English

What is this?

A Claude Code / Cowork plugin that turns an Obsidian vault into a self-maintaining course knowledge base. You drop course slides into raw/; the AI compiles them into an interlinked wiki in wiki/. You never write wiki pages yourself — you curate sources and ask questions. Based on Andrej Karpathy's LLM Wiki pattern.

Student-specific features

• Confidence decay — Concepts auto-downgrade after 30 days without review. Weak areas surface on the dashboard.
• Feynman review mode — AI quizzes you with "why" and "what if" questions to test real understanding.
• Exam prep generation — Auto-generates practice questions from your weakest concepts.
• Cross-course connections — Actively links concepts across different courses.
• Contradiction tracking — Conflicting descriptions get flagged with [!contradiction] callouts.
• Bilingual — All pages in Chinese-English.
• Domain rules — Security: attack-defense pairs. ML: bias-variance. NLP: pretrain/finetune distinction.

Plugin architecture

This is a real Claude Code plugin with modular skills — not a monolithic prompt. Each skill loads only when relevant, which keeps token usage low.

student-llm-wiki/
├── .claude-plugin/
│   ├── plugin.json          ← Plugin manifest
│   └── marketplace.json     ← Distribution config
├── skills/                  ← Modular skills (load on demand)
│   ├── wiki-core/           ←   Architecture + token rules (loads first)
│   ├── wiki-ingest/         ←   INGEST: digest a source
│   ├── wiki-lint/           ←   LINT: health check + confidence decay
│   ├── wiki-review/         ←   REVIEW: Feynman quiz
│   └── exam-prep/           ←   EXAM-PREP: generate questions
├── commands/                ← Slash commands
│   ├── ingest.md            ←   /ingest
│   ├── lint.md              ←   /lint
│   ├── review.md            ←   /review
│   └── exam-prep.md         ←   /exam-prep
├── raw/                     ← Layer 1: your course slides (read-only)
│   ├── COMP4337/  COMP6713/  COMP9417/  INFS5730/  misc/
│   └── .manifest.json       ←   hash dedup tracker
├── wiki/                    ← Layer 2: AI-maintained pages
│   ├── hot.md               ←   context cache (read first)
│   ├── index.md  overview.md  log.md
│   ├── courses/  concepts/  sources/  exam-prep/
├── SCHEMA.md                ← Human-readable design reference
└── Home.md                  ← Obsidian dashboard

Why split into skills?

A single SCHEMA file was loaded fully into context every operation (~3000 tokens). Splitting into skills means /ingest loads only ingest rules, /lint loads only lint rules. Combined with the hot.md cache (≤500 words read at session start) and manifest dedup (skip unchanged files), this cuts token usage dramatically.

Token efficiency

• Hot cache (wiki/hot.md) — AI reads ≤500 words at start instead of full rules.
• Manifest dedup (raw/.manifest.json) — files hashed before ingest; same hash = skip.
• Token budget rules (in wiki-core) — max 3–5 pages read per ingest, surgical edits, batch meta updates.

Quick start

Option A: Install as a plugin (Claude Code / Cowork)

# Add this repo as a marketplace
claude plugin marketplace add IssacW228/student-llm-wiki
# Install the plugin
claude plugin install student-llm-wiki@student-llm-wiki

Then open your vault folder, drop slides into raw/, and use the slash commands.

Option B: Use directly without installing

1. Download/clone this repo. Open the folder as an Obsidian vault. Install the Dataview plugin.
2. Open the folder in Claude Code, or point a Cowork project at it.
3. The AI reads skills/wiki-core/SKILL.md automatically when you start working.
4. Drop course PDFs into raw/{COURSE_CODE}/ and run a command.

Commands

You can also just say "ingest this file" or "quiz me on COMPXXXX" in natural language — the skills trigger automatically.

Adding your own courses

1. Create a folder under raw/ (e.g. raw/MATH1131/).
2. Create wiki/courses/MATH1131-overview.md.
3. If the course has special needs, add a domain rule in skills/wiki-core/SKILL.md.

