Claude Certified Architect – Foundations Practice Test

How to Study for the Claude Certified Architect – Foundations Exam

Anchor your study to the five blueprint domains, weighted as the exam weights them. Agentic Architecture & Orchestration (27%) is the biggest: know the agentic loop cold — continue while stop_reason is "tool_use", append tool results to history, terminate on "end_turn" — plus the hub-and-spoke coordinator pattern, subagent context isolation (subagents never inherit the parent's history; pass findings explicitly in the prompt), parallel spawning via multiple Task calls in one response, and when hooks (deterministic enforcement) beat prompt instructions (probabilistic guidance). The exam loves the hooks-vs-prompt distinction: if a business rule must never be violated, the answer is programmatic enforcement.

For Tool Design & MCP (18%), remember that tool descriptions are the model's primary selection signal — most misrouting questions are answered by 'fix the descriptions' or 'reduce/scope the toolset,' not by routing layers or temperature. Know structured MCP error responses (errorCategory, isRetryable, isError in the result so the model can self-correct), tool_choice semantics (auto/any/forced), and .mcp.json project scoping with ${ENV_VAR} expansion for secrets. For Claude Code (20%), map the memory hierarchy (user ~/.claude/CLAUDE.md is per-machine; project CLAUDE.md ships with the repo — the answer to every 'works on my machine, fails in CI' question), .claude/rules/ with paths globs, skill frontmatter (context: fork isolates verbose output; allowed-tools scopes permissions), plan mode vs direct execution, and headless -p with --output-format json for CI.

For Prompt Engineering & Structured Output (20%), the recurring patterns are: explicit categorical criteria beat vague instructions (and beat 'be conservative') for false-positive problems; few-shot examples fix format inconsistency across varied documents; tool use with a JSON schema guarantees syntax but never semantics; nullable fields prevent fabrication when data is absent; retry-with-error-feedback cannot recover information missing from the source; and the Message Batches API trades a 24-hour window for a 50% discount, with custom_id as the only reliable way to match results. For Context Management & Reliability (15%), know the persistent case-facts block (vs lossy progressive summarization), lost-in-the-middle positioning (documents first, question last), trimming verbose tool outputs, objective escalation triggers (explicit human request, policy gaps — never sentiment or self-reported confidence), and distinguishing access failures from genuinely empty results in multi-agent error propagation. Finally, drill with our timed 60-question mock to build pacing under exam conditions.

The remaining domain is context management, and it is the one candidates most often underestimate because it looks like an optimization concern rather than a correctness one. Start from the fact that the API is stateless: each turn resends the whole conversation, so the transcript is a budget you are spending, not a place where things are quietly remembered. From there, prompt caching follows a single rule that answers most of the questions in this area — the cache matches on a prefix, so stable material belongs first and volatile material last, and any byte that changes early invalidates everything after it. That is why a per-request timestamp near the top of a prompt silently destroys the cache, and why changing the tool set or the model mid-conversation is a full rebuild rather than a partial one. Know the strategies for a conversation that outgrows its window — summarizing earlier turns versus clearing stale tool results — and be ready to say which one preserves the information the task still needs. Verify with the usage figures on each response rather than assuming a cache is working.