Learn Claude Code

Claude Certified Architect – Professional

Anthropic's certification for senior architects who lead Claude solutions from design to rollout. It tests judgement across the whole lifecycle: design, integration, evaluation, governance and working with stakeholders.

Take the 63-question mock examPractise by domain
At a glance

63 questions in 120 minutes, scored out of 1000 with 720 to pass. Formats include multiple choice, select-TWO, Yes/No matching and classification. Integration is the largest domain at 19%, followed by Solution Design and Architecture at 17%. Time limit, question count and pass mark are as widely reported for exam guide v1.0; confirm on Anthropic's certification pages before you book.

The domains and their weights

Solution Design and Architecture17%
Claude Models, Prompting and Context Engineering13%
Integration19%
Evaluation, Testing and Optimization16%
Governance, Safety and Risk Management14%
Stakeholder Communication and Lifecycle Management14%
Developer Productivity and Operational Enablement7%

Solution Design and Architecture 17%

  • Use a fixed workflow when steps are predictable; use an agent when the path depends on what it finds.
  • Let hard constraints such as data residency and compliance decide the design before cost or convenience.
  • Place human checkpoints on consequential, irreversible actions.
  • Design for failure: timeouts, fallbacks and clear escalation.

Practise 17 Solution Design and Architecture questions →

Claude Models, Prompting and Context Engineering 13%

  • Match the model tier to the task: capable models for complex reasoning, fast ones for volume.
  • Split overloaded prompts into sequenced steps with structured hand-offs.
  • Separate instructions and inputs with XML-style tags, and show examples.
  • Budget the context window: retrieve what matters, cache repeated prefixes.

Practise 13 Claude Models, Prompting and Context Engineering questions →

Integration 19%

  • Use tool schemas and structured outputs where other systems consume the result.
  • Retry transient errors with backoff and make writes idempotent.
  • Choose batch for large jobs that can wait and synchronous calls for anything a person is waiting on.
  • Scope MCP servers and credentials to what each integration needs.

Practise 19 Integration questions →

Evaluation, Testing and Optimization 16%

  • Define success in business terms before building, then turn it into a representative eval set.
  • Include edge cases and adversarial inputs, not only typical ones.
  • Re-run evals on every prompt change and model upgrade to catch regressions.
  • Measure cost and latency before and after optimising.

Practise 16 Evaluation, Testing and Optimization questions →

Governance, Safety and Risk Management 14%

  • Treat external content as untrusted: prompt injection can arrive through documents, email and the web.
  • Give tools least privilege and log every consequential action.
  • Keep personal data within the policy and region it belongs to.
  • Red-team before launch and have an incident plan.

Practise 14 Governance, Safety and Risk Management questions →

Stakeholder Communication and Lifecycle Management 14%

  • Explain trade-offs in terms stakeholders care about: risk, cost, accuracy and time.
  • Set honest expectations about accuracy and where people stay in the loop.
  • Roll out in phases with agreed success measures.
  • Plan model upgrades and deprecations as part of the lifecycle.

Practise 14 Stakeholder Communication and Lifecycle Management questions →

Developer Productivity and Operational Enablement 7%

  • Share Claude Code standards through a committed CLAUDE.md, settings and permissions.
  • Package repeatable work as custom commands and skills.
  • Use claude -p for automation in CI with tightly scoped tools.
  • Measure productivity by outcomes, not by lines of code generated.

Practise 7 Developer Productivity and Operational Enablement questions →

Further practice

For a second set of questions written by someone who has sat the exam, Matthew Purcell publishes an independent Architect – Professional (CCAR-P) practice set with answers and rationales.

Try three sample questions

Pick your answer, then open the reveal. The full bank has 100 questions.

Sample 1Solution Design and Architecture
You are designing a claims intake system for Corvane Insurance. Every claim passes through the same five steps: read the form, check the policy number, classify the claim type, extract the loss amount and file it. About 4,000 claims arrive each day. Which architecture fits best?
  1. An autonomous agent that decides which of the five steps to run for each claim
  2. A multi-agent system with one specialist agent assigned to each of the five steps
  3. A fixed workflow of model calls in sequence, with validation between the steps
  4. One very long prompt asking the model to do all five steps in a single reply
Reveal answer
Answer: C. The steps are known and never vary, so a fixed workflow is cheaper, faster and easier to test than an agent. Validation between steps catches errors early. An agent or multi-agent design adds cost and unpredictability without any benefit when the path is fixed.
Sample 2Integration
Northwind Insurance's claims portal calls Claude synchronously. At its 9am peak of 400 requests a minute it gets 429 errors, and when the client retries, some claims end up with duplicate notes. What is the best fix?
  1. Back off with jitter on 429s and make each note write idempotent
  2. Retry at once in a tight loop until every single call goes through
  3. Move every call to the Message Batches API to avoid the limits
  4. Drop failed requests and ask handlers to resubmit them by hand
Reveal answer
Answer: A. Exponential backoff with jitter respects the rate limit, and an idempotency key stops a retried write from creating a second note. Tight retry loops make throttling worse, batches do not suit a portal where people wait, and manual resubmission pushes the problem onto staff.
Sample 3Governance, Safety and Risk Management
You are designing a Claude assistant for Calderwood Insurance that reads claimant emails and can call a refund_premium tool. Claim emails come from the public. What is the most important control against prompt injection here?
  1. Instruct the model in the system prompt to ignore any instructions found in emails
  2. Use the most capable model tier so it is harder to trick with injected text
  3. Strip all punctuation from emails before passing them to the model
  4. Require human approval before refund_premium runs, whatever the model decides
Reveal answer
Answer: D. Untrusted email text can carry injected instructions, so the consequential action needs a control the model cannot talk its way past: a human approval gate. A system prompt warning helps but is probabilistic, punctuation stripping does not remove instructions, and a stronger model is still not a guarantee.

A four-week study plan

  1. Week 1, design and models. Solution design and model, prompt and context decisions. Start from the Architect – Foundations material if it is new to you.
  2. Week 2, integration. The largest domain: APIs, tools, MCP, structured outputs, errors and data flow.
  3. Week 3, evaluation and governance. Evals, regression testing, monitoring, risk, safety and audit.
  4. Week 4, people and mock exams. Stakeholders, rollout and enablement, then the timed mock until you score 800+.

Frequently asked

How is Professional different from Architect – Foundations?

Foundations focuses on building agentic systems with Claude Code, the Agent SDK and MCP. Professional widens to the whole lifecycle: integration with existing systems, evaluation, governance and communicating with stakeholders.

How long is the exam and what is the pass mark?

It is widely reported as 63 questions in 120 minutes, scored from 100 to 1000 with 720 to pass. Check Anthropic's certification pages for the current details.

What question formats does it use?

Multiple choice, multiple response (select TWO), Yes/No matching and classification. Our mock exam includes all four.

Independent practice material, not affiliated with Anthropic. Questions are original and not taken from any real exam. Always confirm exam details on Anthropic's official certification pages.