中文 ｜ English

中文

一个个人 LLM 知识库模板，灵感来自 Andrej Karpathy 的推文 —— 用 LLM 来构建和维护个人 wiki。本模板针对 Claude Code 优化，但任何能读写文件的 AI agent 都可以使用。

状态：空模板。你负责放资料、提问题；LLM 负责整理、交叉引用、回写和持续维护。

Last updated：2026-04-08

直接按你的情况选一条路：

最近更新

• 2026-04-08：新增 python scripts/wiki_index.py --report，生成 wiki/_attention.md 注意力 / 链接结构报告
• 2026-04-08：新增 GitHub Actions Wiki Lint，提交和 PR 会自动检查 wiki 健康状态
• 2026-04-08：新增 scripts/install_wiki.ps1，Windows 小白用户可一键安装
• 2026-04-08：优化 README 首页导流，第一次来的用户更容易知道该怎么开始
• 完整历史见 CHANGELOG.md

这是什么？

一个基于 markdown 的个人 wiki，用户负责策展，LLM 负责维护。把一份研究笔记丢进 wiki/raw/，LLM 会自动把它编译成结构化的 sources/ 页面、更新 concepts/ 和 entities/、建立交叉引用，并标记任何与你已有信念相冲突的地方。

wiki/
├── _schema.md          # 宪法。AI 在每次操作前先读它
├── _log.md             # 操作日志
├── raw/                # 不可变的原始材料（PDF、文章、文字稿）
├── sources/            # 编译后的结构化总结，每个来源一页
├── entities/           # 公司、人、书、项目——任何想追踪的对象
├── concepts/           # 主题、框架、把 entities 联系起来的想法
├── explorations/       # 研究问题的固化答案（回写）
├── decisions/          # 决策日志，含理由 + 取舍
├── rules.md            # 被 3 次以上验证过的规律
├── false-beliefs.md    # 被数据推翻的常识
└── comparisons/        # 并排对比

五层按 变化速率 组织：raw/ 永不变 → sources/ 很少变 → entities/ + concepts/ 周/月级 → explorations/ 随查询而变 → rules.md + false-beliefs.md 季度级。

和 Notion / Obsidian / RAG 的关键区别：传统知识库是写给"某天会回头读"的你的，所以维护成本最终会杀死它；这个模板是写给 每次查询都读一遍 的 LLM 的，维护成本由 LLM 承担——你只负责策展。没有向量数据库，markdown + frontmatter + LLM 上下文窗口已经够用（Karpathy 的洞察：个人 wiki 的体量根本不需要 RAG）。

🪟 Windows 小白：一键安装（推荐）

如果你 第一次接触这类项目 + 不会 Git / Markdown / 命令行，按这 4 步来。

开始前你只需要知道：

• 你不需要先懂 schema、protocol、frontmatter 这些词
• 你不需要会 Markdown；你的工作是"丢文件 + 提问题"
• 推荐环境：Windows + PowerShell + Claude Code

1. 下载这个项目

二选一：

• 会用 git：git clone https://github.com/Benboerba620/karpathy-claude-wiki.git
• 不会用 git：在 GitHub 页面点 Code → Download ZIP，解压

2. 打开 PowerShell，进入项目文件夹

cd 你的项目路径\karpathy-claude-wiki

3. 一键安装到你的项目目录

powershell -ExecutionPolicy Bypass -File .\scripts\install_wiki.ps1 -TargetDir "D:\my-project" -EntityName "AAPL"

这一条命令会自动帮你：复制 wiki/、复制 scripts/wiki_index.py、处理 CLAUDE.md、创建第一个可跟踪的 entity（这里是 AAPL，你可以换成任意股票代码 / 书名 / 人名）、生成索引并跑 lint。

如果你还没有自己的项目目录，把 -TargetDir 指到一个全新的空文件夹即可：

powershell -ExecutionPolicy Bypass -File .\scripts\install_wiki.ps1 -TargetDir "D:\my-wiki-project" -EntityName "AAPL"

