llm-wiki-skill

Language: English | 한국어 | 简体中文

Tutorial: English | 한국어 | 简体中文

Operations Guide: English | 한국어 | 简体中文

A standalone, runtime-agnostic skill for building and maintaining a persistent markdown knowledge base - based on Andrej Karpathy's LLM Wiki pattern.

Compatible with GitHub Copilot CLI, Claude Code, Codex CLI, Gemini CLI, and any agentskills.io-compatible runtime.

---

Why This Exists

llm-wiki-skill takes an idea proven inside NousResearch/hermes-agent and packages it as a reusable, cross-runtime skill package.

In Hermes Agent, llm-wiki is one bundled skill inside a larger agent system. This repository splits that idea into a standalone package with its own:

• SKILL.md for the core workflow
• templates/ for domain-specific schemas
• scripts/wiki-lint.py for zero-dependency health checks
• scripts/wiki-stats.py for quick coverage and activity summaries
• examples/ for a reference wiki
• extensions/ for optional add-ons like arXiv, GitHub, and Obsidian

The result is a cleaner way to adopt the pattern in tools beyond Hermes without bringing along Hermes-specific config and runtime assumptions.

---

The Problem It Solves

Every AI session starts from zero. You re-explain the domain, the agent re-reads the same sources, re-discovers the same relationships. Work doesn't compound.

llm-wiki flips this. The agent compiles knowledge once into interlinked markdown files. Next session: orient in 30 seconds (SCHEMA + index + recent log), then build on what's already there. Cross-references exist. Contradictions are flagged. Synthesis is done.

---

Start Here

• Read the hands-on tutorial: docs/TUTORIAL.md
• Keep the wiki healthy with the operations guide: docs/OPERATIONS.md
• Install the skill into your runtime
• Initialize a wiki for one real domain
• Ingest a few sources before judging the workflow

---

Installation

GitHub Copilot CLI

# Project-level: copy SKILL.md to .github/skills/
cp SKILL.md .github/skills/llm-wiki/SKILL.md

# User-level: available in all sessions
cp SKILL.md ~/.copilot/skills/llm-wiki/SKILL.md

# Project-level
New-Item -ItemType Directory -Force .github\skills\llm-wiki | Out-Null
Copy-Item SKILL.md .github\skills\llm-wiki\SKILL.md

# User-level
New-Item -ItemType Directory -Force $HOME\.copilot\skills\llm-wiki | Out-Null
Copy-Item SKILL.md $HOME\.copilot\skills\llm-wiki\SKILL.md

Claude Code

# User-level
cp SKILL.md ~/.claude/skills/llm-wiki/SKILL.md

Codex CLI

cp SKILL.md ~/.codex/skills/llm-wiki/SKILL.md

Gemini CLI

gemini skills install github:drvoss/llm-wiki-skill

---

Quick Start

# 1. Set your wiki path (add to your shell profile)
$env:LLM_WIKI_PATH = "$HOME/wiki"

# 2. If you're using this repository locally, bootstrap a wiki skeleton
python3 scripts/wiki-init.py $env:LLM_WIKI_PATH --template tech-stack --domain "AI coding tools"

# 3. Ingest your first source
# > Ingest this article into the wiki: [URL or paste text]

# 4. Query the wiki
# > What do we know about agentic coding patterns?

# 5. Run lint to check wiki health
python3 scripts/wiki-lint.py

For a complete scenario - including installing into a project, checking that Copilot sees the skill, watching the wiki grow, and learning better operating habits - use the tutorial: docs/TUTORIAL.md.

---

Three Operations

Operation	Trigger	What happens
Ingest	"Add this to the wiki", paste a URL	Saves raw source -> checks for duplicates -> creates/updates pages -> updates index + log
Query	"What does the wiki say about X?"	Reads index -> finds relevant pages -> synthesizes -> optionally files the answer
Lint	"Check the wiki" or python3 scripts/wiki-lint.py	Finds orphans, broken links, index gaps, frontmatter issues, stale content

---

Structure

llm-wiki-skill/
|- SKILL.md              <- main skill (load this into your AI tool)
|- scripts/
|  |- wiki-init.py       <- bootstrap a wiki skeleton from local templates
|  |- wiki-lint.py       <- standalone lint script (Python 3, no dependencies)
|  \- wiki-stats.py      <- lightweight wiki coverage and activity summary
|- templates/            <- ready-made SCHEMA.md for different domains
|  |- SCHEMA-tech-stack.md
|  |- SCHEMA-codebase.md
|  |- SCHEMA-product-intelligence.md
|  |- SCHEMA-research.md
|  \- pages/             <- starter templates for entity/concept/comparison/query pages
|- docs/                 <- tutorial and operations guides (EN / KO / ZH-CN)
|- examples/
|  \- sample-wiki/       <- working example wiki to explore
\- extensions/           <- optional add-ons (opt-in, separate dependencies)
   |- arxiv/SKILL.md     <- arXiv paper search (curl + python3 stdlib)
   |- github/SKILL.md    <- GitHub issues, PRs, releases, README ingest
   \- obsidian/SKILL.md  <- Obsidian vault + headless sync

---

Wiki Structure (created in your $LLM_WIKI_PATH)

~/wiki/
|- SCHEMA.md        <- domain rules and tag taxonomy (from templates/)
|- index.md         <- one-line summary for every page
|- log.md           <- append-only action log
|- raw/             <- immutable sources (articles, papers, transcripts)
|- entities/        <- people, orgs, tools, products
|- concepts/        <- ideas, techniques, patterns
|- comparisons/     <- side-by-side analyses
\- queries/         <- filed answers worth keeping

---

Extensions

Core SKILL.md has no external dependencies. Extensions are opt-in:

Extension	What it adds
extensions/arxiv	Search arXiv during ingest (no API key)
extensions/github	Ingest GitHub issues, PRs, releases, and README material
extensions/obsidian	Obsidian vault + headless sync

---

Lint Script

# Bootstrap a new wiki from a repository template
python3 scripts/wiki-init.py ~/wiki/ai-coding-tools --template tech-stack --domain "AI coding tools"

# Check wiki health
python3 scripts/wiki-lint.py                      # uses LLM_WIKI_PATH
python3 scripts/wiki-lint.py ~/my-project-wiki    # explicit path
python3 scripts/wiki-lint.py --strict ~/my-project-wiki
python3 scripts/wiki-lint.py --json ~/my-project-wiki
python3 scripts/wiki-lint.py --fix ~/my-project-wiki

# Get a lightweight wiki summary
python3 scripts/wiki-stats.py ~/my-project-wiki

# CI example for GitHub-hosted wikis
cat examples/github-actions/wiki-lint.yml

Checks: broken wikilinks, missing source files, orphan pages, index gaps, missing frontmatter, unknown tags, minimum outbound wiki links, contradiction targets, index header metadata, oversized pages, stale content, log rotation.

wiki-stats.py complements lint by giving a fast summary of page counts, link density, raw source counts, and recent activity without framing everything as a problem.

Exit code 0 = no errors. Exit code 1 = errors found. Exit code 2 = wiki not found.

---

Relationship to Hermes Agent and everything-copilot-cli

This repository started from the llm-wiki skill in NousResearch/hermes-agent, then expanded it into a standalone package for multiple runtimes.

It was also deliberately kept separate from everything-copilot-cli to avoid scope creep. llm-wiki-skill is cross-runtime and includes its own templates, lint tooling, tutorials, examples, and extensions - more than a single SKILL.md file.

---

License

MIT