system-prompts-and-models-of-ai-tools/Claude Code

Claude Code

Type: CLI Tool (Terminal-based AI Assistant)
Availability: Free with Claude API access
Provider: Anthropic
Model: Claude 3.5 Sonnet, Claude 4 (latest)

---

📋 Overview

Claude Code is Anthropic's official terminal-based AI coding assistant. Features:

• Autonomous terminal agent
• Full file system access
• Bash command execution
• Multi-file editing
• Web search integration
• TODO/progress tracking
• AGENTS.md support
• Git workflow integration

Philosophy: Concise, action-oriented, autonomous

---

📂 Files in This Directory

System Prompts:

• claude-code-system-prompt.txt - Complete system instructions

Tools:

• claude-code-tools.json - Comprehensive tool definitions (20+ tools)

---

🔍 Source

• Provider: Anthropic
• Official Documentation: Part of Claude API
• Date Captured: October 2024 (latest update)
• Attribution: Anthropic PBC

---

🎯 Key Features

1. Extreme Conciseness

One of the most concise prompts analyzed:

"Be concise. Answer in 1-3 sentences. No preamble or postamble. After editing, just stop."

2. TODO System

Built-in progress tracking:

Tools: todo_write, todo_read

Workflow:
1. Plan: Create TODO list
2. Execute: Mark items in-progress
3. Complete: Mark done immediately
4. Report: User sees progress

3. AGENTS.md Pattern

Per-project context file:

AGENTS.md contains:
- Commands (npm test, npm run dev)
- Style preferences
- Project notes
- Custom instructions

4. Git Workflow

Structured commit process:

1. Check status + diff in parallel
2. Analyze changes + check for secrets
3. Stage files
4. Commit with formatted message
5. Include attribution footer

5. Defensive Security Only

Strong security stance:

"IMPORTANT: Assist with defensive security tasks only. Refuse to create, modify, or improve code that may be used maliciously."

---

🛠️ Tool Architecture

Core Tools (20+):

File Operations:

• read_file, write_to_file, search_replace
• list_dir, glob (pattern matching)

Search:

• grep_search, semantic_search
• read_web_page (web search integration)

Execution:

• run_bash_command (with background mode)

Progress:

• todo_write, todo_read

Analysis:

• get_linting_diagnostics

Git:

• Via bash with structured workflow

Tool Design Principles:

1. Minimal tool count (20 vs. 30+ in some tools)
2. Clear separation of concerns
3. Composable (tools combine well)
4. Parallel-friendly (independent operations)

---

📊 Unique Patterns

1. No Explanatory Comments

Explicit instruction:

"IMPORTANT: DO NOT ADD ANY COMMENTS unless asked. Only add comments when: 1. User explicitly requests them, 2. Code is complex and requires context"

Rationale: AI can explain in chat, comments clutter code.

2. Verify Before Committing

Required checks:

• Analyze all staged changes
• Draft commit message
• Check for sensitive information
• Include co-author attribution

3. Parallel Execution Default

"Whenever possible, you should call the functions in parallel. If there is no strict dependency, you should call them in parallel."

Performance Impact: 3-10x faster task completion

4. Context from Multiple Sources

Intelligent context gathering:

1. Direct file content
2. Diagnostics/errors
3. AGENTS.md (if exists)
4. Web search (when needed)
5. Codebase semantic search

---

🔐 Security Features

Comprehensive Security Instructions:

1. Never Log Secrets:

   "Never introduce code that exposes or logs secrets and keys. Never commit secrets or keys to the repository."

2. Defensive Security Only:

   • Security analysis: ✅ Allowed
   • Vulnerability explanations: ✅ Allowed
   • Exploit creation: ❌ Forbidden
   • Malicious code: ❌ Forbidden

3. Git Security:

   • Check for secrets before staging
   • Review all changes before commit
   • No force push without explicit permission

4. Bash Command Safety:

   "When you run a non-trivial bash command, you should explain what the command does and why you are running it."

---

💡 Best Practices Extracted

From Claude Code Prompts:

1. Conciseness is King:

   • 1-3 sentence responses
   • No preamble or postamble
   • Stop immediately after task completion

2. TODO Transparency:

   • Create plan before starting
   • Mark in-progress before work
   • Complete immediately after finishing
   • Give user visibility

3. Parallel by Default:

   • Independent operations run simultaneously
   • Serialize only when necessary
   • 3-10x performance improvement

4. No Code Comments:

   • Explain in chat, not in code
   • Only add when explicitly requested
   • Keep code clean

5. Read Before Edit:

   • Always read file first
   • Understand context
   • Make informed changes

6. Verify Changes:

   • Check after editing
   • Run tests when applicable
   • Ensure success

7. Fail Fast:

   • If same error persists, stop after 3 attempts
   • Ask user for guidance
   • Don't loop indefinitely

