Skip to content

Leonxlnx/agentic-ai-prompt-research

Folders and files

NameName
Last commit message
Last commit date

Latest commit

 

History

6 Commits
 
 
 
 

Repository files navigation

Claude Code System Prompts

A research project exploring how modern agentic AI coding assistants work under the hood. This repository contains our best understanding of the prompt architecture, agent coordination patterns, and security mechanisms that power tools like Claude Code.

Everything here is based on behavioral observation, output analysis, community discussions, and publicly shared information. These are reconstructed approximations, not verbatim copies. The actual implementation may differ significantly.

What This Project Is

This is an educational deep-dive into the design patterns behind agentic coding assistants. We analyze how these systems:

  • Assemble dynamic system prompts at runtime
  • Coordinate multiple specialized sub-agents
  • Classify and auto-approve tool calls safely
  • Manage context windows through intelligent compaction
  • Handle memory, skills, and user preferences

The goal is to help AI engineers, researchers, and builders learn from these architectural patterns and apply them in their own projects.

What This Project Is Not

This is not a leak, dump, or direct copy of any proprietary system. The prompts documented here are our best reconstructions based on observable behavior. They represent one interpretation of how these systems likely work.

Documented Patterns

Core Identity

# Pattern Description
01 Main System Prompt How the master prompt is dynamically assembled from modular sections
02 Simple Mode Minimal prompt variant for lightweight operation
03 Default Agent Prompt Base instructions inherited by all sub-agents
04 Cyber Risk Instruction Security boundaries between authorized and prohibited actions

Orchestration

# Pattern Description
05 Coordinator System Prompt Multi-worker orchestration with phased workflows
06 Teammate Prompt Addendum Communication protocols for multi-agent collaboration

Specialized Agents

# Pattern Description
07 Verification Agent Adversarial testing agent that validates implementations
08 Explore Agent Read-only codebase exploration with no-modify constraints
09 Agent Creation Architect Generates new agent configurations from requirements
10 Status Line Setup Agent Terminal status line configuration across shells

Security and Permissions

# Pattern Description
11 Permission Explainer Risk assessment before tool approval
12 Auto Mode Classifier Multi-stage security classifier for autonomous tool execution

Tool Descriptions

# Pattern Description
13 Tool-Specific Prompts How individual tools (Bash, Edit, Agent, etc.) describe themselves

Utility Patterns

# Pattern Description
14 Tool Use Summary Generating concise labels for completed tool batches
15 Session Search Semantic search across past conversation sessions
16 Memory Selection Selecting relevant memory files for query context
17 Auto Mode Critique Reviewing user-written classifier rules
20 Session Title Lightweight title generation for session management
29 Agent Summary Background progress updates for sub-agents
30 Prompt Suggestion Predicting likely user follow-up commands

Context Window Management

# Pattern Description
21 Compact Service Conversation summarization strategies for long sessions
22 Away Summary Brief session recaps for returning users

Dynamic Behaviors

# Pattern Description
18 Proactive Mode Autonomous background operation with pacing controls
23 Chrome Browser Automation Browser extension integration patterns
24 Memory Instruction Hierarchical memory loading and override semantics

Skill Patterns

# Pattern Description
19 Simplify Skill Multi-agent parallel code review pattern
25 Skillify Skill Interview-based skill creation workflow
26 Stuck Skill Session diagnostic and recovery patterns
27 Remember Skill Memory organization and promotion workflow
28 Update Config Skill Configuration management patterns

Architectural Observations

Dynamic Prompt Assembly

Based on our analysis, the system prompt appears to be assembled through a pipeline of modular builders:

Prompt Assembly Pipeline
    |
    |   Cacheable Prefix (stable across sessions)
    |-- Identity and safety instructions
    |-- Permission and hook configuration
    |-- Code style and error handling rules
    |-- Tool preferences and usage patterns
    |-- Tone, style, and output rules
    |
    |   Cache Boundary
    |
    |   Dynamic Suffix (changes per session)
    |-- Available agents and skills
    |-- Memory file contents
    |-- Environment context (OS, directory, git state)
    |-- Language and output preferences
    |-- Active MCP server instructions
    |-- Context window management directives

Security Classification

The auto-approval system appears to use a multi-stage approach:

  1. A base classifier with predefined rules for safe and unsafe operations
  2. User-configurable overrides that can extend or restrict the defaults
  3. A fast first pass, with extended reasoning as fallback for ambiguous cases

Memory Hierarchy

Loading Order (earliest = lowest priority):
    |
    |-- Enterprise/managed configuration
    |-- User global preferences
    |-- Project-level instructions (shared)
    |-- Project rules directory
    |-- Local overrides (private, not committed)
    |
    |   Supports transitive file inclusion
    |   Conditional injection via path-based filtering

Use Cases

This research is useful for:

  • AI engineers building their own agentic coding tools
  • Prompt engineers studying production-grade prompt architectures
  • Security researchers understanding how autonomous AI tools manage permissions
  • Students and educators learning about multi-agent system design

Repository Structure

claude-code-system-prompts/
    README.md
    prompts/
        01-30 documented patterns (see catalog above)

Disclaimer

This is an independent research project. All content represents our analysis and approximations based on publicly observable behavior. This project is not affiliated with, endorsed by, or connected to Anthropic. All trademarks belong to their respective owners. If any content owner has concerns, please open an issue and we will address it promptly.

About

Research into how agentic AI coding assistants work — reconstructed prompt patterns, agent coordination, and security classification

Topics

Resources

Stars

Watchers

Forks

Releases

No releases published

Packages

 
 
 

Contributors