4. 打开 Claude Code，对它说

读一下 wiki/_schema.md 和 CLAUDE.md，然后按 ingest 协议把我放进 wiki/raw/ 的文件摄入到 wiki 里。

小白常见问题

• 不懂 Git，能用吗？ 可以，下载 ZIP 解压也行。
• 不懂 Markdown，能用吗？ 可以，你的工作是"丢文件 + 提问题"，整理交给 AI。
• 必须先懂 schema / protocol 吗？ 不必。先装起来、跑起来，后面再慢慢看。
• 没装 Python，会失败吗？ 不一定。安装脚本会尽量继续执行；只是会跳过索引生成，之后装好 Python 再运行 python scripts/wiki_index.py 即可。
• 只想先试试，不想污染现有项目？ 把 -TargetDir 指到一个全新的空文件夹。

🍎 macOS / Linux 小白：一键安装

和 Windows 版完全等价的 bash 脚本，行为一样。

1. 下载

git clone https://github.com/Benboerba620/karpathy-claude-wiki.git
cd karpathy-claude-wiki

不会用 git：从 GitHub 页面 Code → Download ZIP，解压，cd 进去。

2. 一键安装

bash scripts/install_wiki.sh --target-dir ~/my-project --entity-name AAPL

参数说明：

• --target-dir：你的项目目录（不存在会自动创建）
• --entity-name：第一个想追踪的 entity 名称（股票代码 / 书名 / 人名都行）
• --force：如果目标目录已经有 wiki，覆盖
• --skip-index：跳过最后的索引 + lint

macOS 默认装了 Python 3，如果你之前删过，脚本会优雅跳过索引那一步，提醒你之后手动跑。

3. 打开 Claude Code，对它说

读一下 wiki/_schema.md 和 CLAUDE.md，然后按 ingest 协议把我放进 wiki/raw/ 的文件摄入到 wiki 里。

手动安装（5 分钟）

适合已经会用 git / 命令行 / Markdown 的人。

1. 克隆

git clone https://github.com/Benboerba620/karpathy-claude-wiki.git
cd karpathy-claude-wiki

2. 把 CLAUDE.md 复制到你的项目根目录

# 🍎 macOS / Linux
cp CLAUDE.md ../my-project/CLAUDE.md

# 🪟 Windows PowerShell
Copy-Item .\CLAUDE.md ..\my-project\CLAUDE.md

如果项目已经有CLAUDE.md，不要覆盖。把本仓库里的协议内容手动合并进去（或者用上面的一键安装脚本，它会自动 append）。

3. 把第一份原始材料丢进 raw/

# 🍎 macOS / Linux
cp ~/Downloads/some-research-paper.md wiki/raw/papers/

# 🪟 Windows PowerShell
Copy-Item "$HOME\Downloads\some-research-paper.md" .\wiki\raw\papers\

4. 打开 Claude Code，对它说

"读一下 wiki/_schema.md 和 CLAUDE.md，然后按 ingest 协议把 wiki/raw/papers/some-research-paper.md 摄入到 wiki 里。"

Claude 接下来会：

• 创建 wiki/sources/<日期>-<slug>.md，写一份结构化总结
• 识别其中提到的 entities 和 concepts，创建或更新对应页面
• 用 [[wikilinks]] 添加交叉引用
• 追加一行到 _log.md

完成第一次 ingest 之后，根据效果调整 schema，然后继续摄入更多内容。

让 AI agent 帮你装

打开 Claude Code（或 Cursor / Cline 等任何能读取 URL + 文件的 agent），把这条消息发给它：

帮我装这个：https://github.com/Benboerba620/karpathy-claude-wiki/blob/main/INSTALL-FOR-AI.md

Agent 会按 INSTALL-FOR-AI.md 里的 6 阶段协议走完整个流程：澄清你的领域、克隆模板、（如果需要）按你的领域定制、合并 CLAUDE.md、创建第一个 entity、生成索引、交付。

