Production-ready Claude Code configuration with multi-agent orchestration, persistent memory, auto-discovered skills, and automation hooks for building fullstack applications.
- Multi-Agent Orchestration: 8 specialized agents coordinated by a central orchestrator (Ezio)
- Persistent Memory: Agents retain context and learnings across sessions
- Auto-Discovered Skills: 60+ technical patterns activated by context keywords
- Mandatory Rules: 6 protocols enforcing quality, boundaries, and honest feedback
- Hook Automation: Logging, session archiving, context sharing, and response tracking
- Slash Commands:
/handover,/prepare-pr,/discard-changes
| Layer | Technologies |
|---|---|
| Backend | Python 3.11+, FastAPI, PostgreSQL |
| AI/Agents | Claude Code SDK, Google ADK, Vertex AI (Gemini) |
| Frontend | React + TypeScript, Vite, Tailwind CSS, TanStack Query |
| Cloud | Cloud-agnostic (Europe regions preferred) |
| Infrastructure | Terraform, GitHub Actions |
┌─────────────────────────────────────────────────────────────┐
│ Planning System (Ezio) │
│ User Request → Plan → Delegate → Synthesize → Report │
└─────────────────────────┬───────────────────────────────────┘
│
┌─────────────────┼──────────────────┐
▼ ▼ ▼
┌────────────────┐ ┌───────────────┐ ┌───────────────┐
│ Scout: Research│ │ Specialists: │ │ QA Agents: │
│ & Exploration │ │ Implementation│ │ Testing │
└────────────────┘ └───────────────┘ └───────────────┘
The orchestrator (Ezio) plans and coordinates but never executes directly. Specialists handle implementation with their own context windows.
| Nickname | Role | Responsibilities |
|---|---|---|
| Ezio | Main Orchestrator | Plans tasks, delegates to specialists, synthesizes results |
| Scout | General Worker | Research, codebase exploration, routine tasks |
| Sage | Solution Architect | Architecture decisions, design reviews, ADRs |
| Kai | AI Engineer | Backend, APIs, AI agents (Python/FastAPI/ADK) |
| Iris | Frontend Engineer | UI components, React/TypeScript, accessibility |
| Devo | DevOps Engineer | Infrastructure, deployment, CI/CD (Terraform) |
| Vera | QA Tester | Testing strategy, test automation, coverage |
| Luna | Frontend QA | Component tests, E2E, accessibility audits |
| Tier | Location | Purpose | Load Behavior |
|---|---|---|---|
| Memory | .claude/memory/ |
Project context, lessons learned | Every session |
| Docs | docs/ |
Detailed plans, ADRs, guides | On-demand |
| Skills | .claude/skills/ |
Technical patterns, code examples | Auto-discovered |
.
├── CLAUDE.md # Universal context (auto-loaded)
├── PROJECT_GUIDELINES.md # Coding standards
│
├── .claude/
│ ├── agents/ # 8 agent definitions
│ ├── commands/ # Slash commands (/handover, /prepare-pr, /discard-changes)
│ ├── rules/ # 6 mandatory protocols (auto-loaded)
│ │ ├── boundaries.md
│ │ ├── compression-protocol.md
│ │ ├── honest-feedback-protocol.md
│ │ ├── memory-protocol.md
│ │ ├── pre-work-protocol.md
│ │ └── quality-gates.md
│ │
│ ├── skills/ # 60+ patterns (auto-discovered)
│ │ ├── google-adk-patterns/
│ │ ├── deployment/
│ │ ├── testing-strategy/
│ │ ├── frontend-patterns/
│ │ ├── llm-evaluation/
│ │ ├── llm-observability/
│ │ └── ...
│ │
│ ├── memory/ # Per-agent persistent files
│ ├── hooks/ # Automation scripts (Python/Bash)
│ ├── context/ # Shared inter-agent state
│ └── settings.json # Hooks & permissions config
│
└── docs/
├── ONBOARDING.md # Comprehensive setup guide
├── SKILLS_AND_AGENTS_GUIDE.md
├── rules-reference.md
└── adr/ # Architecture Decision Records
These are auto-loaded and enforce consistent behavior:
| Rule | Purpose |
|---|---|
| boundaries | Three-tier permission system (Always Do / Ask First / Never Do) |
| compression-protocol | Summarize findings, use file:line references |
| honest-feedback | Challenge ideas, state confidence levels |
| memory-protocol | STAR format for lessons, keep memory lean |
| pre-work-protocol | Check skills before implementation |
| quality-gates | Code review checklists (cloud-agnostic) |
Skills are discovered by keyword mentions. Major domains:
| Domain | Files | Coverage |
|---|---|---|
google-adk-patterns/ |
8 | ADK agents, events, state, tools, memory |
deployment/ |
7 | Cloud Run, Terraform, OAuth, Cloud Build |
testing-strategy/ |
13 | Unit, integration, E2E, accessibility, security |
frontend-patterns/ |
9 | React, routing, state, forms, animations |
llm-evaluation/ |
6 | DeepEval, RAGAS, Vertex AI evaluation |
llm-observability/ |
6 | Langfuse, Phoenix, OpenLLMetry tracing |
Example: Mentioning "ADK event persistence" auto-discovers relevant patterns from google-adk-patterns/.
