Live Example / 实际案例: llm-wiki-kb — A real knowledge base built with this skill, covering AI coding tools, Agent development, Prompt engineering and more.
Interactive Knowledge Graph — 76 nodes · 291 edges · 15 categories / 交互式知识图谱可视化
A Claude Code Skill based on Karpathy's LLM Wiki methodology. Instead of re-deriving knowledge from raw text every time (like RAG), LLM Wiki pre-compiles articles into structured, queryable wiki pages. Every new article makes the whole wiki richer. Every good answer gets saved back as a new knowledge node. Compound knowledge, not repeated inference.
npx skills add https://github.com/zyw-Wayne/wei-llm-wikiwiki init ~/my-wiki # Set wiki root (persisted)
wiki ingest https://example.com/article # Ingest an article
wiki query "What are the key takeaways?" # Query the knowledge baseOne command, auto-detects source type:
| Source | Example | Method |
|---|---|---|
| WeChat Articles | https://mp.weixin.qq.com/s/... |
wechat-article-down skill |
| GitHub Doc Repos | https://github.com/owner/repo |
GitHub MCP scan + batch read |
| Web Pages | https://blog.example.com/... |
chrome-devtools extraction + local images |
| Local Files | ~/notes/research.md |
Direct read (md/html/txt) |
# Mixed sources in one command
wiki ingest https://mp.weixin.qq.com/s/aaa ~/notes/b.md https://blog.example.com/c| Command | Description |
|---|---|
wiki init <path> |
Initialize wiki root, deploy knowledge graph HTML |
wiki ingest <source> |
Fetch article → store in raw/ → compile into wiki/ pages |
wiki query "question" |
Search wiki index → synthesize structured answer → auto-save insights |
wiki evolve [category] |
Audit knowledge coverage, track maturity with evolution vectors (🔴→🟡→🟢) |
wiki lint |
Health check: contradictions, stale content, orphan pages, missing cross-refs |
wiki graph |
Deploy interactive knowledge graph visualization (see preview below) |
wiki refresh |
Regenerate WIKI.md metadata from actual wiki state |
wiki log / wiki list |
Aggregated operation log / one-screen knowledge overview |
Not just storage — active gap detection.
wiki evolve agent # Audit "Agent Development" category
wiki evolve # Update all previously evaluated categories
wiki evolve --all # Full evaluation of all categories5-dimensional assessment: Breadth · Depth · Practicality · Timeliness · Cross-references
Evolution vectors track each knowledge direction: 🔴 Blank → 🟡 Basic coverage → 🟢 Mature, with specific search leads and suggestions.
Passive evolution: During wiki query, if knowledge gaps or factual contradictions are discovered, evolution profiles update automatically.
<wiki-root>/
├── raw/ # Original articles (read-only)
│ └── <title>/
│ ├── article.md
│ └── images/ # Downloaded images from web articles
├── wiki/ # LLM-compiled knowledge pages
│ ├── index.md # Content directory (query entry point)
│ ├── log.md # Operation log (append-only)
│ └── *.md # Concept / topic / query archive pages
├── evolve/ # Knowledge evolution tracking
│ ├── index.md # Dashboard: all category scores
│ └── <category>.md # Per-topic evolution profile
├── knowledge-graph.html # Interactive knowledge graph visualization
└── WIKI.md # Wiki schema
Collection mode: One parent directory manages multiple independent topic wikis. Auto-detects sub-wikis, generates a collection registry. Select target wiki interactively or specify path directly.