这个模板不包含什么

• 没有向量数据库。Markdown + frontmatter + LLM 上下文窗口已经够用。
• 没有 GUI。Obsidian、VSCode 或任意 markdown 编辑器都行。也可以一个都不用——直接和 LLM 对话。
• 没有领域内容。这里故意是空的。Yibo 的投资 wiki 用了 2 个月之后有 34 个 entities 和 111 个 sources；你的会长成完全不同的样子。

致谢

• Andrej Karpathy — 原始想法与框架
• Karpathy 的 gist — "想法文件"
• Stewart Brand《How Buildings Learn》 — shearing layers 原则
• Claude Code — 本模板针对其设计的 agent 框架

协议

MIT。Fork 它、改它、发布它。做出有意思的东西欢迎告诉作者。

关于作者

更多投资思考、研究方法与系统化协作的文章，欢迎关注微信公众号 奔波儿r：

English

A personal LLM knowledge-base template, inspired by Andrej Karpathy's tweet on building personal wikis with LLMs. Optimized for Claude Code but works with any AI agent that can read & write files.

Status: Empty template. You decide what goes in; the LLM handles the bookkeeping.

Last updated: 2026-04-08

Pick the path that matches you:

Recent Updates

• 2026-04-08: Added python scripts/wiki_index.py --report to generate the attention / link-graph report in wiki/_attention.md
• 2026-04-08: Added GitHub Actions Wiki Lint for automatic wiki health checks on pushes and PRs
• 2026-04-08: Added scripts/install_wiki.ps1 for beginner-friendly Windows setup
• 2026-04-08: Improved README routing so first-time visitors can pick the right install path faster
• Full history: CHANGELOG.md

What is this?

A markdown-based personal wiki where you curate, the LLM maintains. Drop a research note into wiki/raw/ and the LLM compiles it into structured sources/, updates concepts/ and entities/, builds cross-references, and flags contradictions with what you already believe.

wiki/
├── _schema.md          # The constitution. AI reads this before every action.
├── _log.md             # Operation log.
├── raw/                # Immutable original materials (PDFs, articles, transcripts).
├── sources/            # Compiled summaries, one page per source.
├── entities/           # Companies, people, books, projects — anything you track.
├── concepts/           # Themes, frameworks, ideas that connect entities.
├── explorations/       # Crystallized answers to research questions.
├── decisions/          # Decision log with reasoning + trade-offs.
├── rules.md            # Rules confirmed by 3+ instances.
├── false-beliefs.md    # Conventional wisdom that data has refuted.
└── comparisons/        # Side-by-side comparisons.

The five layers are organized by rate of change: raw/ never changes → sources/ rarely → entities/ + concepts/ weekly–monthly → explorations/ per-query → rules.md + false-beliefs.md quarterly.

Why this is different from Notion / Obsidian / RAG: traditional KBs are written for the future you who'll come back and read them — the maintenance cost eventually kills them. This template is written for the LLM that re-reads it on every query. Maintenance is the LLM's job; yours is curation. No vector database — markdown + frontmatter + the LLM's context window is enough. (Karpathy's insight: a personal wiki is small enough to not need RAG.)

🪟 Windows beginner: one-shot install (recommended)

If you're new to projects like this + don't know git / Markdown / CLI, follow these 4 steps.

Before you start, you only need to know:

• You don't need to understand schema, protocol, or frontmatter
• You don't need to know Markdown — your job is "drop files, ask questions"
• Recommended environment: Windows + PowerShell + Claude Code

1. Download the project

Either:

• With git: git clone https://github.com/Benboerba620/karpathy-claude-wiki.git
• Without git: on GitHub, click Code → Download ZIP, then unzip it

2. Open PowerShell and cd into the project folder

cd path\to\karpathy-claude-wiki

3. One-shot install into your project directory

powershell -ExecutionPolicy Bypass -File .\scripts\install_wiki.ps1 -TargetDir "D:\my-project" -EntityName "AAPL"

This single command copies wiki/, copies scripts/wiki_index.py, sets up CLAUDE.md, scaffolds your first trackable entity (here AAPL — replace with any ticker / book / person), and generates the index + lint report.

