wiki-knowledge-base-skill

Build and maintain a local obsidian knowledge base with based on Karpathy's LLM Wiki.

English

Install

Codex

Clone the whole repository, expose skills/ to Codex skill discovery, then restart Codex.

git clone https://github.com/Playitcooool/wiki-knowledge-base-skill.git ~/.codex/vendor_imports/wiki-knowledge-base-skill
mkdir -p ~/.agents/skills/wiki-knowledge-base
ln -s ~/.codex/vendor_imports/wiki-knowledge-base-skill/skills ~/.agents/skills/wiki-knowledge-base/skills

Claude Code

Local plugin directory:

claude --plugin-dir /path/to/wiki-knowledge-base-skill

Local or self-hosted marketplace:

/plugin marketplace add /path/to/wiki-knowledge-base-skill

After publishing:

/plugin marketplace add https://github.com/Playitcooool/wiki-knowledge-base-skill
/plugin install kb@knowledge-base

Cursor

For local testing, install the repository as a local Cursor plugin. The plugin is not published in Cursor Marketplace yet.

git clone https://github.com/Playitcooool/wiki-knowledge-base-skill.git ~/src/wiki-knowledge-base-skill
mkdir -p ~/.cursor/plugins/local
ln -s ~/src/wiki-knowledge-base-skill ~/.cursor/plugins/local/kb

Then reload Cursor and use:

/kb:ingest

Use

Use:

/kb:ingest

Typical intents:

/kb:ingest preview this folder first
/kb:ingest bootstrap this folder and build the knowledge base
/kb:ingest update pages/ from raw/
/kb:ingest check whether PDF conversion is ready

Behavior:

• Always check what would change first
• Auto-apply when the check is clean
• Ask before risky writes, especially delete operations
• Dependency question: run doctor logic
• Greenfield folder: initialize raw/, pages/, pages/index.md, and log.md when needed

The model can also infer this skill from user intent even without the explicit slash command, but /kb:ingest is the clearest path.You can check the graph in obsidian.

For direct local execution from this repository, run:

python3 skills/knowledge-base-maintainer/scripts/doctor.py
python3 skills/knowledge-base-maintainer/scripts/kb-ingest.py --root .
python3 skills/knowledge-base-maintainer/scripts/kb-ingest.py --root . --apply

Dependencies

• Out of the box: md, txt
• Requires pandoc: html, docx
• Basic PDF fallback:

pip install -r skills/knowledge-base-maintainer/requirements.txt

• Enhanced PDF and OCR:

pip install -r skills/knowledge-base-maintainer/requirements-optional.txt

Conversion path:

• HTML and DOCX: pandoc
• PDF: docling -> mineru -> pypdf

Page Categories

Generated pages are written under pages/ with one primary category:

• research: papers, reports, technical investigations, experiment writeups
• guides: explainers, comparisons, tool docs, workflow guides
• notes: reading notes, informal notes, meeting notes, or anything still to be confirmed

If classification is unclear, the system falls back to notes.

Architecture

Architecture inspiration: Andrej Karpathy, LLM Wiki

graph TD
    A["raw source files"] --> B["/kb:ingest"]
    B --> C["knowledge-base-maintainer"]
    C --> D["kb-ingest.py"]
    D --> E["sync_kb.py"]
    E --> F["convert_source.py"]
    F --> F1["pandoc for html and docx"]
    F --> F2["docling then mineru then pypdf"]
    E --> G["pages/category/slug.md"]
    E --> H["pages/index.md"]
    E --> I["log.md"]

Loading

中文

安装

Codex

建议直接按整仓安装。克隆仓库后，把 skills/ 暴露给 Codex 的 skill discovery，然后重启 Codex。

git clone https://github.com/Playitcooool/wiki-knowledge-base-skill.git ~/.codex/vendor_imports/wiki-knowledge-base-skill
mkdir -p ~/.agents/skills/wiki-knowledge-base
ln -s ~/.codex/vendor_imports/wiki-knowledge-base-skill/skills ~/.agents/skills/wiki-knowledge-base/skills

Claude Code

本地 plugin 目录方式：

claude --plugin-dir /path/to/wiki-knowledge-base-skill

本地或自托管 marketplace：

/plugin marketplace add /path/to/wiki-knowledge-base-skill

发布后也可以直接从 GitHub 安装：

/plugin marketplace add https://github.com/Playitcooool/wiki-knowledge-base-skill
/plugin install kb@knowledge-base

Cursor

如果是本地测试，把仓库作为本地 Cursor plugin 安装。当前还没有发布到 Cursor Marketplace。

git clone https://github.com/Playitcooool/wiki-knowledge-base-skill.git ~/src/wiki-knowledge-base-skill
mkdir -p ~/.cursor/plugins/local
ln -s ~/src/wiki-knowledge-base-skill ~/.cursor/plugins/local/kb

然后重载 Cursor，统一入口仍然是：

/kb:ingest

使用

统一入口是：

/kb:ingest

典型调用：

/kb:ingest 先预检当前文件夹
/kb:ingest 初始化当前目录并构建知识库
/kb:ingest 根据 raw/ 更新 pages/
/kb:ingest 检查 PDF 转换依赖是否就绪

行为规则：

• 始终先检查会发生什么变化
• 检查结果干净时自动 apply
• 遇到高风险写入，尤其是 delete 时先询问用户
• 用户在问依赖或转换能力时：先走 doctor 逻辑
• 绿地目录下：按需初始化 raw/、pages/、pages/index.md、log.md

即使用户不显式输入 /kb:ingest，模型也可以根据意图命中这个 skill，但显式命令最稳定。 完成之后可以在obsidian看到节点之间的依赖关系。

如果你是在这个仓库里直接运行脚本，请使用：

python3 skills/knowledge-base-maintainer/scripts/doctor.py
python3 skills/knowledge-base-maintainer/scripts/kb-ingest.py --root .
python3 skills/knowledge-base-maintainer/scripts/kb-ingest.py --root . --apply

依赖

• 开箱可用：md、txt
• 需要 pandoc：html、docx
• 基础 PDF fallback：

pip install -r skills/knowledge-base-maintainer/requirements.txt

• 增强 PDF 和 OCR：

pip install -r skills/knowledge-base-maintainer/requirements-optional.txt

转换链路：

• HTML / DOCX：pandoc
• PDF：docling -> mineru -> pypdf

页面分类

生成后的页面会写入 pages/，并且只归入一个主分类：

• research：论文、报告、技术调研、实验记录
• guides：解释文、对比文、工具文档、工作流指南
• notes：读书笔记、零散笔记、会议记录，或者暂时无法确认分类的内容

如果分类不明确，系统会默认回落到 notes。

架构

架构参考：Andrej Karpathy, LLM Wiki

graph TD
    A["raw 源文件"] --> B["/kb:ingest"]
    B --> C["knowledge-base-maintainer"]
    C --> D["kb-ingest.py"]
    D --> E["sync_kb.py"]
    E --> F["convert_source.py"]
    F --> F1["html 和 docx 走 pandoc"]
    F --> F2["pdf 走 docling 再 mineru 再 pypdf"]
    E --> G["pages/category/slug.md"]
    E --> H["pages/index.md"]
    E --> I["log.md"]

Loading

License

MIT. See LICENSE.