Credits

• Andrej Karpathy's LLM Wiki — the original pattern
• claude-obsidian — hot cache, manifest dedup, skill decomposition
• Obsidian — the knowledge platform

License

MIT

中文

这是什么？

一个 Claude Code / Cowork 插件，把 Obsidian 知识库变成一个自维护的课程知识系统。你把课件丢进 raw/，AI 自动编译成互相链接的 wiki 到 wiki/。你不需要写任何 wiki 页面——只负责选材料和提问。基于 Karpathy 的 LLM Wiki 模式。

学生专属功能

• Confidence 衰减 — 概念超过30天未复习自动降级，薄弱环节自动浮现
• 费曼复习模式 — AI 用"为什么""如果...会怎样"提问，测试真正理解
• 自动出题 — 根据最薄弱的概念生成练习题
• 跨课程连接 — 主动链接不同课程的概念
• 矛盾标注 — 冲突的描述用 [!contradiction] 标记
• 中英双语 — 所有页面双语
• 域规则 — 安全:攻防配对 | ML:bias-variance | NLP:预训练/微调区分

插件架构

这是一个真正的 Claude Code 插件，采用模块化 skill——不是单体提示词。每个 skill 只在相关时加载，大幅降低 token 消耗。

student-llm-wiki/
├── .claude-plugin/
│   ├── plugin.json          ← 插件清单
│   └── marketplace.json     ← 分发配置
├── skills/                  ← 模块化技能（按需加载）
│   ├── wiki-core/           ←   架构 + token规则（最先加载）
│   ├── wiki-ingest/         ←   消化课件
│   ├── wiki-lint/           ←   健康检查 + confidence衰减
│   ├── wiki-review/         ←   费曼复习
│   └── exam-prep/           ←   出题
├── commands/                ← 斜杠命令
│   ├── ingest.md  lint.md  review.md  exam-prep.md
├── raw/                     ← 第一层：课件（只读）
│   ├── COMP4337/  COMP6713/  COMP9417/  INFS5730/  misc/
│   └── .manifest.json       ←   hash去重追踪
├── wiki/                    ← 第二层：AI维护的页面
│   ├── hot.md               ←   上下文缓存（首读）
│   ├── index.md  overview.md  log.md
│   ├── courses/  concepts/  sources/  exam-prep/
├── SCHEMA.md                ← 人类可读的设计参考
└── Home.md                  ← Obsidian 仪表盘

为什么拆成 skill？

单体 SCHEMA 每次操作都被完整读入上下文（约3000 tokens）。拆成 skill 后，/ingest 只加载 ingest 规则，/lint 只加载 lint 规则。配合 hot.md 缓存（每次开始只读≤500词）和 manifest 去重（跳过未变文件），token 消耗大幅下降。

Token 节省说明

• Hot cache（wiki/hot.md）：每次只读≤500词缓存，不读完整规则
• Manifest 去重（raw/.manifest.json）：文件 hash 校验，相同则跳过
• Token 预算规则（在 wiki-core）：每次最多读3-5页，局部编辑，批量更新meta

快速开始

方式一：作为插件安装（Claude Code / Cowork）

# 把本仓库添加为 marketplace
claude plugin marketplace add IssacW228/student-llm-wiki
# 安装插件
claude plugin install student-llm-wiki@student-llm-wiki

然后打开你的知识库文件夹，把课件放进 raw/，使用斜杠命令。

方式二：不安装直接用

1. 下载/克隆本仓库，用 Obsidian 打开作为知识库，安装 Dataview 插件
2. 用 Claude Code 打开该文件夹，或让 Cowork 项目指向它
3. 开始工作时 AI 会自动读取 skills/wiki-core/SKILL.md
4. 把课件 PDF 放进 raw/{课程代码}/，运行命令

命令

也可以直接用自然语言说"消化这个文件"或"考考我 COMPXXXX"——skill 会自动触发。

添加自己的课程

1. 在 raw/ 下创建文件夹（如 raw/MATH1131/）
2. 创建 wiki/courses/MATH1131-overview.md
3. 如有特殊需求，在 skills/wiki-core/SKILL.md 加一条域规则

致谢

• Karpathy LLM Wiki — 原始模式
• claude-obsidian — hot cache、manifest去重、skill拆分
• Obsidian — 知识平台

许可证

MIT