Reading is not enough. You become a Claude Architect by building. This chapter gives you structured exercises that cover every domain, plus the resources to go deeper.
This single project touches all five domains. Build it end to end.
| Component | Domain |
|---|
| Coordinator + 2 subagents (lookup + resolution) | Domain 1: Orchestration |
| MCP tools for customer lookup, order lookup, refund processing | Domain 2: Tool Design |
CLAUDE.md with team standards, .claude/rules/ for test files | Domain 3: Configuration |
| Explicit escalation criteria, few-shot examples for edge cases | Domain 4: Prompt Engineering |
| Persistent case facts block, structured error propagation | Domain 5: Context Management |
Customer Message
│
▼
┌─────────────────┐
│ Coordinator │
│ │
│ - Decompose │
│ - Route │
│ - Synthesise │
└───────┬──────────┘
┌────┴────┐
▼ ▼
┌────────┐ ┌──────────┐
│ Lookup │ │Resolution│
│ Agent │ │ Agent │
└────────┘ └──────────┘
- Happy path: customer requests refund for defective product, under $500, account verified
- Prerequisite gate test: attempt refund without account verification — should be blocked
- Hook test: refund request above $500 — should redirect to human escalation
- Multi-concern request: customer has a billing issue AND a product question — both handled
- Ambiguous customer: search returns multiple matches — agent asks for clarification
- Explicit human request: customer says "I want a human" — immediate escalation, no investigation
Build a coordinator with three subagents:
| Subagent | Role | Tools (4-5 max) |
|---|
| Web Search | Find current information | search_web, fetch_page, extract_quotes |
| Document Analysis | Analyse uploaded documents | read_document, extract_data, summarize_section |
| Synthesis | Combine findings into report | verify_fact, format_output |
| Component | What to Build |
|---|
| Non-interactive review | -p flag with --output-format json |
| Independent review instance | Separate from code generation session |
| Multi-pass architecture | Per-file analysis + cross-file integration |
| Incremental reviews | Only report new or unaddressed issues |
| Component | What to Build |
|---|
| JSON schema with tool_use | Required, optional, nullable fields, enums with "other" |
| Validation-retry loop | Send back errors, model self-corrects |
| Batch processing | Message Batches API for bulk documents |
| Confidence routing | Low-confidence fields routed to human review |
| Domain | Weight | Single Most Important Concept |
|---|
| 1. Agentic Architecture | 27% | Subagents have isolated context — pass everything explicitly |
| 2. Tool Design & MCP | 18% | Tool descriptions are the primary selection mechanism |
| 3. Claude Code Config | 20% | CLAUDE.md hierarchy — project-level for team standards |
| 4. Prompt Engineering | 20% | Be explicit — concrete criteria beat vague instructions |
| 5. Context & Reliability | 15% | Never summarise transactional data — use persistent case facts |
| Resource | What It Covers |
|---|
| Agent SDK Overview | Agentic loop mechanics, subagent patterns |
| Building Agents with the Claude Agent SDK | Hooks, orchestration, sessions |
| Agent SDK Python repo + examples | Hands-on code: hooks, custom tools, fork_session |
| MCP Integration for Claude Code | Server scoping, env var expansion, project vs user config |
| MCP specification + community servers | Protocol understanding, when to use community vs custom |
| Claude Agent SDK TypeScript repo | Tool definition patterns, structured errors |
| Claude Code official docs | CLAUDE.md hierarchy, rules directory, slash commands |
| Anthropic Prompt Engineering docs | Few-shot patterns, explicit criteria, structured output |
| Anthropic API Tool Use documentation | tool_use, tool_choice config, JSON schema enforcement |
- Building with the Claude API — core API mechanics
- Introduction to Model Context Protocol — MCP fundamentals
- Claude Code in Action — practical CLI usage
- Claude 101 — general Claude capabilities
| Resource | What It Covers |
|---|
| Claude Code CLI Cheatsheet | Commands, skills, hooks, CI/CD flags |
| Creating the Perfect CLAUDE.md | Real team configuration patterns, MCP integration |
| Everything Claude Code repo | Context management patterns, scratchpad files, compaction |
Keep these in mind across all domains:
| Anti-Pattern | Why It Is Wrong | Correct Approach |
|---|
| Parsing natural language for loop termination | Ambiguous and unreliable | Check stop_reason field |
| Arbitrary iteration caps | Cuts off useful work or wastes iterations | Let model signal via stop_reason |
| Assuming subagents share memory | They do not | Pass everything explicitly in prompt |
| Prompt-based enforcement for high-stakes | Non-zero failure rate | Programmatic hooks and gates |
| Vague tool descriptions | Causes misrouting | Specific descriptions with boundaries |
| "Be conservative" as criteria | Inconsistently applied | Define explicit categories and examples |
| Sentiment-based escalation | Frustration ≠ complexity | Use the three valid triggers |
| Summarising transactional data | Loses critical numbers and dates | Persistent case facts block |
| Single-pass review of many files | Attention dilution | Multi-pass: per-file + cross-file |
| Same session for generation and review | Reasoning bias | Independent review instance |
| Batch API for blocking workflows | No latency SLA | Synchronous for blocking, batch for overnight |
Now go build. You do not need a certificate to be a Claude Architect. You need the knowledge and the projects to prove it.