- Clear division: You curate sources and ask questions; LLM handles all maintenance — summaries, cross-references, archival, updates, contradiction detection
- Compound knowledge: Good query answers are saved back to wiki as reusable knowledge nodes
- Preserve contradictions: Marked with
> ⚠️ Contradiction:, never forcefully unified - Smart routing: Auto-detects wiki-root (persisted config → current dir → subdirectory scan → ask user)
基于 Karpathy LLM Wiki 方法论的 Claude Code 技能。不同于 RAG 每次从原文重新推导知识，LLM Wiki 预编译文章为结构化、可查询的 wiki 页面。每篇新文章让整个 wiki 更丰富，每个好答案存回 wiki 成为新的知识节点。知识复利积累，不是每次重新推导。
npx skills add https://github.com/zyw-Wayne/wei-llm-wikiwiki init ~/my-wiki # 指定知识库根目录（持久化）
wiki ingest https://mp.weixin.qq.com/s/xxxxx # 摄入一篇文章
wiki query "这篇文章的核心观点是什么？" # 查询知识库一个命令，自动识别来源类型并分派获取方式：
| 来源类型 | 示例 | 获取方式 |
|---|---|---|
| 微信公众号 | https://mp.weixin.qq.com/s/... |
调用 wechat-article-down 技能下载 |
| GitHub 文档仓库 | https://github.com/owner/repo |
GitHub MCP 扫描 + 批量读取，整仓库合并 |
| 普通网页 | https://blog.example.com/... |
chrome-devtools 抓取正文 + 图片本地化 |
| 本地文件 | ~/notes/research.md |
直接读取（支持 md/html/txt） |
# 混合来源批量摄入
wiki ingest https://mp.weixin.qq.com/s/aaa ~/notes/b.md https://blog.example.com/c| 命令 | 功能 |
|---|---|
wiki init <路径> |
初始化知识库根目录，部署知识图谱 HTML |
wiki ingest <来源> |
获取文章 → 存入 raw/ → 编译为 wiki/ 知识页面 |
wiki query "问题" |
检索 wiki 索引 → 综合结构化回答 → 自动存档有价值的洞察 |
wiki evolve [分类] |
覆盖度审计，用进化向量追踪知识成熟度（🔴→🟡→🟢） |
wiki lint |
健康检查：矛盾、过时内容、孤儿页面、缺失交叉引用 |
wiki graph |
部署交互式知识图谱可视化（见下方预览） |
wiki refresh |
根据实际状态重新生成 WIKI.md 元信息 |
wiki log / wiki list |
聚合操作日志 / 一屏知识概览 |
不只是"存文章"——主动识别知识缺口。
wiki evolve agent # 审计「Agent 开发」分类的覆盖度
wiki evolve # 更新所有已评估分类
wiki evolve --all # 全量评估所有分类五维评估体系：广度 · 深度 · 实用性 · 时效性 · 交叉引用
进化向量追踪每个知识方向的成熟度：🔴 空白 → 🟡 有基础覆盖 → 🟢 成熟，并给出具体的搜索线索和补充建议。
被动进化：wiki query 过程中若发现知识盲区或事实矛盾，自动更新进化档案并提示。
<wiki-root>/
├── raw/ # 原始文章（只读）
│ └── <文章标题>/
│ ├── article.md
│ └── images/ # 在线文章的本地图片
├── wiki/ # LLM 编译的知识页面
│ ├── index.md # 内容目录（查询入口）
│ ├── log.md # 操作日志（追加式）
│ └── *.md # 各概念/主题/查询存档页面
├── evolve/ # 知识进化追踪
│ ├── index.md # 总控面板：所有分类评分
│ └── <分类名>.md # 每个主题的进化档案
├── knowledge-graph.html # 交互式知识图谱可视化
└── WIKI.md # 知识库 schema
Collection 模式：一个父目录管理多个独立的主题知识库。自动检测子目录中的知识库，生成 collection 类型的注册表。操作时自动列出供选择，也可直接指定路径。
- 分工明确：用户筛选来源、提问、判断方向；LLM 负责摘要、交叉引用、归档、更新、检查矛盾
- 知识复利：好的查询答案存回 wiki，成为可复用的知识节点
- 矛盾保留：用
> ⚠️ 矛盾：标注，不强行统一，保留认知张力 - 被动进化：查询中自动发现盲区，无需手动审计
- 智能路由：自动检测 wiki-root（持久化配置 → 当前目录 → 子目录扫描 → 询问用户）
MIT