Start here
Seven guides assume you already know what a RAG pipeline looks like, what an evaluation set is for, and what governance means once a system is live. This page is where those assumptions get earned — and where the most useful fact about this exam gets named early: it is a judgment test, and that changes how you study for it.
One honest opening. The author of this guide booked the Claude Certified Architect — Professional (CCAR-P) three hours before sitting it, prepared nothing, and passed. That is not a flex — it is the single most useful thing you can know about this exam, because it tells you what the exam is: a test of organization-level design judgment. The Foundations exam below it is where Anthropic hid the CLI flags, parameter values and file locations, and it was the harder of the two sittings. The Professional asks a different question entirely: can you make design decisions about AI systems — integration, evaluation, governance, stakeholders — and defend them? If you have shipped AI past the pilot stage, most of its questions map onto decisions you have already made, some of them badly. That experience is the preparation.
This guide treats the exam as what it is: a judgment exam about design. Every page teaches its domain from the ground up, then collects the facts the exam quotes at design level — caching economics, retrieval trade-offs, protocol choices — then drills the traps: the wrong answers that sound sensible until you have shipped.
01What the exam is
CCAR-P is the advanced tier of Anthropic’s architect track: 63 multiple-choice and multiple-response questions in 120 minutes (about 135 minutes of seat time with check-in and survey), $175, delivered through Pearson VUE online (OnVUE) or at a test centre. Multiple-response items tell you how many answers to select. You need a scaled 720 out of 1,000 to pass, the credential lasts 12 months, and renewal — unlike the initial exam — is a free, non-proctored assessment on the Anthropic Partner Academy, as long as you renew on time. Let it lapse and you retake the full exam at the full fee.
Two structural facts worth knowing before you book. The exam is criterion-referenced: you are measured against a fixed standard set by a panel of experts, not against other candidates — there is no curve to game. And your score report breaks out percent-correct by domain, which makes a failed attempt a study plan: it tells you exactly which of the seven domains to rebuild before the retake.
The rules of conduct are strict and cheap to violate: closed book (no notes, documentation, translation tools or AI assistants), a government photo ID whose name matches your registration exactly, English only, and a confidentiality agreement you accept before the first question. Cancel or reschedule at least 24 hours ahead — inside 24 hours you forfeit the fee. If you don’t pass: wait 14 days before the second attempt, 30 before the third, 90 before the fourth, up to four attempts in a rolling twelve months.
02The seven domains — study where the marks are
| Domain | Weight | What it tests |
|---|---|---|
| 01 Solution design and architecture | 17% | Turning business problems into AI solutions, end-to-end architecture with feedback loops, choosing between workflow and agentic patterns, multi-agent orchestration, decomposition, and business value pillars. |
| 02 Claude models, prompting and context engineering | 13% | Model selection trade-offs, system prompts and guardrails, zero-shot / few-shot / chain-of-thought, context window budgets, and reuse: prompt caching, modular prompts, Skills. |
| 03 Integration | 19% | The heaviest domain: tool and agent configuration discipline, authentication and authorization gaps, accuracy–latency trade-offs, observability at scale, RAG pipeline design, retrieval strategy, MCP vs API/CLI vs agent-to-agent, progressive discovery. |
| 04 Evaluation, testing and optimization | 16% | Metrics across accuracy, latency, cost, safety and security; datasets and mixed methodologies; A/B testing; diagnosing failures; token and cost optimization; production monitoring. |
| 05 Governance, safety and risk management | 14% | Layered guardrails, failure modes of LLM systems, human-in-the-loop design, regulatory compliance (GDPR, HIPAA, FedRAMP), and ethics: bias, fairness, transparency. |
| 06 Stakeholder communication and lifecycle management | 14% | The domain unique to the Professional tier: structured discovery, communicating trade-offs, expectation alignment and SLAs, documentation, and the discovery-to-iteration lifecycle. |
| 07 Developer productivity and operational enablement | 7% | Configuring Claude tools like Claude Code for teams, AI-assisted developer workflows, and supporting debugging and operations. |
Read the table twice, once for the weights and once for a pattern. Integration alone is nearly a fifth of the exam, with RAG pipeline design called out explicitly; add evaluation and the two design domains and the first four carry two-thirds of the marks. The pattern: the domains a developer would call “soft” — governance, stakeholder communication — carry 28% between them, more than prompting and context engineering. This is the professional tier’s signature: it tests the halves of the job that shipping teaches and reading alone does not.
03How it differs from Foundations
If you came through CCAR-F — or read our guide to it — calibrate like this. Foundations asks “what does this feature do, exactly?” Professional asks “when should this feature be used, what does it cost, and what happens when it fails at scale?” Same platform, different verb. Foundations questions name a flag or file and ask for its purpose; Professional questions hand you a scenario — a support agent with too many tools, a RAG system that broke after a document refresh, an executive promised 99% accuracy — and ask which decision is right or which investigation comes first.
The official sample questions set the tone. One shows an agent that can issue refunds and delete accounts when the staff only need to read tickets and draft replies; least privilege says remove the tools, not log them or wrap them in confirmations. Another shows an 8,000-token policy document resent on every request; the answer is stable-prefix ordering plus prompt caching, not truncation. A third shows confident wrong answers appearing right after a document refresh; you investigate retrieval, not model weights. Every page of this guide builds toward answering that class of question in ninety seconds.
And the felt difference, from the sitting behind this guide: the wrong answers are obviously wrong, because they are the decisions experience has already taught you not to make. Foundations walks you out convinced you failed; Professional feels like a warm-up afterwards. Do not relax on that account — the exam is fair, not easy, and “I would just do X” is not an answer until you can say why X and not the two alternatives that sound nearly as good.
04The prep that actually works
- —Read the exam guide first — linked from the official page and downloadable as a PDF. It is the authoritative source for domains, weights and the objectives every question is written against. If its domain list reads like your last twelve months of work, book the exam.
- —Skim for gaps, not coverage: docs.claude.com on retrieval patterns, evaluation, and governance, plus the MCP spec if integrations are the thin spot in your experience. You are patching holes in experience, not taking a course.
- —Do a practice set if you can find one. Fewer exist for CCAR-P than for the foundation exams — Preporato has a CCAR-P guide and the community sites are still catching up — but any scenario-style questions will calibrate you to the wording.
- —Prepare your environment, not your memory. OnVUE logistics, government ID with a matching name, Pearson’s system test the day before — the only cheap way to fail this exam is a technical disqualification before question one.