llm-wiki-bootstrap

Compare original and translation side by side

🇺🇸

Original

English

🇨🇳

Translation

Chinese

Karpathy LLM Wiki Bootstrap

Scaffold a complete LLM Wiki framework from zero: directory structure, schema (the AI instruction file), index, log, and operational workflows. The output is a ready-to-use wiki that the user and their LLM agent can immediately start populating.

Guiding principle: The wiki is a persistent, compounding artifact. The LLM writes and maintains all content. The human curates sources, directs analysis, and asks questions. This skill builds the scaffolding; the user and their LLM fill it with knowledge over time.

从零搭建完整的LLM Wiki框架：目录结构、Schema（AI指令文件）、索引、日志和操作工作流。输出结果是一个可直接使用的wiki，用户及其LLM Agent可立即开始填充内容。

指导原则：wiki是一个持久化、可不断积累的成果。所有内容由LLM编写和维护，人类负责整理来源、指导分析并提出问题。本技能仅搭建框架，用户及其LLM会随着时间推移为其填充知识。

Workflow

工作流

Six sequential phases. Each phase completes before the next begins.

六个连续阶段，需完成前一阶段后再进入下一阶段。

Phase 1: Gather Requirements

阶段1：收集需求

BLOCKING — must complete before any file creation.

Use

AskUserQuestion

with ALL questions in ONE call:

Q1 — header: "Domain", question: "What is this wiki about?"

Options: "Research topic", "Book / media", "Personal (goals, health, learning)", "Business / team", "Other (describe below)"

Q2 — header: "Wiki name", question: "Short name for the wiki root directory?"

Free text. Suggest:

{domain}-wiki

(e.g.

ml-wiki

lotr-wiki

health-wiki

)

Q3 — header: "Agent", question: "Which LLM agent will maintain this wiki?"

Options:
- "Claude Code" — generates
```
CLAUDE.md
```
- "OpenAI Codex" — generates
```
AGENTS.md
```
- "Copilot (VS Code)" — generates
```
.github/copilot-instructions.md
```
- "Other / generic" — generates
```
SCHEMA.md
```

Q4 — header: "Editor", question: "Primary editor for browsing the wiki?"

Options: "Obsidian (recommended)", "VS Code", "Other / plain files"

Q5 — header: "Source types", question: "What kind of sources will you add?"

Multi-select. Options: "Web articles", "PDFs / papers", "Books (chapter by chapter)", "Meeting notes / transcripts", "Personal notes / journals", "Images / diagrams", "Data files (CSV, JSON)", "Other"

Q6 — header: "Output location", question: "Where to create the wiki?"

Options:
- "Current directory" — create
```
{wiki-name}/
```
  here
- "Custom path" — user types absolute path

Store all answers as session variables for subsequent phases.

阻塞性任务——必须完成此阶段才能创建任何文件。

通过一次调用

AskUserQuestion

提出所有问题：

Q1 — 标题：“领域”，问题：“此wiki的主题是什么？”

选项：“研究课题”、“书籍/媒体”、“个人（目标、健康、学习）”、“商务/团队”、“其他（请在下方描述）”

Q2 — 标题：“Wiki名称”，问题：“wiki根目录的简短名称？”

自由文本输入。建议格式：
```
{domain}-wiki
```
（例如：
```
ml-wiki
```
、
```
lotr-wiki
```
、
```
health-wiki
```
）

Q3 — 标题：“Agent”，问题：“将维护此wiki的LLM Agent是哪一个？”

选项：
- "Claude Code" — 生成
```
CLAUDE.md
```
- "OpenAI Codex" — 生成
```
AGENTS.md
```
- "Copilot (VS Code)" — 生成
```
.github/copilot-instructions.md
```
- "其他/通用" — 生成
```
SCHEMA.md
```

Q4 — 标题：“编辑器”，问题：“浏览wiki的主要编辑器？”

选项：“Obsidian（推荐）”、“VS Code”、“其他/纯文本文件”

Q5 — 标题：“来源类型”，问题：“你将添加哪些类型的来源？”

多选。选项：“网页文章”、“PDF/论文”、“书籍（按章节）”、“会议记录/转录稿”、“个人笔记/日志”、“图片/图表”、“数据文件（CSV、JSON）”、“其他”

Q6 — 标题：“输出位置”，问题：“在哪里创建wiki？”

选项：
- "当前目录" — 在此处创建
```
{wiki-name}/
```
- "自定义路径" — 用户输入绝对路径

将所有答案存储为会话变量，供后续阶段使用。

Phase 2: Create Directory Structure

阶段2：创建目录结构

Based on Phase 1 answers, create the directory tree:

