LLM Wiki

LLM Wiki — Second Brain for Claude Code + Obsidian

Inspired by Andrej Karpathy's LLM Wiki pattern (gist). This skill turns Claude Code (or any agent CLI) into a disciplined wiki maintainer that incrementally builds and maintains a persistent, interlinked Obsidian vault as you feed it sources. The knowledge compounds — cross-references, contradictions, and synthesis are already there when you query.

Core principle

Most LLM+docs workflows are RAG: retrieve fragments at query time, synthesize from scratch, forget. The wiki is compounding: sources are read once, integrated into a persistent markdown knowledge base, and kept current. You curate and ask; the LLM reads, files, cross-references, and maintains.

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

When to use

• Personal: track goals, health, psychology, journaling, self-improvement
• Research: deep dives over weeks on a topic — papers, articles, reports, evolving thesis
• Book companion: file chapters as you read; build a fan-wiki-style companion for characters, themes, plot threads
• Business/team: internal wiki fed by Slack, meeting notes, calls — LLM does maintenance nobody else wants to do
• Competitive analysis, due diligence, trip planning, course notes, hobby deep-dives

Do NOT use when: you need one-shot Q&A over a fixed document (use RAG), you don't plan to add sources over time, or you don't want Obsidian in the loop.

Architecture (three layers)

vault/
├── raw/                    # Layer 1 — IMMUTABLE source of truth
│   ├── <source files>      # Articles, papers, PDFs, images, data
│   └── assets/             # Downloaded images from clipped articles
├── wiki/                   # Layer 2 — LLM-owned knowledge base
│   ├── index.md            # Content catalog (LLM updates every ingest)
│   ├── log.md              # Append-only timeline (## [YYYY-MM-DD] <op> | <title>)
│   ├── entities/           # Person/Org/Place pages
│   ├── concepts/           # Ideas, theories, frameworks
│   ├── sources/            # One summary page per ingested source
│   ├── comparisons/        # Cross-source analysis pages
│   └── synthesis/          # High-level syntheses, theses, overviews
├── CLAUDE.md               # Schema + conventions (Claude Code)
└── AGENTS.md               # Same content, for Codex/Cursor/Antigravity

• Layer 1 (raw/) — you own. LLM only reads; never writes.
• Layer 2 (wiki/) — LLM owns. It creates, updates, and cross-references pages. You read it.
• Layer 3 (CLAUDE.md / AGENTS.md) — the schema. Conventions, workflows, frontmatter rules. Co-evolved by you and the LLM.

Three core operations

1. Ingest — LLM reads a source, discusses takeaways with you, writes a source summary, updates 10-15 relevant pages, updates index, appends to log. See references/ingest-workflow.md.
2. Query — LLM starts from the task and the nearest .agents/memory.yaml. Activated projects use memory context; opt-in current_contract repository sources precede durable_memory, while QMD failures degrade only to declared entry_pages. Vaults without a manifest keep an explicit legacy/non-pilot catalog route. Good answers get filed back into the wiki so explorations compound. See references/query-workflow.md.
3. Lint — Health check: contradictions, stale claims, orphan pages, missing cross-refs, concepts mentioned but lacking their own page, data gaps to fill with web search. See references/lint-workflow.md.

Quick start

# 1. Initialize a vault (in Obsidian's vault directory)
python scripts/init_vault.py --path ~/vaults/research --topic "LLM interpretability"

# 2. Drop a source into raw/, then ingest
/wiki-ingest ~/vaults/research/raw/anthropic-monosemanticity.pdf

# 3. Ask questions (answers can be re-filed into the wiki)
/wiki-query "how does monosemanticity compare to mechanistic interpretability?"

# 4. Periodic health check
/wiki-lint

# 5. See the timeline
/wiki-log --last 10

Slash commands (this plugin ships)

Sub-agents (this plugin ships)

Python tools (scripts/)

All tools are standard library only (no pip installs). Run with python scripts/<tool>.py --help.

wiki/index.md is a hash-attested managed section. New vaults receive exact markers automatically. Regeneration is idempotent, preserves content outside the managed section, and blocks if the markers or managed content drift. A legacy unmarked index must be reviewed and migrated explicitly; never adopt or overwrite it silently.

Cross-tool compatibility

The vault's schema lives in CLAUDE.md (Claude Code) or AGENTS.md (Codex/Cursor/Antigravity/OpenCode). The same content works in both. This plugin ships both templates. For per-tool setup instructions see references/cross-tool-setup.md.

CLAUDE.md       → Claude Code
AGENTS.md       → Codex CLI, Cursor, Antigravity, OpenCode, Gemini CLI
.cursorrules    → legacy Cursor (pre-AGENTS.md)

The scripts are pure Python stdlib → run identically everywhere. Only the loader file changes per tool.

Obsidian setup (recommended)

• Obsidian Web Clipper — browser extension; converts web articles to markdown and drops them in raw/
• Download images locally — Settings → Files and links → Attachment folder path = raw/assets/. Settings → Hotkeys → bind "Download attachments for current file" to Ctrl+Shift+D
• Graph view — see hubs/orphans; essential for spotting structural problems
• Marp plugin — Markdown-based slide decks directly from wiki pages
• Dataview plugin — dynamic tables/lists over page frontmatter (tags, dates, source counts)
• Git — the vault is a plain markdown repo; version it

Full setup walkthrough: references/obsidian-setup.md

Why this works (vs plain RAG)

At ~100 sources / hundreds of pages, index.md + filesystem search is enough. Past that, layer in a local search tool like qmd or use scripts/wiki_search.py.

Related skills (chains via context: fork)

This skill is marked context: fork so other skills can chain into it:

• para-memory-files — PARA-method memory; complementary as long-term personal memory that feeds sources into the wiki
• obsidian-vault (mattpocock) — lightweight Obsidian note helper; this skill is the maintained-wiki layer on top
• rag-design — when wiki outgrows ~500 pages, use rag-design to bolt on a retrieval layer
• mcp-design — expose the wiki as an MCP tool
• agent-communication — for multi-agent wiki maintenance (ingestor + linter + librarian)

Reference docs

• references/wiki-schema.md — full vault layout, page frontmatter, naming conventions
• references/page-formats.md — entity, concept, source, comparison, synthesis templates
• references/ingest-workflow.md — the detailed ingest flow the wiki-ingestor agent follows
• references/query-workflow.md — query patterns, citation format, re-filing answers
• references/lint-workflow.md — health-check heuristics
• references/obsidian-setup.md — Obsidian plugins, hotkeys, vault config
• references/cross-tool-setup.md — per-tool setup (Codex, Cursor, Antigravity, etc.)
• references/memex-principles.md — Bush's Memex, why the LLM changes the maintenance math

Templates (assets/)

• CLAUDE.md.template, AGENTS.md.template, .cursorrules.template — schema loaders per tool
• index.md.template, log.md.template — starter index and log
• page-templates/ — entity, concept, source-summary, comparison, synthesis
• example-vault/ — small worked example you can study or copy

Iron rule

The LLM never edits files in raw/. Ever. Sources are immutable. All LLM writes go to wiki/. If you need to correct a source, do it in raw/ yourself — then re-ingest.