Claude Code System Prompts — Full Extraction

What it is

This is a complete reverse-engineering of Claude Code's internal prompt architecture. It documents 110+ dynamic strings that change based on environment and configuration, extracted directly from compiled source code across 135+ versions (v2.0.14 onward).

The extraction covers agent prompts (explore, plan, batch, PR review, security review), system prompts for all modes (auto, learning, minimal, plan), 18+ built-in tool descriptions, and utility prompts. It is an essential reference for anyone building Claude Code skills, custom agents, or studying how Anthropic instructs Claude Code.

How it saves time or tokens

Understanding Claude Code's prompt architecture helps you write skills and instructions that align with its existing behavioral patterns rather than fighting against them. The extraction reveals how Claude Code prioritizes tools, delegates to sub-agents, and handles safety checks. This knowledge lets you write more effective custom prompts and avoid patterns that conflict with the system prompts. Without this reference, prompt engineering for Claude Code requires extensive trial and error.

How to use

1. Visit the repository and browse by category:

git clone https://github.com/Piebald-AI/claude-code-system-prompts
cd claude-code-system-prompts
ls -la
# agent-prompts/  system-prompts/  tool-descriptions/  system-reminders/

1. Check the CHANGELOG.md for version-by-version changes to understand how prompts evolved.

1. To customize Claude Code's behavior locally, use the tweakcc tool:

# tweakcc allows safe modification of Claude Code prompts
npx tweakcc

Example

The explore agent sub-prompt structure:

# Agent: Explore
Role: Fast codebase exploration
Capabilities:
  - Read files
  - Search with grep/glob
  - Navigate directory structures
Constraints:
  - Cannot modify files
  - Cannot run commands
  - Optimized for speed over depth

# Agent: Plan
Role: Software architecture planning
Capabilities:
  - Read files
  - Analyze dependencies
  - Generate implementation plans
Constraints:
  - Cannot modify files
  - Cannot run commands
  - Provides structured plans for the main agent

Related on TokRepo

• Prompt library — Browse more prompts and skills for AI agents.
• Featured workflows — Discover curated AI tools and frameworks on TokRepo.

Common pitfalls

• Trying to override system prompts with custom instructions that contradict them leads to inconsistent behavior. Work with the existing prompt patterns, not against them.
• Assuming prompts are static. Claude Code's prompts change between versions. Check the CHANGELOG before building tools that depend on specific prompt behaviors.
• Using extracted prompts verbatim in other AI systems. The prompts are optimized for Claude's specific capabilities and may not work well with other models.