{wiki-root}/
├── raw/                    # Immutable source documents
│   └── assets/             # Images, attachments (if image sources selected)
├── wiki/                   # LLM-maintained markdown pages
│   ├── index.md            # Content catalog
│   ├── log.md              # Chronological operation log
│   └── overview.md         # High-level synthesis (starts empty)
├── {schema-file}           # AI instruction file (name from Q3)
└── .gitignore              # Ignore OS files, keep everything else

Conditional directories:

Condition	Add
Q5 includes images/diagrams	`raw/assets/`
Q4 = Obsidian	`.obsidian/` is NOT created (Obsidian auto-creates it)
Any source type selected	`raw/` with a `.gitkeep`

根据阶段1的答案，创建目录树：

{wiki-root}/
├── raw/                    # 不可变的源文档
│   └── assets/             # 图片、附件（若选择了图片来源）
├── wiki/                   # LLM维护的Markdown页面
│   ├── index.md            # 内容目录
│   ├── log.md              # 按时间顺序的操作日志
│   └── overview.md         # 高层级综述（初始为空）
├── {schema-file}           # AI指令文件（名称来自Q3）
└── .gitignore              # 忽略系统文件，保留其他所有内容

条件目录：

条件	添加内容
Q5包含图片/图表	`raw/assets/`
Q4 = Obsidian	不创建 `.obsidian/` （Obsidian会自动创建）
选择了任何来源类型	带有 `.gitkeep` 文件的 `raw/` 目录

Phase 3: Generate Schema File

阶段3：生成Schema文件

The schema is the most critical output. It instructs the LLM agent how to operate on the wiki. Generate it using the template at references/templates/schema.md.

Customization rules:

Variable	Source
`{WIKI_NAME}`	Q2 answer
`{DOMAIN_DESCRIPTION}`	Q1 answer (expanded to 1-2 sentences)
`{SOURCE_TYPES}`	Q5 answers, comma-separated
`{SCHEMA_FILENAME}`	Determined by Q3
`{EDITOR}`	Q4 answer
`{DATE}`	Current date in YYYY-MM-DD

After generating, adapt section details:

If domain is "Book / media" → add character, timeline, and plot-thread page types
If domain is "Research" → add paper-summary and claim-tracking page types
If domain is "Personal" → add journal-entry and goal-tracking page types
If domain is "Business / team" → add decision-log and meeting-summary page types

Schema是最重要的输出，它指导LLM Agent如何操作wiki。使用references/templates/schema.md中的模板生成。

自定义规则：

变量	来源
`{WIKI_NAME}`	Q2的答案
`{DOMAIN_DESCRIPTION}`	Q1的答案（扩展为1-2句话）
`{SOURCE_TYPES}`	Q5的答案，用逗号分隔
`{SCHEMA_FILENAME}`	由Q3决定
`{EDITOR}`	Q4的答案
`{DATE}`	当前日期（格式：YYYY-MM-DD）

生成后，调整章节细节：

如果领域是“书籍/媒体”→添加角色、时间线和情节线页面类型
如果领域是“研究”→添加论文摘要和主张追踪页面类型
如果领域是“个人”→添加日志条目和目标追踪页面类型
如果领域是“商务/团队”→添加决策日志和会议摘要页面类型

Phase 4: Generate Initial Wiki Files

阶段4：生成初始Wiki文件

Create three seed files using templates in

references/templates/

4.1 index.md — from references/templates/index.md

Start with the universal sections (Sources, Entities, Concepts, Comparisons, Synthesis)
Add domain-specific sections matching the page types injected in Phase 3:
- Research → Papers, Claims, Methods, Datasets
- Book / media → Characters, Timelines, Themes, Locations
- Personal → Journal, Goals, Habits, Lessons
- Business → Decision Logs, Meetings, Projects, Stakeholders
Leave content sections empty with placeholder comments

4.2 log.md — from references/templates/log.md

Write the first entry: wiki creation event with current date

4.3 overview.md — from references/templates/overview.md

Minimal stub that explains the wiki's purpose and domain

使用

references/templates/

中的模板创建三个初始文件：

4.1 index.md — 来自references/templates/index.md

从通用章节开始（来源、实体、概念、对比、综述）
添加与阶段3中注入的页面类型匹配的领域特定章节：
- 研究→论文、主张、方法、数据集
- 书籍/媒体→角色、时间线、主题、地点
- 个人→日志、目标、习惯、经验教训
- 商务→决策日志、会议、项目、利益相关者
内容部分留空，添加占位符注释

4.2 log.md — 来自references/templates/log.md

写入第一条记录：包含当前日期的wiki创建事件

4.3 overview.md — 来自references/templates/overview.md

极简 stub，说明wiki的用途和领域

Phase 5: Editor Configuration

阶段5：编辑器配置

If Q4 = Obsidian:

