llm-wiki Skill: A Karpathy-Method Personal Knowledge Base for AI CLI Agents

The Problem llm-wiki Addresses

Most AI CLI sessions start from scratch. The agent reads the documents you point it to, reasons about them, and produces answers, but nothing accumulates. The next session repeats that process. llm-wiki addresses this by applying a methodology attributed to Andrej Karpathy: compile knowledge once into a structured wiki, then let the agent query and extend that wiki rather than re-deriving conclusions from raw sources.

The target user is an engineer or researcher who runs an AI agent as part of their regular workflow and who wants a growing, queryable knowledge base rather than a fresh-start assistant. The repository describes the core distinction plainly: knowledge is compiled once and continuously maintained, not re-derived on every query.

How the Skill and Workbench Modes Differ

The repository is a monorepo containing two entry points that read and write the same knowledge base format.

The Skill mode is described as mature and stable. It installs the wiki functionality directly into a supported AI CLI. Once installed, you interact with the knowledge base through your agent: asking it to initialize a wiki, ingest an article, or answer a question using the accumulated pages.

The Workbench mode, under `workbench/` in the repository, is a locally running application with a conversation interface and an interactive knowledge graph viewer. The README marks this as under development and aimed at developers. The development server starts with `npm run dev`. A desktop application is planned but not yet available. The two modes share a graph rendering engine located in `packages/graph-engine/`, meaning the offline HTML graph exported by the Skill and the live graph in the Workbench use the same rendering logic.

Installing the Skill and Running the First Ingest

Installation uses a shell script with a platform flag. The supported platforms are Claude Code, Codex, OpenClaw, and Hermes:

bash install.sh --platform claude

Optional content adapters for web pages, X/Twitter, WeChat public accounts, YouTube, and Zhihu are not installed by default. Enable them explicitly:

bash install.sh --platform claude --with-optional-adapters

After installation, ask the agent to initialize a knowledge base. Once initialized, point the agent at content to ingest:

> "Help me initialize a knowledge base" > "Help me digest this: <link>"

The repository notes that Node 22.19.0 or later is required, as stated in the `engines` field of package.json. The install script also ships a PowerShell variant, `install.ps1`, for Windows environments.

Each supported platform has its own entry point documentation: Claude Code uses `platforms/claude/CLAUDE.md`, Codex uses `platforms/codex/AGENTS.md`, OpenClaw uses `platforms/openclaw/README.md`, and Hermes uses `platforms/hermes/README.md`. Claude Code installations also receive an `/llm-wiki-upgrade` command for updating the Skill.

Knowledge Structure: Pages, Links, and Confidence Levels

The wiki stores knowledge as structured pages for entities and topics, linked with bidirectional `[[double-bracket]]` syntax. Ingesting a source creates entity pages, topic pages, and a source summary, all cross-referenced.

Each claim in the wiki carries a confidence annotation: EXTRACTED for information taken directly from source material, INFERRED for conclusions drawn from it, AMBIGUOUS for cases where the source is unclear, and UNVERIFIED for claims that need checking. This annotation system makes the epistemic status of each piece of knowledge visible at a glance.

The ingest process uses SHA256 deduplication with a caching layer described as self-healing. The README states that even a weaker model will not miss cache entries due to this design. A `purpose.md` file guides the agent when organizing and querying the wiki, giving it a stated research direction to follow.

The Offline Knowledge Graph

The Skill generates a self-contained HTML file that opens directly in a browser with no server required. The graph supports search, community filtering, focus mode, node visual layering, hover preview, lightweight summaries, zoom, pan, and a minimap. All of this runs offline.

The graph uses Sigma and Graphology as its primary rendering path, with a DOM/SVG fallback for environments where Sigma fails. The README describes the fallback as a safety net rather than a feature path: it activates only when Sigma itself encounters an error, not when the shared rendering result fails.

What llm-wiki Does Not Handle and Where It Falls Short

The README explicitly limits support to four AI CLIs: Claude Code, Codex, OpenClaw, and Hermes. Engineers using other agents cannot install the Skill without manual adaptation.

Xiaohongshu (Little Red Book) content can only be ingested by manual paste. The automatic extractors for web pages, X/Twitter, WeChat public accounts, YouTube, and Zhihu must be enabled at install time; they are not on by default, and the README notes that if automatic extraction fails, the fallback is a manual prompt.

The Workbench is marked as development-stage software aimed at developers rather than general users. Running it requires `npm run dev` and Node 22.19.0 or later. A desktop application does not yet exist.

The privacy boundary documentation notes that prompts, selected references, retrieval fragments, tool outputs, and generated content may be sent to whichever model provider the user configures. The knowledge files themselves stay on the local machine, but any query that touches the agent goes through the configured API.

How It Compares to Obsidian with an AI Plugin

Obsidian is a local-first Markdown note-taking application that stores notes as plain files and supports a plugin ecosystem. Several plugins connect Obsidian to AI APIs for summarization and linking. The difference in approach is where the AI fits. In Obsidian with an AI plugin, the user takes notes and optionally asks the AI to process them. The note-taking itself is manual.

In llm-wiki, the agent does the ingestion and page creation. The user points the agent at source material and the agent produces the structured wiki entries. The knowledge base is an artifact of agent work rather than user writing. The related search data shows interest in combining llm-wiki with Obsidian, but the README does not describe an Obsidian integration.

Maintenance and License

The last push to the repository was on 2026-07-27. The repository is not archived. The repository has no GitHub releases.

The package.json at the root lists MIT as the license. A CHANGELOG.md exists at the top level. The repository root also contains a TODOS.md, suggesting active development work is tracked in the repository rather than only through issues.

Editorial conclusion

llm-wiki suits engineers who use an AI CLI daily and want knowledge to persist across sessions as structured pages rather than re-summarized documents. The Skill mode is described as stable; the Workbench is explicitly in development and aimed at developers. Before deploying, check whether your AI CLI is among the four supported platforms and confirm you have Node 22.19.0 or later. The license file is MIT according to package.json; verify the LICENSE file in the repository before using in a commercial context.

Frequently asked questions

Official sources

1. Issues
2. README
3. sdyckjq-lab/llm-wiki-skill on GitHub