   └── references/   ← Registries, digests, bibliographies
├── output/           ← Generated artifacts (playbooks, study guides, timelines)
├── inventory/        ← Durable tracking records (collection items, tasks)
├── datasets/         ← Manifests for large/external data
├── inbox/            ← Drop zone for unprocessed material
├── _index.md         ← Master index with stats
└── log.md            ← Append-only activity log

Raw is immutable. Once a source is ingested — a URL, a PDF, a book pointer, a pasted excerpt — it never changes. Articles in wiki/ are compiled from raw sources, not copied. Every article carries a compiled-from: and sources: frontmatter trail back to its evidence. This separation means you can always ask: where did this claim come from?

Articles are synthesized, not transcribed. A compiled article explains, contextualizes, and cross-references. It links to related concepts with dual-format links (Name for Obsidian + [Name](/knowledge/concepts/path) for everything else), rates its own confidence (high / medium / low), and declares its volatility (hot / warm / cold). The agent, not the human, maintains these cross-references — adding See Also back-links, updating indexes, and running structural lint checks after every operation.

The Pipeline

The core operations form a pipeline that any Lemon Wiki follows:

Research is the most distinctive operation. In standard mode, five parallel agents search simultaneously from different angles (academic, technical, applied, news, contrarian). In --deep mode, eight agents add historical context, adjacent fields, and data sources. In --retardmax mode, ten agents skip planning and ingest aggressively — "act first, think later."

Question-mode research (/wiki:research "What are the 24 critical concepts...?") decomposes the question into sub-questions, assigns one agent per sub-question, synthesizes the findings, and generates a playbook as the deliverable.

Key Design Principles

One topic, one wiki. Each research area gets its own isolated wiki with its own indexes, articles, and log. The Buckyverse .wiki/ is a single-topic local wiki; the hub architecture (~/wiki/topics/) supports many topic wikis side by side, each openable as its own Obsidian vault.

Navigation by index. Every directory has an _index.md. Agents read indexes first, never scan blindly. This is what makes the system fast — an agent entering the wiki reads the master _index.md, finds the relevant category index, and navigates to specific articles. No filesystem crawling.

Confidence scoring. Articles are rated high (multiple corroborating sources, verified), medium (single good source or unverified synthesis), or low (stub, thin evidence). The agent sets confidence when compiling; humans can override.

Dual-linking. Every cross-reference appears twice: Name ([Name](/knowledge/concepts/path)). The wikilink powers Obsidian's graph view and backlinks panel; the markdown link works in GitHub, any text editor, and the Astro reader. The wiki is not locked into any tool.

Structural guardian. After every operation, the agent checks wiki integrity — are indexes up to date? Do all links resolve? Are stats accurate? Trivial issues are fixed silently; structural problems are reported.

Activity log. log.md is append-only and grep-friendly. Every research session, compile pass, ingest, lint fix, and output generation gets a timestamped one-line entry. The log is the operational history of the wiki.

The Tooling

The Lemon Wiki system is the llm-wiki plugin, created by nvk and inspired by Andrej Karpathy's LLM wiki concept. It ships as:

• A Claude Code plugin (the primary runtime — claude plugin install wiki@llm-wiki)
• An OpenAI Codex plugin (marketplace installable)
• An OpenCode instruction file (URL-fetchable, auto-updating)
• A portable AGENTS.md for any other LLM agent

Zero external dependencies. The entire system runs on the agent's built-in file and shell tools. The wiki itself is plain markdown files in directories — no database, no server, no build step for the wiki layer. (The Astro reader that publishes wiki content to the web is a separate concern.)

This Wiki

The Buckyverse .wiki/ is a project-local Lemon Wiki (the --local variant, living inside the repo rather than in the hub). It was initialized on 2026-06-27 by migrating the legacy llm-wiki/ directory, and has since grown to ~400 compiled articles across concepts, topics, and references, plus raw sources, inventory records, datasets, and generated output artifacts. Its content covers R. Buckminster Fuller's work, legacy, and related design science — but the infrastructure that manages it is topic-agnostic. The same plugin runs wikis on nutrition, bitcoin, woodworking, or any other subject.

See Also

• Wiki Maintenance & Tooling Lessons — operational guidance compiled from building this wiki
• Collection Registry Overview — the inventory/collection subsystem
• Design Science — Fuller's methodology; the wiki is a design science tool in practice

Sources

• llm-wiki README — the plugin repository and documentation
• Andrej Karpathy's LLM wiki concept — the original idea
• This wiki's own _index.md, log.md, and operational history