Claude Code Prompt

English | 中文

A comprehensive, reverse-engineered documentation of all system prompts, tool prompts, and product architecture for Claude Code -- Anthropic's official AI coding assistant.

---

Overview

This project provides a complete analysis and documentation of Claude Code's prompt engineering system. Every prompt — from core system instructions to individual tool definitions — has been extracted, categorized, and documented with Mermaid architecture diagrams.

What's inside:

• Full system prompt breakdown (13 sections)
• All tool prompts (20+ tools)
• Agent & Coordinator multi-worker orchestration system
• Output style system
• Safety & security instructions
• Prompt assembly pipeline & caching mechanism

Table of Contents

Document	Description
00-architecture	Product architecture overview with 10+ Mermaid diagrams
01-system-prompt	Core system prompt -- 13 sub-sections covering identity, task guidance, tool usage, etc.
02-tool-prompts-core	Core file tools: Bash, Read, Write, Edit
03-tool-prompts-search	Search tools: Glob, Grep, WebSearch, WebFetch
04-tool-prompts-workflow	Workflow tools: PlanMode, Skill, Config, Cron, Worktree, Team, ToolSearch
05-agent-and-coordinator	Agent system (Fork/Subagent) & Coordinator multi-worker mode
06-output-styles	Output styles: Default, Explanatory, Learning
07-safety-and-security	Cyber risk & safety instructions
08-companion-and-auxiliary	Companion system, Hooks, Language, XML Tags, Memory
09-prompt-assembly	Prompt assembly pipeline, caching boundaries & section registration

Architecture at a Glance

graph TB
    subgraph Entry["User Entry Points"]
        CLI["CLI"]
        Desktop["Desktop App"]
        WebApp["Web App"]
        IDE["IDE Extensions"]
    end

    subgraph Core["Core Engine"]
        SP["System Prompt Assembly"]
        TP["Tool Pool (20+ tools)"]
        PM["Permission Manager"]
    end

    subgraph Agent["Agent System"]
        Fork["Fork Mode"]
        Sub["Subagent Mode"]
        Coord["Coordinator Mode"]
    end

    Entry --> Core
    Core --> Agent
    Core -->|"Claude API"| LLM["LLM (Opus/Sonnet/Haiku)"]

Loading

Key Insights

Prompt Assembly Pipeline

Claude Code's system prompt is assembled from 17 modular sections, split by a Dynamic Boundary marker into:

• Static sections (cacheable globally across organizations)
• Dynamic sections (session-specific, recomputed per turn)

This design enables efficient prompt caching while maintaining per-session customization.

Agent Architecture

Claude Code supports three agent modes:

Mode	Use Case
Fork	Inherits parent context, shares prompt cache. Best for research & implementation.
Subagent	Fresh context with specialized tools. Best for independent verification.
Coordinator	Orchestrates multiple Workers in parallel. Research → Synthesis → Implementation → Verification.

Tool Deferral System

Not all tools are loaded at startup. MCP tools and tools marked shouldDefer are lazily loaded via ToolSearch — only their names appear initially, and full schemas are fetched on demand.

Use as a Claude Code Skill

This project includes a ready-to-use Claude Code Skill (/claude-code-prompt) that provides prompt engineering best practices as a reference during AI/Agent development.

Installation

Copy the skill file to your Claude Code commands directory:

# Clone the repo
git clone https://github.com/mm7894215/claude-code-prompt.git

# Install the skill globally
cp -r claude-code-prompt/skills/claude-code-prompt ~/.claude/skills/

Usage

In any Claude Code session:

/claude-code-prompt                      # Full reference overview
/claude-code-prompt --topic=agent        # Agent/multi-worker patterns
/claude-code-prompt --topic=tools        # Tool prompt design patterns
/claude-code-prompt --topic=system       # System prompt architecture
/claude-code-prompt --topic=coordinator  # Multi-worker orchestration
/claude-code-prompt --topic=safety       # Safety instruction patterns

What It Provides

When you're building AI assistants, agent systems, or tool-using LLM applications, this skill gives you:

• System Prompt Architecture -- Modular section design, cache boundary patterns, key behavioral principles
• Tool Prompt Design -- Routing rules, sandbox patterns, read-before-write invariants, deferred loading
• Agent Orchestration -- Fork vs Subagent decisions, prompt writing rules, parallel execution patterns
• Coordinator Patterns -- Phase-based workflows, synthesis-first principle, continue vs spawn decisions
• Safety Design -- Layered safety architecture, dual-use policies, reversibility framework, prompt injection defense

Disclaimer

Important

This project is an unofficial, independently created documentation based on publicly available source code analysis. It is not affiliated with, endorsed by, or sponsored by Anthropic.

The content is provided for educational and research purposes only. Prompt content may change at any time as Claude Code evolves.

Contributing

Contributions are welcome! Please see CONTRIBUTING.md for guidelines.

If you find inaccuracies or have updates based on newer versions of Claude Code, feel free to open an issue or submit a pull request.

License

This project is licensed under the MIT License.

Star History

If you find this project useful, please consider giving it a star!