---

🎯 Comparison to Other Tools

vs. Cursor:

• Claude Code: Terminal-first, CLI tool
• Cursor: Visual IDE, GUI-focused

vs. GitHub Copilot:

• Claude Code: Autonomous agent, multi-file
• Copilot: Inline completions, autocomplete

vs. Windsurf:

• Claude Code: More mature, simpler architecture
• Windsurf: Cascade architecture, newer patterns

vs. Devin:

• Claude Code: Developer tool, terminal-based
• Devin: Autonomous developer, full project ownership

---

📈 Evolution & Updates

Observable Changes:

• October 2024: Latest captured version
• Increased emphasis on conciseness
• TODO system maturity
• Git workflow refinement
• Security instructions expanded

Future Directions:

• Likely continued conciseness optimization
• More tool additions
• Better context management
• Enhanced parallel execution

---

🎓 Learning Resources

From Claude Code:

Prompt Engineering:

• How to achieve extreme conciseness
• Tool architecture for parallel execution
• Security instruction design
• TODO/progress tracking patterns

Software Engineering:

• Git workflow best practices
• Code review patterns
• Testing strategies
• Security-first development

AI Design:

• Multi-tool orchestration
• Context management
• Error handling
• User communication

---

📊 Key Statistics

Metric	Value
Total Tools	20+
Prompt Length	~15,000 tokens
Response Target	1-3 sentences
Parallel Execution	Default
Security Rules	10+ explicit
Git Commands	Structured workflow

---

🔍 Technical Deep Dive

Tool Definitions (claude-code-tools.json):

Example: read_file

{
  "name": "read_file",
  "description": "Read the contents of a file",
  "input_schema": {
    "type": "object",
    "properties": {
      "path": {
        "type": "string",
        "description": "The absolute path to the file"
      }
    },
    "required": ["path"]
  }
}

Design Principles:

• Clear descriptions
• Explicit parameter types
• Required vs. optional fields
• Absolute paths (no ambiguity)

---

🎯 Use Cases

Ideal For:

1. Terminal-based workflows (developers who live in CLI)
2. Multi-file refactoring (autonomous agent handles complexity)
3. Git-heavy workflows (structured commit process)
4. Security-sensitive projects (defensive security only)
5. Quick iterations (extreme conciseness = fast responses)

Not Ideal For:

1. Visual debugging (no GUI)
2. Beginners (terminal-focused)
3. Windows users (bash-centric, though adaptable)

---

🌍 Impact & Influence

Claude Code's Influence on Other Tools:

1. TODO System:

   • Adopted by Amp, Windsurf
   • Now a standard pattern

2. AGENTS.md:

   • Spreading across tools
   • Becoming de facto standard

3. Conciseness Mandate:

   • Influenced Cursor, others
   • Token economics driving design

4. No Comments Philosophy:

   • Controversial but spreading
   • AI explanations > code comments

5. Parallel Execution:

   • Performance insight
   • Now emphasized in many tools

---

📚 Academic Interest

Research Topics:

1. Conciseness in AI Assistants:

   • Impact on user experience
   • Token cost savings
   • Information density

2. Multi-Tool Orchestration:

   • Parallel vs. serial execution
   • Performance benchmarks
   • Tool dependency graphs

3. Security in AI Assistants:

   • Defensive security only approach
   • Secret detection patterns
   • Malicious use prevention

4. Context Management:

   • AGENTS.md pattern effectiveness
   • Long-term memory vs. per-project context
   • Optimal context size

---

🔗 Related Resources

• Anthropic Documentation: https://docs.anthropic.com
• Claude API: https://console.anthropic.com
• Community: https://discord.gg/anthropic
• Research Papers: Anthropic publications on AI safety

---

🤝 Contributing

Found updates or improvements? See CONTRIBUTING.md

To Add:

• Newer versions of prompts
• Tool definition updates
• Usage examples
• Comparative analysis

---

⚖️ License & Attribution

Provider: Anthropic PBC
Documentation Purpose: Educational and research
Usage: These prompts are documented under fair use for:

• Understanding AI assistant design
• Comparative analysis
• Educational purposes
• Research

Source Attribution: Anthropic Claude Code

---

🙏 Acknowledgments

• Anthropic Team for Claude and Claude Code
• Community for sharing insights
• Contributors to this repository

---

📞 Contact

For questions about Claude Code specifically:

• Anthropic Support: support@anthropic.com
• Documentation: https://docs.anthropic.com

For this repository:

• Issues: GitHub Issues
• Discussions: GitHub Discussions

---

Last Updated: 2025-01-02
Claude Code Version: October 2024
README Version: 1.0