· AI Engineering · 4 min read
The Study Approach That Works
The exam tests architectural judgment — your ability to choose correctly in trade-off situations. The fastest path to that judgment is hands-on practice, not document reading. Read the docs for concepts, then immediately implement them in real code.
Target: one working implementation per domain, not just notes.
8-Week Roadmap
Week 1 — Foundations and Model Selection (Domain 1)
Goal: Know the model tiers cold; answer every Domain 1 question correctly.
| Day | Activity |
|---|---|
| Mon | Read the Claude model overview docs; study Opus / Sonnet / Haiku capability differences |
| Tue | Build a model router: classify request complexity, pick the cheapest sufficient model |
| Wed | Study extended thinking — enable it, test it, understand when it helps vs. hurts |
| Thu | Do 20 Domain 1 practice questions; review every wrong answer |
| Fri | Read the Anthropic pricing page; practice cost estimation for common usage patterns |
Week 1 deliverable: A Python function that correctly routes 10 different prompt types to the right model with written justifications.
Week 2 — Prompt Engineering Theory (Domain 2)
Goal: Write a correct system prompt for any persona on the first try.
| Day | Activity |
|---|---|
| Mon | Study system prompt structure: persona, constraints, output format, assumed context |
| Tue | Write system prompts for 5 different personas (support bot, code reviewer, data extractor, safety classifier, document analyst) |
| Wed | Study few-shot prompting — when it helps, how many examples, how to structure them |
| Thu | Study chain-of-thought — explicit vs. structured CoT; when to use each |
| Fri | Study XML structuring for multi-part inputs; practice on a real prompt |
Week 2 deliverable: 5 working system prompts, each tested against 10 real inputs in the API.
Week 3 — Prompt Engineering Hands-On (Domain 2)
Goal: Diagnose and fix broken prompts; pass Domain 2 practice questions at 85%+.
| Day | Activity |
|---|---|
| Mon | Take a bad prompt that produces inconsistent output — iterate until it is reliable |
| Tue | Practice few-shot: build a JSON extraction prompt that works 99%+ of the time |
| Wed | Practice CoT: pick a multi-step reasoning task and design a structured thinking scaffold |
| Thu | Do 25 Domain 2 practice questions; review every wrong answer |
| Fri | Read the Anthropic prompt engineering guide end to end; fill in any gaps |
Week 3 deliverable: One production-quality JSON extraction prompt with retry logic and validation.
Week 4 — Context, Memory, and Caching (Domain 3)
Goal: Implement prompt caching and understand exactly when to use RAG vs. in-context.
| Day | Activity |
|---|---|
| Mon | Study the context decision tree: when to pass in-context, when to cache, when to use RAG |
| Tue | Implement prompt caching on an existing project; measure cost difference using the usage field |
| Wed | Build a basic RAG pipeline: chunk a PDF, embed, retrieve top-k, pass to Claude |
| Thu | Study conversation history management: sliding window vs. periodic summarization |
| Fri | Do 20 Domain 3 practice questions; note which caching rules trip you up |
Week 4 deliverable: A cached document Q&A system with measurable cost metrics.
Week 5 — Tool Use (Domain 4, Part 1)
Goal: Build a working agentic loop from scratch with correct tool definitions.
| Day | Activity |
|---|---|
| Mon | Study tool definition anatomy: name, description, input_schema, required fields |
| Tue | Write tool definitions for 5 tools; test that Claude calls them correctly |
| Wed | Implement the full agentic loop: tool_use → execute → tool_result → loop |
| Thu | Study stop_reason values; handle max_tokens, end_turn, and tool_use correctly |
| Fri | Do 15 Domain 4 practice questions on tool use; review wrong answers |
Week 5 deliverable: A working agentic loop that can call 3 real tools (e.g., web search, calculator, database query).
Week 6 — Multi-Agent Systems (Domain 4, Part 2)
Goal: Build a two-agent system with inter-agent validation.
| Day | Activity |
|---|---|
| Mon | Study orchestrator–worker pattern; draw it as an architecture diagram |
| Tue | Implement an orchestrator that delegates to two worker agents |
| Wed | Add input validation between orchestrator and workers; test with adversarial inputs |
| Thu | Study when NOT to use multi-agent; practice the “is this simpler as one context?” decision |
| Fri | Do 15 Domain 4 practice questions on multi-agent; review all wrong answers |
Week 6 deliverable: A two-agent system with schema validation and a written justification for why it needs to be multi-agent.
Week 7 — Safety and Production Deployment (Domain 5)
Goal: Build production guardrails; understand the full deployment checklist.
| Day | Activity |
|---|---|
| Mon | Study Constitutional AI and Anthropic’s safety philosophy |
| Wed | Implement input sanitization (PII redaction) and output validation (topic classifier) |
| Thu | Study prompt injection: read the attack patterns, implement defenses |
| Thu | Implement retry with exponential backoff; add streaming to a user-facing app |
| Fri | Build a cost monitoring logger; do 20 Domain 5 practice questions |
Week 7 deliverable: A Claude application with full production hardening: sanitization, validation, retry, streaming, cost logging.
Week 8 — Mock Exam and Gap Review
Goal: Score ≥ 80% on the mock exam; complete the capstone architecture document.
| Day | Activity |
|---|---|
| Mon | Take the full 60-question mock exam under timed conditions (120 min) |
| Tue | Review every wrong answer; identify which domains need more work |
| Wed | Targeted gap review on weak domains |
| Thu | Write the capstone architecture document (see course capstone lesson) |
| Fri | Final review of the production deployment checklist; light review of all domains |
Resource List
| Resource | Priority |
|---|---|
| Anthropic model overview | Essential |
| Prompt engineering guide | Essential |
| Prompt caching guide | Essential |
| Tool use guide | Essential |
| Multi-agent systems | Essential |
| Anthropic usage policy | Important |
| Anthropic pricing | Important |
The One Rule That Separates Passers From Failers
You must write code, not just read about it. Candidates who read the docs but never implement caching consistently get the edge-case questions wrong. Candidates who have actually watched their cache hit rate in the usage field answer those questions instantly.
Build something real for each domain. That is the study plan.