Claude Code System Prompts

An independent research project documenting the internal prompt architecture and agentic behavior of Claude Code, Anthropic's AI-powered software engineering assistant.

This documentation was produced through behavioral analysis, output observation, and study of publicly available discussions. It serves as an educational resource for understanding how modern agentic AI coding assistants are designed.

Overview

Claude Code uses a sophisticated multi-layered prompt architecture. The main system prompt is not a static string but is dynamically assembled at runtime from modular section-builder functions. A boundary marker splits it into a globally cacheable prefix and a session-specific suffix, enabling prompt caching across API calls.

Beyond the core identity prompt, the system includes specialized agent prompts, a multi-worker coordinator, a 2-stage security classifier for auto-approving tool calls, and a suite of utility prompts for memory selection, session search, and tool summarization.

Prompt Catalog

Core Identity

#	Prompt	Description
01	Main System Prompt	Dynamically assembled master prompt covering identity, behavior, tool guidance, tone, and efficiency
02	Simple Mode	Minimal 4-line prompt activated by CLAUDE_CODE_SIMPLE
03	Default Agent Prompt	Base prompt inherited by all sub-agents
04	Cyber Risk Instruction	Security boundaries for authorized vs. prohibited actions

Orchestration

#	Prompt	Description
05	Coordinator System Prompt	Multi-worker orchestrator with 4-phase workflow and concurrency rules
06	Teammate Prompt Addendum	Communication protocol for swarm and team mode

Specialized Agents

#	Prompt	Description
07	Verification Agent	Adversarial testing specialist that tries to break implementations
08	Explore Agent	Read-only codebase exploration with strict no-modify constraints
09	Agent Creation Architect	Designs new agent configurations from user requirements
10	Status Line Setup Agent	Configures terminal status line across shell environments

Security and Permissions

#	Prompt	Description
11	Permission Explainer	Explains tool risk levels before user approval
12	Auto Mode Classifier	2-stage security classifier for auto-approving tool calls

Tool Descriptions

#	Prompt	Description
13	Tool-Specific Prompts	All 30+ tool descriptions including Bash, Edit, Agent, and fork semantics

Utility and Helpers

#	Prompt	Description
14	Tool Use Summary	Generates git-commit-style labels for completed tool batches
15	Session Search	Semantic search across past conversation sessions
16	Memory Selection	Selects relevant memory files for query context
17	Auto Mode Critique	Reviews user-written auto-mode classifier rules
20	Session Title	Haiku-powered 3-7 word session title generator
29	Agent Summary	Periodic progress updates for sub-agents in coordinator mode
30	Prompt Suggestion	Predicts user follow-up commands for clickable suggestions

Context Window Management

#	Prompt	Description
21	Compact Service	Multi-variant conversation summarization with analysis/summary blocks
22	Away Summary	1-3 sentence session recap for returning users

Dynamic Sections

#	Prompt	Description
18	Proactive Mode	Autonomous agent with tick-based pacing and terminal focus awareness
23	Chrome Browser Automation	Browser extension integration: GIF recording, tab management, dialog handling
24	Memory Instruction	CLAUDE.md loading, @include directives, and frontmatter globs

Bundled Skills

#	Prompt	Description
19	Simplify Skill	Three-agent parallel review for code reuse, quality, and efficiency
25	Skillify Skill	Interview-based skill creation, generates SKILL.md files
26	Stuck Skill	Internal diagnostic for frozen or slow sessions
27	Remember Skill	Promotes auto-memory entries into CLAUDE.md files
28	Update Config Skill	Manages settings.json, hooks, and permission arrays

Architecture

Prompt Assembly

The main system prompt is built through a pipeline of section builders:

System Prompt Assembly
    |
    |   Static Prefix (globally cached)
    |-- Identity and Cyber Risk
    |-- Permission modes, hooks, reminders
    |-- Code style, security, error handling
    |-- Reversibility, blast radius
    |-- Tool preferences, parallel calls
    |-- Tone and style rules
    |-- Output efficiency patterns
    |
    |   CACHE BOUNDARY
    |
    |   Dynamic Suffix (session-specific)
    |-- Agent tools, skills, verification
    |-- Memory file content
    |-- Model overrides
    |-- Environment info (CWD, OS, git state)
    |-- Language preferences
    |-- Custom output styles
    |-- MCP server instructions
    |-- Context window management

Auto Mode Classifier

The auto-approval system uses a separate classifier prompt assembled from:

1. Base prompt with classifier instructions
2. Default rules with allow, deny, and environment sections
3. User overrides that replace sections entirely
4. 2-stage classification: Stage 1 runs fast, Stage 2 uses extended thinking if uncertain

Context Window Pipeline

User Message
    |
    v
[Micro-Compaction]  -- Cache-aware tool result deletion
    |
    v
[Compact Service]   -- Full/partial summarization
    |
    v
[Prompt Suggestion] -- Predict next user command
    |
    v
[Away Summary]      -- Session recap if user was idle

Memory System

Memory Loading Order (first loaded = lowest priority):
    |
    |-- Enterprise managed config
    |-- User global config
    |-- Project config (shared, checked in)
    |-- Project rules directory
    |-- Local config (private, git-ignored)
    |
    |   @include directives resolve transitively (max depth: 5)
    |   Frontmatter paths field enables conditional injection

Environment Variables

Variable	Effect
CLAUDE_CODE_SIMPLE	Activates the minimal 4-line system prompt
USER_TYPE=ant	Enables internal-only sections and model overrides
Feature flags	Gate proactive mode, verification agents, fork subagents, and more

Repository Structure

claude-code-system-prompts/
    README.md
    prompts/
        01_main_system_prompt.md        19_simplify_skill.md
        02_simple_mode.md               20_session_title.md
        03_default_agent_prompt.md      21_compact_service.md
        04_cyber_risk_instruction.md    22_away_summary.md
        05_coordinator_system_prompt.md 23_chrome_browser_automation.md
        06_teammate_prompt_addendum.md  24_memory_instruction.md
        07_verification_agent.md        25_skillify.md
        08_explore_agent.md             26_stuck_skill.md
        09_agent_creation_architect.md  27_remember_skill.md
        10_statusline_setup_agent.md    28_update_config_skill.md
        11_permission_explainer.md      29_agent_summary.md
        12_yolo_auto_mode_classifier.md 30_prompt_suggestion.md
        13_tool_prompts.md
        14_tool_use_summary.md
        15_session_search.md
        16_memory_selection.md
        17_auto_mode_critique.md
        18_proactive_mode.md

Purpose

This project exists as an educational resource for AI researchers, developers building agentic systems, and anyone interested in understanding the design patterns behind production-grade AI coding assistants.

The documented patterns cover topics relevant to the broader AI engineering community:

• Multi-agent orchestration and coordination
• Security classification for autonomous tool use
• Context window management and conversation compaction
• Memory systems with hierarchical override semantics
• Prompt caching strategies for latency optimization

Disclaimer

This is an independent research project. The content represents analysis and observations of Claude Code's behavior and architecture. This project is not affiliated with, endorsed by, or connected to Anthropic in any way. All trademarks belong to their respective owners.