一个用于构建卡帕西风格大型语言模型知识库的OpenClaw / Codex特工技能。

实验性技能——会随着时间不断迭代。 请在GitHub issues中反馈意见。

灵感来自Andrej Karpathy的LLM维基Gist以及社区基于此构建的工作。

这是什么

这种模式不是RAG（每次查询都重新检索原始文档），而是让LLM将原始源代码编译成一个持久的、交叉链接的Markdown维基。每传递一次compile、ingest、query、lint和audit，维基内容都会更丰富。知识会随着时间积累。

• 你拥有：原材料采购、提出好问题、指导、提交AI错误的反馈。
• LLM拥有：所有写作、交叉核对、归档、簿记以及根据你的反馈采取行动。

该技能在本仓库中附带两个配套工具：

• plugins/obsidian-audit/ — 一个Obsidian插件：在任意页面中选择文本，留下带有严重程度的评论，该注释作为锚定的标记文件写入audit/。
• web/ — 本地Node.js预览服务器：通过 mermaid、KaTeX 和维基链接渲染维基，允许你从浏览器中选择 + 文件反馈，并显示每页未完成的审核。

llm-wiki

An OpenClaw / Codex Agent Skill for building Karpathy-style LLM knowledge bases.

Experimental skill — will iterate over time. Please send your feedbacks in github issues.

Inspired by Andrej Karpathy’s llm-wiki Gist and the community’s work building on it.

What this is

Instead of RAG (re-retrieving raw docs on every query), this pattern has the LLM compile raw sources into a persistent, cross-linked Markdown wiki. Every compile, ingest, query, lint, and audit pass makes the wiki richer. Knowledge compounds over time.

• You own: sourcing raw material, asking good questions, steering direction, filing feedback on things the AI got wrong.
• LLM owns: all writing, cross-referencing, filing, bookkeeping, and acting on your feedback.

The skill comes with two companion tools in this repo:

• plugins/obsidian-audit/ — an Obsidian plugin: select text in any page, leave a comment with severity, the comment is written into audit/ as an anchored markdown file.
• web/ — a local Node.js preview server: renders the wiki with mermaid, KaTeX, and wikilinks, lets you select + file feedback from the browser, and shows open audits per page.

Both tools share a single TypeScript library (audit-shared/) so audit files written from Obsidian and the web viewer are byte-identical in shape.

Install

# Copy the skill into your agent's skills directory
cp -r llm-wiki/ ~/.claude/skills/llm-wiki/
# or for Codex
cp -r llm-wiki/ ~/.codex/skills/llm-wiki/

Then reference it in your agent config, or simply paste llm-wiki/SKILL.md into your agent context.

Quick start

# 1. Scaffold a new wiki
python3 llm-wiki/scripts/scaffold.py ~/my-wiki "My Research Topic"

# 2. Add a source
cp my-article.md ~/my-wiki/raw/articles/

# 3. Tell your agent: "ingest raw/articles/my-article.md"

# 4. Ask questions: "what does the wiki say about X?"

# 5. Run lint periodically
python3 llm-wiki/scripts/lint_wiki.py ~/my-wiki

# 6. File a comment from the web viewer or Obsidian plugin, then process it
python3 llm-wiki/scripts/audit_review.py ~/my-wiki --open
# then tell the agent: "audit: process the open comments"

Repo contents

llm-wiki-skill/
├── llm-wiki/                    ← The skill
│   ├── SKILL.md                 ← Main skill file (read by agent)
│   ├── references/
│   │   ├── schema-guide.md      ← CLAUDE.md schema template
│   │   ├── article-guide.md     ← Article writing (divide & conquer, mermaid, KaTeX)
│   │   ├── log-guide.md         ← log/ folder convention
│   │   ├── audit-guide.md       ← audit file format + processing workflow
│   │   └── tooling-tips.md      ← Obsidian, qmd, plugin + web
│   └── scripts/
│       ├── scaffold.py          ← Bootstrap new wiki directory
│       ├── lint_wiki.py         ← 7-pass health check (links, audit, log shape)