Do NOT create

.obsidian/

or modify Obsidian settings. Instead, append a

## Obsidian Setup

section to the schema file with recommendations:

Set "Attachment folder path" to
```
raw/assets/
```
in Settings → Files and links
Install recommended plugins: Dataview (frontmatter queries), Marp (slide decks if needed)
Use graph view to inspect wiki structure
Bind "Download attachments" hotkey if using Web Clipper

If Q4 = VS Code:

Create

.vscode/settings.json

with markdown-friendly defaults:

json

{
  "files.exclude": { "**/.DS_Store": true },
  "editor.wordWrap": "on",
  "markdown.preview.breaks": true
}

Do NOT append an Obsidian section to the schema file.

If Q4 = Other / plain files:

Skip editor configuration entirely. Do NOT append any editor-specific section to the schema file. Do NOT create

.vscode/

如果Q4 = Obsidian：

不要创建

.obsidian/

或修改Obsidian设置。而是在Schema文件末尾添加

## Obsidian 设置

章节，提供建议：

在设置→文件与链接中，将“附件文件夹路径”设置为
```
raw/assets/
```
安装推荐插件：Dataview（前置元数据查询）、Marp（如需幻灯片）
使用图谱视图检查wiki结构
如果使用Web Clipper，绑定“下载附件”快捷键

如果Q4 = VS Code：

创建

.vscode/settings.json

，配置Markdown友好的默认设置：

json

{
  "files.exclude": { "**/.DS_Store": true },
  "editor.wordWrap": "on",
  "markdown.preview.breaks": true
}

不要在Schema文件中添加Obsidian章节。

如果Q4 = 其他/纯文本文件：

完全跳过编辑器配置。不要在Schema文件中添加任何编辑器特定章节，也不要创建

.vscode/

目录。

Phase 6: Summary and Next Steps

阶段6：总结与后续步骤

After all files are created, output a brief summary:

Wiki scaffolded at {wiki-root}/

Structure:
  raw/          → Drop source documents here
  wiki/         → LLM-maintained pages (index, log, overview)
  {schema-file} → AI instructions for wiki operations

Next steps:
  1. Open {wiki-root}/ in {editor}
  2. Add your first source to raw/
  3. Tell your LLM agent: "Read {schema-file}, then ingest raw/{filename}"

所有文件创建完成后，输出简短总结：

Wiki已搭建在 {wiki-root}/

结构说明：
  raw/          → 在此处放入源文档
  wiki/         → LLM维护的页面（索引、日志、综述）
  {schema-file} → wiki操作的AI指令

后续步骤：
  1. 在{editor}中打开{wiki-root}/
  2. 向raw/中添加你的第一个源文档
  3. 告诉你的LLM Agent：“阅读{schema-file}，然后处理raw/{filename}”

Post-Bootstrap Operations

搭建后操作

The schema file generated in Phase 3 defines three core operations. These are documented in detail in:

references/workflows/ingest.md — source ingestion protocol
references/workflows/query.md — query and answer-filing protocol
references/workflows/lint.md — wiki health-check protocol

The schema file embeds condensed versions of these workflows. The reference files here contain the full rationale and edge cases for skill maintainers.

阶段3生成的Schema文件定义了三个核心操作，详细文档位于：

references/workflows/ingest.md — 源文档处理协议
references/workflows/query.md — 查询与答案归档协议
references/workflows/lint.md — wiki健康检查协议

Schema文件中嵌入了这些工作流的精简版本，此处的参考文件包含了技能维护者所需的完整原理和边缘情况说明。

Design Principles

设计原则

These inform all generated content:

LLM-native — all instructions are written as executable protocols for language models, not prose for humans
Source of truth separation — raw sources are immutable; the wiki is derived and regenerable
Incremental compilation — each source is integrated once; knowledge compounds, not re-derived
Convention over configuration — sensible defaults, minimal required decisions
Editor-agnostic core — markdown files work everywhere; editor-specific features are optional layers
Git-friendly — plain text, no binary blobs in wiki/, version history for free

这些原则指导所有生成内容：

LLM原生 — 所有指令均为语言模型可执行的协议，而非面向人类的散文
真相源分离 — 原始源文档不可变；wiki是派生且可重新生成的
增量编译 — 每个源文档仅集成一次；知识不断积累，无需重新推导
约定优于配置 — 合理的默认设置，最小化必要决策
编辑器无关核心 — Markdown文件可在任何地方使用；编辑器特定功能为可选层
Git友好 — 纯文本，wiki/目录中无二进制大对象，免费获得版本历史",