用 AI 把课件变成你自己的知识库 Turn your course slides into a personal wiki with AI
你把课件 PDF 丢进 raw/ 文件夹,AI 自动帮你整理成结构化的 wiki 笔记,存在 wiki/ 里。你用 Obsidian 浏览这些笔记,用 AI 工具来复习、出题、检查薄弱环节。你不需要手写任何笔记。
基于 Andrej Karpathy 的 LLM Wiki 模式。
1. Obsidian(用来浏览笔记)
- 下载地址:https://obsidian.md/download
- 免费,支持 Windows / macOS / Linux / iOS / Android
- 安装后还需要在社区插件里安装 Dataview(用于首页动态面板)
2. 选一个 AI 工具(任选其一即可)
| 工具 | 适合人群 | 获取方式 |
|---|---|---|
| Claude Code CLI | 熟悉命令行 | 安装说明 |
| Claude Code Web | 不想装任何东西 | claude.ai/code |
| Cursor | 喜欢 IDE 界面 | cursor.com |
| Trae | 国内用户 | trae.ai |
| Cowork | 团队协作 | 参考 Cowork 文档 |
| GitHub Copilot | 已有订阅 | VS Code 插件市场 |
3. 你的课件(PDF 或其他文本格式)
git clone https://github.com/IssacW228/student-llm-wiki.git
cd student-llm-wiki然后打开 Obsidian → 「打开本地仓库」→ 选择 student-llm-wiki 文件夹。
必须安装 Dataview 插件:Obsidian 设置 → 社区插件 → 浏览 → 搜索
Dataview→ 安装并启用。
Claude Code CLI(命令行)
- 确保已安装 Claude Code CLI 并登录
- 在项目目录运行:
claude
- 配置文件(
.claude/)会自动加载,无需额外设置 - 直接输入命令或自然语言即可开始
Claude Code Web(网页版,无需安装)
- 打开 claude.ai/code
- 将此仓库连接到 Claude Code
- 配置文件(
.claude/)自动识别,直接使用
Cursor
- 用 Cursor 打开
student-llm-wiki文件夹 - 规则文件已在
.cursor/rules/wiki.mdc,自动生效 - 在 Cursor 聊天框直接输入命令,无需额外配置
Trae
- 用 Trae 打开
student-llm-wiki文件夹 - Skills 已在
.trae/skills/,自动识别 - 在对话框直接输入命令
Cowork
- 将 Cowork 项目指向
student-llm-wiki文件夹 COWORK-INSTRUCTIONS.md会自动加载- 直接开始对话
GitHub Copilot / OpenAI Codex
- 用支持 Copilot 的编辑器打开项目文件夹
AGENTS.md会被自动读取- 直接在对话框输入命令
- 在
raw/下新建文件夹,用你的课程代号命名(如raw/MATH1001/) - 把课件 PDF 复制进去
- 在 AI 工具里输入:
ingest raw/MATH1001/L1.pdf - AI 会自动生成概念页、课程总览、来源摘要
- 回到 Obsidian,打开
Home.md查看结果
同一份文件不会被重复导入(通过文件 hash 去重)。
| 命令 | 作用 |
|---|---|
ingest raw/XXXX/L1.pdf |
导入课件,自动生成笔记 |
lint |
检查笔记健康度:断链、孤立页、过期概念 |
review XXXX |
费曼复习模式:AI 提问,你来回答 |
exam-prep XXXX |
根据薄弱概念自动出练习题 |
也可以用自然语言,比如「帮我复习这门课」或「把这个文件导入知识库」。
- 在
raw/下创建对应课程文件夹(如raw/PHYS1001/) - 把课件放进去
- 运行
ingest,课程总览页会自动创建 Home.md首页会动态显示新课程,无需手动修改
student-llm-wiki/
├── raw/ ← 放课件(只读,AI 不会修改这里)
│ └── XXXX/ ← 按课程分文件夹
├── wiki/ ← AI 生成的笔记(自动维护)
│ ├── concepts/ ← 概念页
│ ├── courses/ ← 课程总览
│ ├── sources/ ← 课件摘要
│ └── exam-prep/ ← 练习题
├── Home.md ← Obsidian 首页仪表盘
│
│ 以下为 AI 工具配置文件,无需修改:
├── .claude/ ← Claude Code 配置
├── .cursor/ ← Cursor 配置
├── .trae/ ← Trae 配置
├── AGENTS.md ← Copilot/Codex 配置
└── COWORK-INSTRUCTIONS.md ← Cowork 配置
- Andrej Karpathy's LLM Wiki — 原始模式
- claude-obsidian — hot cache、manifest 去重、skill 拆分
- Obsidian — 知识平台
MIT
You drop course PDF slides into the raw/ folder. The AI automatically organizes them into structured wiki notes in wiki/. You browse the notes with Obsidian, and use AI tools to review, generate practice questions, and surface weak spots. You never write notes yourself.
