Skip to content

bwads001/claude-code-agents

Repository files navigation

Claude Code Agent Orchestration System

This repository contains a comprehensive multi-agent orchestration system for Claude Code that transforms it into a workflow-aware development environment with specialized agents, automated quality gates, and intelligent coordination patterns.

🚀 Quick Start

Installation

  1. Clone this repository to your Claude Code configuration directory:

    git clone https://github.com/bwads001/claude-code-agents.git ~/.claude
  2. Make hooks executable:

    chmod +x ~/.claude/hooks/*.py
  3. Start using specialized workflows:

    claude  # Your alias should include --dangerously-skip-permissions
    /output-style execution  # Switch to implementation mode
    "Implement user authentication using the backend-database-engineer"

For New Projects

# Copy project template
cp ~/.claude/CLAUDE_template.md ./CLAUDE.md

# Customize the template (replace [PLACEHOLDERS] and complete TODOs)
# Create ai-docs/ directory structure
# Start development with specialized agents

🧠 Why Agentic Workflows Matter

Our Philosophy: "Human-Guided Context Orchestration"

While platforms like Augment Code use ML-driven vector search to automatically discover context at scale, we take a deliberately different approach: explicit, human-guided context management through specialized agents and documentation.

Traditional Claude Code usage suffers from context pollution and workflow inefficiency:

The Problem: Context Chaos

  • Main thread cluttered with lint errors, log files, and repetitive tasks
  • Agent prompts bloated with quality gate instructions and tool usage patterns
  • Cognitive load scattered across orchestration, implementation, and validation
  • Knowledge fragmented across conversations without persistent context
  • Manual context building - forgetting to include crucial codebase docs and patterns
  • Auto-compression context loss - vital information disappears when context limits hit
  • Inconsistent code patterns - agents can't follow established conventions without manual prompting

The Solution: Clear Context Separation

🎯 Main Thread Focus = Orchestration Only

# Traditional approach - context pollution"Let me run lint, check types, create component, test it, commit..."

# Agentic approach - pure orchestration  "I'll use frontend-ui-specialist → code-quality-reviewer → git-workflow-specialist"

🤖 Specialized Agents = Domain Expertise

  • No repetitive rules - agents focus purely on their expertise
  • Context injection hooks give agents "cheat codes" with automatic project context
  • Quality automation hooks remove lint/typecheck cruft from agent definitions
  • Knowledge persistence through ai-docs/ system across conversations

🔄 Workflow Modes = Context Clarity

# Research Mode: Pure discovery, no implementation pressure
# Planning Mode: Requirements focus, no rushed decisions  
# Execution Mode: Implementation focus, plans already decided

📚 Documentation Hooks = Automatic Context Building

Problem Solved: No more manual context assembly or forgotten references

Agents automatically receive:

  • Project structure - complete codebase layout without manual discovery
  • Available ai-docs listing - "here are your architectural guides, library patterns, etc."
  • Git context - current branch, recent changes, workflow state
  • Domain-specific focus - backend agents get DB patterns, frontend gets component guides
  • Established patterns - automatically injected from ai-docs/ based on task type

Before (Manual Context Building):

# You manually include in every prompt:"Here's the project structure: [paste], here are our DB patterns: [paste], 
   here's our component style: [paste], follow these lint rules: [paste]..."

After (Automatic Context Injection):

# Hook automatically provides context:
✅ Agent receives: "📁 Project structure, 📚 Available guides: 
   drizzle-patterns.md, component-architecture.md, 🌿 Branch: feature/auth"

The 4 C's of Context Management

Principle Traditional Claude Agentic Orchestration
Clear Mixed concerns, scattered focus Dedicated specialists, pure domains
Concise Bloated prompts with repeated rules Automated context injection
Complete Manual discovery each conversation Persistent knowledge base
Consistent Ad-hoc quality checks Automated hooks and standards

Context Compression Resistance

Problem Solved: Vital context survives conversation length limits

Issue Traditional Approach Agentic Solution
Context Loss Auto-compression drops your codebase docs ai-docs/ referenced by filename, re-injected per agent
Pattern Drift Agents forget established patterns mid-conversation Hooks re-inject patterns automatically each invocation
Reference Forgetting You forget to include DB schema in backend tasks Context injection hook provides relevant ai-docs automatically
Consistency Decay Code style drifts as conversation continues Fresh context injection prevents pattern degradation

Key Innovation: Instead of cramming context into prompts, references to ai-docs/ are injected, keeping conversations lightweight while preserving access to comprehensive knowledge.

Measurable Benefits

  • 60% fewer tokens - no repeated quality gate instructions or manual context
  • 80% less discovery time - automated context injection with established patterns
  • Zero lint noise - quality automation handles mundane checks
  • 100% pattern consistency - hooks re-inject standards automatically per agent
  • Zero context loss - ai-docs/ references survive compression, can be re-read on demand

🏗️ System Architecture

This system implements a three-layer orchestration architecture:

1. Output Styles → Workflow Mode

Location: output-styles/
Purpose: Set workflow-specific orchestration patterns

  • Research Mode: Discovery and learning focus
  • Planning Mode: Requirements and design focus
  • Execution Mode: Implementation and delivery focus

2. Universal Orchestration → Agent Coordination

Location: agents/CLAUDE.md
Purpose: Agent delegation patterns that work across all projects

