Turn Your Article Collection into an LLM‑Powered Wiki with a Single Skill

Problem

WeChat favorites often accumulate hundreds of technical articles that become impossible to locate later. Keyword search fails and trend analysis is impractical, turning the collection into a chaotic repository rather than a usable knowledge base.

LLM Wiki solution

Traditional Retrieval‑Augmented Generation (RAG) retrieves raw fragments for each query without organizing the source material. LLM Wiki “compiles knowledge” by extracting entities (companies, products, people), summarizing concepts, and building cross‑document references into a structured Wiki. Subsequent queries consult the organized Wiki, making retrieval fast and accurate.

Core workflow (Karpathy’s original design)

Ingest : Incrementally process new material, update Wiki pages, and maintain internal links without rebuilding the whole knowledge base.

Query : Answer questions across multiple pages with source citations; contradictions are flagged with a warning symbol.

Archive : Store valuable query results back into the knowledge base, enriching it over time.

Lint : Periodically audit for outdated content, orphan pages, and missing links.

Main components

Youdao Note : Cloud‑synced note storage accessible on mobile, desktop, web, and iPad.

youdaonote‑llm‑wiki skill : Packages the LLM Wiki methodology into a plug‑in that can be triggered with a single sentence in supported AI agents.

Setup: Install Youdao Note CLI

Install the CLI (macOS/Linux via the official script, Windows via manual download), bind the account API key, and verify configuration with: youdaonote list Successful output lists the notes, confirming the CLI is ready.

Finding and installing the skill

Download the Youdao Lobster client.

Open Lobster and go to the Skill Market.

Search for youdaonote-llm-wiki.

Click the skill card and install with one click.

In the chat box, type “Help me create a knowledge base”.

Creating a knowledge base (example)

In Lobster, send the prompt:

"I have many technical articles stored as Markdown in /Users/xxxxx/Desktop/llm-wiki-demo/notes/. Create a knowledge base named ‘Public Account Wiki’ to structure this content."

The agent suggests the English name wechat-article-wiki, which is confirmed. Ingestion starts automatically and processes 26 articles.

Resulting directory structure

The knowledge base contains five subfolders and three system notes: raw/ – original material. entities/ – extracted entities (companies, products, people). concepts/ – abstract concepts. comparisons/ – side‑by‑side comparison pages. queries/ – archived query results. schema.md – structure definition. index.md – index of all pages. log.md – change log.

Sample generated pages

Entity page entities/companies/anthropic.md aggregates mentions of Anthropic across multiple articles, including background, product updates, and event references, with backlinks to the source articles.

Concept page concepts/skills.md lists core actions of various developers and frameworks when releasing Skills, each citation indicating the original article.

Cross‑page query with citations

"In wechat-article-wiki, summarize the key trends for AI programming tools in 2026, citing sources."

The agent pulls relevant pages from index.md, concepts, entities, and comparisons, returns a multi‑sentence answer with source links, and flags any contradictory statements with a warning symbol.

Archiving the query

The answer is stored back into queries/ with a simple command, enabling the knowledge base to grow with each valuable interaction.

"Store the previous answer in wechat-article-wiki's queries/ folder."

Syncing new articles from AI coding tools

The skill also works in AI coding assistants such as Codex, Claude Code, and Cursor. Install the skill via npx skills add or a GUI manager, then issue a natural‑language command like “Sync this new article into wechat-article-wiki”. The note appears instantly in Youdao Note.

Differences from Karpathy’s original approach

Storage location : Karpathy – local Markdown files; youdaonote‑llm‑wiki – cloud‑synced notes.

Viewing method : Karpathy – local editor; youdaonote‑llm‑wiki – any device (mobile, desktop, web, iPad).

Setup cost : Karpathy – write scripts and configure prompts; youdaonote‑llm‑wiki – install skill and start using.

Maintenance : Karpathy – self‑maintain scripts and prompt versions; youdaonote‑llm‑wiki – skill updates automatically.

Target audience : Karpathy – tech‑savvy geeks; youdaonote‑llm‑wiki – anyone wanting knowledge retention.

Karpathy’s method is solid for developers who can build the system themselves; the skill packages the same methodology into a no‑code product.

Who can benefit

Authors of public accounts or blogs who need searchable, aggregated knowledge bases.

People with large collections of saved technical articles that have never been organized.

Developers tracking a specific technology direction (AI programming tools, Node.js ecosystem, front‑end engineering).

Learners who want to take structured notes while studying a new framework.