Based on Andrej Karpathy's LLM Wiki pattern.
1. Obsidian (for browsing your notes)
- Download: https://obsidian.md/download
- Free. Available on Windows / macOS / Linux / iOS / Android
- After installing, also install the Dataview community plugin (required for the Home dashboard)
2. An AI tool (pick just one)
| Tool | Best for | Get it |
|---|---|---|
| Claude Code CLI | Terminal users | Install guide |
| Claude Code Web | No install needed | claude.ai/code |
| Cursor | IDE-style interface | cursor.com |
| Trae | — | trae.ai |
| Cowork | Team use | See Cowork docs |
| GitHub Copilot | Existing subscribers | VS Code marketplace |
3. Your course slides (PDF or other text formats)
git clone https://github.com/IssacW228/student-llm-wiki.git
cd student-llm-wikiOpen Obsidian → "Open folder as vault" → select the student-llm-wiki folder.
Install the Dataview plugin: Obsidian Settings → Community plugins → Browse → search
Dataview→ Install and enable.
Claude Code CLI
- Make sure Claude Code CLI is installed and you're logged in
- Run in the project folder:
claude
- The config folder (
.claude/) is auto-detected — no extra setup needed - Type commands or plain English to get started
Claude Code Web (no install needed)
- Go to claude.ai/code
- Connect this repository to Claude Code
- The
.claude/config is auto-detected — start typing commands directly
Cursor
- Open the
student-llm-wikifolder in Cursor - Rules are pre-configured in
.cursor/rules/wiki.mdc— auto-applied - Type commands in the Cursor chat panel
Trae
- Open the
student-llm-wikifolder in Trae - Skills are pre-configured in
.trae/skills/— auto-detected - Type commands in the chat panel
Cowork
- Point your Cowork project at the
student-llm-wikifolder COWORK-INSTRUCTIONS.mdloads automatically- Start chatting
GitHub Copilot / OpenAI Codex
- Open the project in an editor with Copilot support
AGENTS.mdis read automatically- Type commands in the chat panel
- Create a folder under
raw/named after your course (e.g.raw/MATH1001/) - Copy your PDF into that folder
- In your AI tool, type:
ingest raw/MATH1001/L1.pdf - The AI generates concept pages, a course overview, and a source summary
- Switch to Obsidian and open
Home.mdto see your new notes
The same file won't be ingested twice — files are tracked by hash.
| Command | What it does |
|---|---|
ingest raw/XXXX/L1.pdf |
Import a slide deck and generate notes |
lint |
Check wiki health: broken links, orphan pages, stale concepts |
review XXXX |
Feynman review mode: AI quizzes you |
exam-prep XXXX |
Auto-generate practice questions from weak concepts |
You can also use plain English: "quiz me on this course" or "import this file into the wiki".
- Create a folder under
raw/for the course (e.g.raw/PHYS1001/) - Put your slides in it
- Run
ingest— a course overview page is created automatically - The
Home.mddashboard updates dynamically — no manual edits needed
student-llm-wiki/
├── raw/ ← Your slides (read-only — AI never modifies this)
│ └── XXXX/ ← One folder per course
├── wiki/ ← AI-generated notes (auto-maintained)
│ ├── concepts/ ← Concept pages
│ ├── courses/ ← Course overviews
│ ├── sources/ ← Slide summaries
│ └── exam-prep/ ← Practice questions
├── Home.md ← Obsidian dashboard
│
│ AI tool config — no need to touch these:
├── .claude/ ← Claude Code config
├── .cursor/ ← Cursor config
├── .trae/ ← Trae config
├── AGENTS.md ← Copilot/Codex config
└── COWORK-INSTRUCTIONS.md ← Cowork config
- Andrej Karpathy's LLM Wiki — the original pattern
- claude-obsidian — hot cache, manifest dedup, skill decomposition
- Obsidian — the knowledge platform
MIT