  • Mandatory agent delegation rules
  • Quality gate coordination
  • Agent selection guidance

3. Project Context → Domain Specifics

Location: Project root CLAUDE.md
Purpose: Tech stack, business domain, and project-specific patterns

  • Technology stack configuration
  • Business requirements and compliance
  • Project-specific quality commands and workflows

🤖 Specialized Agents

Location: agents/
Each agent is an expert in specific domains:

Agent Purpose Use When
project-manager Planning docs, workflow orchestration Start/end of features, planning lifecycle
backend-database-engineer Database, APIs, server logic Any backend work - don't do server work yourself
frontend-ui-specialist UI components, styling, design Any frontend work - don't create components yourself
feature-architect-planner Complex feature planning Breaking down large features, architecture decisions
code-quality-reviewer Manual QA, integration testing Quality gates - automated tools handled by hooks
documentation-specialist Research, library guides, ai-docs Library research, knowledge management
git-workflow-specialist Git operations, branches, commits Any git work - don't handle git yourself
file-refactor-organizer Code organization, large files Files >300 lines, structure improvements

Reference: agents/README.md for complete agent documentation

🔧 Quality Automation System

Location: hooks/
Automated quality gates that run invisibly during development:

Hook Trigger Purpose
agent-context-injection Before agent runs Auto-inject project context and ai-docs
agent-performance-monitor Before agent runs Track usage patterns, suggest optimizations
agent-result-validator After agent completes Validate text responses meet quality standards
file-content-validator After code edits Prevent debugging cruft and compatibility bloat

Benefits:

  • No manual quality checks - hooks handle lint/typecheck automatically
  • Anti-bloat protection - blocks "backwards compatible" and debugging remnants
  • Consistent standards - same quality gates across all work
  • Agent optimization - performance monitoring and suggestions

Reference: hooks/README.md for complete hook system documentation

📚 Knowledge Management

Location: ai-docs/
Structured documentation that agents can automatically discover and use:

Directory Purpose When Agents Use
architecture/ System design patterns Before architectural decisions
business-context/ Domain requirements Understanding business needs
guides/ Implementation patterns Implementing features, integrations
library-guides/ Third-party integrations Adding dependencies, using libraries
planning/ Active development plans Planning workflow coordination
troubleshooting/ Known issues, solutions Debugging, problem-solving

Agent Integration:

  • Context injection hook automatically lists available ai-docs for agents
  • documentation-specialist maintains and creates knowledge base content
  • All agents check ai-docs before implementation to follow established patterns

Reference: ai-docs/README.md for complete documentation system guide

🎯 Workflow Examples

Research → Planning → Execution Flow

# 1. Research Phase: Understand the domain
/output-style research  
"Help me understand authentication patterns in Next.js applications"

# 2. Planning Phase: Design the solution
/output-style planning
"Design a comprehensive user authentication system with role-based access"

# 3. Execution Phase: Build the feature  
/output-style execution
"Implement the authentication system from the planning documents"

Agent Coordination Example

# Execution mode automatically coordinates multiple specialists:
"Build a user dashboard with analytics"

# Behind the scenes:
# 1. project-manager: Moves plan from backlog → active
# 2. backend-database-engineer: Creates analytics queries and API endpoints  
# 3. frontend-ui-specialist: Builds dashboard components and charts
# 4. code-quality-reviewer: Validates UX and runs integration tests
# 5. git-workflow-specialist: Handles commits, branches, and PR creation
# 6. project-manager: Moves plan to completed, coordinates human handoff

📖 Key Documentation

🚦 Best Practices

Context Clarity Principles

  1. Keep main thread pure - orchestration only, zero implementation details
  2. Let hooks handle repetitive tasks - no manual lint/typecheck in conversations
  3. Use agents for all specialized work - backend, frontend, git, quality, documentation
  4. Switch output styles intentionally - research → planning → execution workflow

Agent Coordination

  • Trust the delegation rules - agents are experts with automatic context injection
  • Batch related tasks - frontend + backend + quality for complete features
  • Review agent reports for integration guidance and next steps
  • Update ai-docs when agents discover new patterns or solutions

Workflow Mode Selection

Mode When to Use Main Thread Focus Agent Behavior
Research Unknown domain, new tech Ask questions, gather info Deep exploration, documentation creation
Planning Feature design, architecture Requirements, acceptance criteria Comprehensive planning, trade-off analysis
Execution Implementation ready Orchestrate specialists Follow plans, coordinate delivery

🤝 Contributing

Repository Structure

  • Main branch: Stable, production-ready configurations
  • Issues: Bug reports, feature requests, and improvements
  • Pull requests: Contributions following established patterns

Collaboration

  • BTCEnoch: Read-only collaborator for review and feedback
  • Community contributions welcome through GitHub issues and PRs
  • Documentation improvements especially appreciated for public usage

This system transforms Claude Code from a coding assistant into a comprehensive development orchestration platform that combines specialized expertise, automated quality gates, and intelligent workflow coordination to create more effective development experiences.

Get started: Copy this configuration to your ~/.claude/ directory and begin experiencing the power of coordinated multi-agent development.

About

🤖 Multi-agent orchestration system for Claude Code featuring specialized agents, workflow modes, automated quality gates, and comprehensive development tooling

Topics

Resources

License

Stars

Watchers

Forks

Releases

No releases published

Packages

 
 
 

Contributors