一个自成体系的Claude代码技能，赋予Claude对自身架构的源代码级知识——涵盖所有内部子系统的50节课程，可通过三种方式搜索。

技能版本： 2.0.0 |截取自： Claude Code v2.1.88 |日期： 2026-03-31 |许可： MIT

---

目录

• 这是什么
• 为什么这个技能有用
• 工作原理
• 先决条件
• 安装
• 使用示例
• 采样输出
• 充分利用它
• RuFlo 与 RuVector 积分
• 故障排除
• 里面是什么
• 版本追踪
• 平台兼容性
• 许可证

---

这是什么

这是一个 Claude Code 技能（Claude Code 自动加载的本地知识包），包含对 Claude Code 内部架构的完整逆向工程。50 节详细课程涵盖了所有主要子系统——从启动序列到未发布的功能。当你输入 /claude-code-internals hooks 或 /claude-code-internals permissions 时，Claude 不会猜测或幻觉。它会阅读实际的架构文档，搜索索引参考资料，并给你源代码层面的答案，包含代码示例和类型定义。

Claude Code Internals

A self-contained Claude Code skill that gives Claude source-level knowledge of its own architecture — 50 lessons covering every internal subsystem, searchable three ways.

Skill Version: 2.0.0 | Captured from: Claude Code v2.1.88 | Date: 2026-03-31 | License: MIT

---

Table of Contents

• What This Is
• Why This Skill Is Useful
• How It Works
• Prerequisites
• Installation
• Usage Examples
• Sample Output
• Getting the Most Out of It
• RuFlo & RuVector Integration
• Troubleshooting
• What’s Inside
• Version Tracking
• Platform Compatibility
• License

---

What This Is

This is a Claude Code skill (a local knowledge package that Claude Code loads automatically) containing a complete reverse-engineering of Claude Code’s internal architecture. 50 detailed lessons cover every major subsystem — from the boot sequence to unreleased features. When you type /claude-code-internals hooks or /claude-code-internals permissions, Claude doesn’t guess or hallucinate. It reads actual architecture documentation, searches through indexed reference material, and gives you source-level answers with code examples and type definitions.

Without this skill, Claude knows how to use Claude Code but doesn’t know how Claude Code works internally. With it, Claude becomes an expert on its own implementation — the query engine’s retry logic, the 27 hook event types, the 7-phase permission pipeline, the compaction algorithm, the agent spawn lifecycle, all of it.

Why This Skill Is Useful

The Core Problem

Claude Code is a powerful tool, but Claude doesn’t understand its own internals. Ask it “what hook events are available?” and it’ll give you a partial, sometimes wrong answer. Ask it “why did compaction eat my context?” and it’ll speculate. Ask it “how do permission modes actually work?” and you’ll get a general answer that misses the 23 Bash security validators and the 7-phase decision pipeline.

This is because Claude’s training data doesn’t include Claude Code’s source code. It knows the public docs, but not the implementation details that matter when you’re configuring complex behavior.

What Changes With This Skill

• Claude stops guessing. Every answer comes from indexed architecture documentation, not training data. When you ask about hooks, Claude reads the actual hook system lesson that documents all 27 event types, exit code semantics, and 5 command types.