No project directory yet? Point -TargetDir at a new empty folder:

powershell -ExecutionPolicy Bypass -File .\scripts\install_wiki.ps1 -TargetDir "D:\my-wiki-project" -EntityName "AAPL"

4. Open Claude Code and say:

Read wiki/_schema.md and CLAUDE.md, then ingest the file I dropped in wiki/raw/ following the ingest protocol.

Beginner FAQ

• No git? Can I still use it? Yes — the ZIP download works.
• No Markdown experience? Fine. Your job is "drop files, ask questions"; the AI does the formatting.
• Do I need to understand schema / protocol first? No. Install it, run it, learn as you go.
• No Python? The installer will gracefully skip the index step. Install Python later, then run python scripts/wiki_index.py.
• I just want to try it without polluting an existing project. Point -TargetDir at a new empty folder.

🍎 macOS / Linux beginner: one-shot install

Behaviour-equivalent bash version of the PowerShell installer.

1. Download

git clone https://github.com/Benboerba620/karpathy-claude-wiki.git
cd karpathy-claude-wiki

No git? Use Code → Download ZIP on GitHub, unzip, cd in.

2. One-shot install

bash scripts/install_wiki.sh --target-dir ~/my-project --entity-name AAPL

Flags:

• --target-dir — your project directory (auto-created if missing)
• --entity-name — first entity to scaffold (a ticker / book / person)
• --force — overwrite an existing wiki at the target
• --skip-index — skip the final index + lint step

macOS ships Python 3 by default; if you've removed it the script will skip the index step gracefully and tell you to run it later.

3. Open Claude Code and say:

Read wiki/_schema.md and CLAUDE.md, then ingest the file I dropped in wiki/raw/ following the ingest protocol.

Manual install (5 minutes)

For people already comfortable with git / CLI / Markdown.

1. Clone

git clone https://github.com/Benboerba620/karpathy-claude-wiki.git
cd karpathy-claude-wiki

2. Copy CLAUDE.md to your project root

# 🍎 macOS / Linux
cp CLAUDE.md ../my-project/CLAUDE.md

# 🪟 Windows PowerShell
Copy-Item .\CLAUDE.md ..\my-project\CLAUDE.md

If your project already has a CLAUDE.md, don't overwrite it. Merge the wiki protocols in manually (or use one of the one-shot installers above — they auto-append).

3. Drop a first source into raw/

# 🍎 macOS / Linux
cp ~/Downloads/some-research-paper.md wiki/raw/papers/

# 🪟 Windows PowerShell
Copy-Item "$HOME\Downloads\some-research-paper.md" .\wiki\raw\papers\

4. Open Claude Code and say:

"Read wiki/_schema.md and CLAUDE.md, then ingest wiki/raw/papers/some-research-paper.md following the ingest protocol."

That's it. Claude will:

• Create wiki/sources/<date>-<slug>.md with a structured summary
• Detect the entities and concepts mentioned, create or update those pages
• Add cross-references using [[wikilinks]]
• Append to _log.md

After your first ingest, refine the schema based on what worked, then ingest more.

Let an AI agent install it for you

Open Claude Code (or Cursor / Cline / any agent that can fetch URLs and write files) and paste:

Install this for me: https://github.com/Benboerba620/karpathy-claude-wiki/blob/main/INSTALL-FOR-AI.md

The agent will follow the 6-phase protocol in INSTALL-FOR-AI.md: clarify your domain, clone the template, customize for your domain (if needed), merge CLAUDE.md, scaffold your first entity, generate the index, and hand off.

What's NOT in this template

• No vector database. Markdown + frontmatter + LLM context window is enough.
• No GUI. Use Obsidian, VSCode, or any markdown editor. Or none at all — talk to the LLM.
• No domain content. Intentionally empty. Yibo's investing wiki has 34 entities and 111 sources after 2 months; yours will look different.

Credit & inspiration

• Andrej Karpathy — original idea & framing
• Karpathy's gist — the "idea file"
• Stewart Brand's How Buildings Learn — shearing layers principle
• Claude Code — the agent harness this was designed for

License

MIT. Fork it, change it, ship it. Tell me what you build.

If you build something interesting on top of this, open an issue or PR.