Hooks provide observability, workflow automation, and protocol enforcement. All hooks are optional — customize via .claude/settings.json.
| Event | Script | Purpose |
|---|---|---|
SessionStart |
setup-environment.sh |
Environment setup, Node/Python activation, orchestrator reminder |
PreToolUse[Task] |
task_context_tracker.py |
Track agent task lifecycle (start) |
PostToolUse[*] |
tool_trace_logger.py |
Log all tool calls to JSONL |
PostToolUse[Task] |
task_context_tracker.py |
Track agent task lifecycle (end) |
PostToolUse[Task] |
context_sharing.py |
Share context between agents |
PostToolUse[Task] |
agent_response_logger.py |
Log agent responses and deliverables |
SessionEnd |
session_analytics.py |
Generate session metrics report |
SessionEnd |
session_archiver.py |
Archive session data for future reference |
| Script | Recommended Event | Purpose |
|---|---|---|
inject_shared_context.py |
UserPromptSubmit |
Injects shared context from previous agents into prompts |
Generated by hooks, safely deletable, and git-ignored:
| Location | File | Generated By |
|---|---|---|
.claude/logs/ |
tool-trace.jsonl |
tool_trace_logger.py |
.claude/logs/ |
session-metrics.jsonl |
session_analytics.py |
.claude/logs/ |
latest-session-report.txt |
session_analytics.py |
.claude/logs/ |
.active_tasks.json |
task_context_tracker.py |
.claude/context/ |
shared-context.json |
context_sharing.py |
- Disable: Remove entry from
.claude/settings.json - Enable: Add to appropriate event in
settings.json - Modify: Edit scripts in
.claude/hooks/(standard stdin/stdout) - Hook types:
command(shell) orprompt(LLM-powered, Stop events only)
- Node.js 18+ (for frontend)
- Python 3.11+ (for backend)
- Claude Code CLI installed
# Clone the repository
git clone https://github.com/YOUR_USERNAME/claude-code-setup.git
cd claude-code-setup
# Open Claude Code — everything auto-loads
claudeCLAUDE.mdloads universal context automatically- Rules in
.claude/rules/enforce protocols on every interaction - Skills in
.claude/skills/are auto-discovered by keyword mentions - Hooks in
.claude/hooks/run on lifecycle events (session start/end, tool use)
Example interaction:
You: "Add a new API endpoint for user profiles"
Ezio: [Plans task with TodoWrite]
→ [Delegates to Scout for codebase research]
→ [Delegates to Kai for implementation]
→ [Reports completion with summary]
- Create
.claude/agents/new-agent.mdwith YAML frontmatter - Add to the agent map in
CLAUDE.md - Create
.claude/memory/memory-new-agent.md
- Create file in
.claude/skills/with frontmatter:--- name: Pattern Name description: What this covers tags: [keyword1, keyword2] ---
- Add technical patterns with code examples
- Skills are auto-discovered by tag keywords
- Create
.claude/rules/new-rule.md - Rules auto-load and enforce mandatory behavior
- Use for protocols that must always apply
| Document | Purpose |
|---|---|
| ONBOARDING.md | Comprehensive setup walkthrough |
| TUTORIAL.md | Hands-on tutorial with examples |
| SKILLS_AND_AGENTS_GUIDE.md | Knowledge system deep dive |
| rules-reference.md | Complete rules documentation |
- Build exactly what's needed, nothing more
- Prefer open source; use managed services only when significantly beneficial
- Start monolithic, scale when necessary
- Agents challenge ideas before validating
- State confidence levels explicitly
- Disagree respectfully when warranted
- Orchestrator context stays strategic
- Raw data processing delegated to agents
- Compression protocol prevents context bloat
MIT