Claude Certified Architect: Foundations

Practice kit for Anthropic's Claude Certified Architect: Foundations exam. Content built from the official exam guide PDF.

Pass score: 720 / 1000. Scenario-based. Tests architectural judgment under production constraints. Not API memorization.

▶ Launch the practice exam - runs in your browser, nothing to install.

Why AI PMs should get this cert

Most AI PMs can describe what a model does. Few can specify how an agent system should behave when a tool call fails, when context runs out, or when a subagent goes off-track. This exam tests exactly that gap.

What you'll actually be able to do after passing:

1. Write specs engineers don't have to translate. You'll know the difference between a hooks-enforced constraint and a prompt-based one - and which one to specify for a given risk level.
2. Debug production failures faster. When an agent misbehaves, you'll know whether to look at stop_reason handling, context isolation, or tool description quality - instead of waiting for eng to explain it.
3. Make better build-vs-configure calls. MCP vs custom tool, tool_choice: any vs auto, project-level vs user-level config - these are PM decisions disguised as engineering ones.
4. Evaluate AI vendors without getting played. You'll recognize when a vendor is describing probabilistic behavior (prompts) as if it were guaranteed behavior (hooks).
5. Hold your ground in technical design reviews. Knowing the architecture means you can push back when a proposed design is fragile - not just defer to the team.

The cert takes about 2 weeks to prep if you use this kit. The signal it sends is disproportionate to the effort.

What's in here

practice-exam.html        - interactive exam, all 5 domains, no setup needed
cheat-sheet/
  domain1.md              - Agentic Architecture & Orchestration (27%)
  domain2.md              - Tool Design & MCP Integration (18%)
  domain3.md              - Claude Code Config & Workflows (20%)
  domain4.md              - Prompt Engineering & Structured Output (20%)
  domain5.md              - Context Management & Reliability (15%)
  sample-questions.md     - All 12 official sample questions verbatim, with explanations

How to use this

1. Read the cheat sheet for a domain: scan Key Concepts, commit the distractor-recognition cribsheet to memory, read the mnemonics.
2. Run the domain's quiz: open practice-exam.html, click the domain card, hit Start Quiz.
3. Review every miss: the explanation tells you why the wrong answers are wrong, not just what's right. That's where the pattern recognition happens.
4. Prioritize Domain 1: 27% weight, and its concepts (hooks, loop termination, context isolation) appear as distractors in D2-D5 questions. If D1 is weak, everything else suffers.
5. Run the full exam: when you're hitting 80%+ per domain, take the full exam cold.

Exam format

The 6 official scenarios

1. Customer Support Resolution Agent
2. Code Generation with Claude Code
3. Multi-Agent Research System
4. Developer Productivity with Claude
5. Claude Code for CI/CD
6. Structured Data Extraction

What each domain actually tests

D1: Agentic Architecture (27%)
Stop_reason handling. Subagent context isolation. When to use hooks vs. prompts. PreToolUse vs PostToolUse. Idempotency keys. Failure tracing (missing topics = coordinator decomposition, not subagents). Task decomposition patterns.

D2: Tool Design & MCP (18%)
Tool descriptions are the routing mechanism. Minimal descriptions break everything. isError: false for zero results vs isError: true for actual failures. tool_choice: "any" vs "auto". Scoped tool access (4-5 per agent, not 18). .mcp.json project-level vs ~/.claude.json personal.

D3: Claude Code Config (20%)
Path prefix is the whole answer: ~/.claude/ = personal, never git-tracked; .claude/ = shared via git. Divergent team behavior = user-level config never committed. context: fork for noisy skills. -p flag for CI. Path-specific rules for file patterns spread across the codebase.

D4: Prompt Engineering (20%)
Explicit categorical criteria > "be conservative." Few-shot for borderline cases, not obvious ones. tool_use eliminates syntax errors, not semantic errors. Retry fails when the information isn't in the source document. Batch API: 50% savings, no latency SLA, no multi-turn tool calling.

D5: Context Management (15%)
Case facts block to preserve specific values outside summarized history. Three escalation triggers: explicit human request, policy gap, can't make progress. Lost-in-the-middle: put key findings at the start. Conflicting sources: annotate both, flag contested. Don't pick one.

The one frame that unlocks this exam

Prompts are probabilistic. Hooks are deterministic.

When a single failure causes financial loss, a security breach, or a compliance violation: use hooks. Always. Doesn't matter how specific or well-worded the prompt is. The exam tests this distinction relentlessly across all 5 domains.

What's NOT on the exam

Constitutional AI, RLHF, fine-tuning, embedding models, computer use, vision API, streaming/SSE, OAuth/API key rotation, rate limiting, quotas, pricing, prompt caching implementation details, specific cloud provider configs, token counting.

Don't waste time on any of it.

Other community resources

Credits

• Official exam guide: Anthropic (Claude Certified Architect: Foundations Exam Guide)
• Community cheat sheets and question patterns: Hamza Farooq - github.com/hamzafarooq/claude-certified-architect