A static website explaining how Claude Code works internally, built for educational purposes. The goal is to help everyday users understand the tool better so they can use it more effectively.
Live at: claude-code-explain.helmcode.com
This site documents the internal architecture of Claude Code based on direct source code analysis: the system prompt structure, built-in tools, agent types, memory system, permission modes, context management, pricing, and more.
All findings come from reading the source code. The analysis/ directory contains the raw research notes.
- Framework: Astro 5 (static output)
- Styling: Tailwind CSS v4
- Languages: English and Spanish
- Deployment: Cloudflare Pages via GitHub Actions
npm install
npm run dev # local dev server
npm run build # production build to dist/| # | Section | What it covers |
|---|---|---|
| 01 | System Prompt | 6-layer hierarchy, 7 content sections, cache boundary |
| 02 | Tools | 45 built-in tools across 9 categories |
| 03 | Agents | 6 built-in agent types with models and tool sets |
| 04 | Memory | File-based persistent memory, 4 types, MEMORY.md index |
| 05 | Permissions | 6 modes, YOLO 2-stage AI classifier |
| 06 | Context & Compaction | 200K context window, auto-compact, micro-compaction |
| 07 | Costs | Pricing per model, fast mode premium |
| 08 | Retry Logic | Exponential backoff, error code handling |
| 09 | Proactive Mode | KAIROS autonomous mode, tick loop |
| 10 | Hidden Features | Feature flags, undocumented capabilities |
| 11 | Practical Tips | 10 tips derived from the source code |
This site is an independent educational project. It is not affiliated with or endorsed by Anthropic. All analysis is based on reading publicly available source code and may not reflect future versions of Claude Code.