№ CertLearn With Darin · Practice exam
Claude Certified Architect Professional, practice exam.
A full-length, interactive practice exam for CCAR-P. Every timed attempt draws sixty-three questions from a 89-question bank, weighted to the official blueprint and in the three formats you'll see on the day, with an explanation for every answer and results broken down by domain and by exam objective, so you know exactly where to spend your next study hour.
Two companions to this practice exam, written after a near miss on the real thing: the mastery guide (the reflexes the questions reward, a cue-to-answer decoder, playbooks for all 32 objectives, a Claude Code team-config cheat sheet and a three-week plan) and eleven hands-on labs that make the patterns muscle memory.
Prefer to listen or ask questions? The companion NotebookLM notebook has this practice set loaded as a source, so you can chat with the material, quiz yourself on a weak domain, and play the audio overview on a commute.
Five things the exam keeps testing
Work through the whole bank and the same five ideas carry most of the marks. Read these first, then watch for them in the questions. The numbers in brackets point to the questions that test each one (each question shows its number while you take the exam). Skip to the practice exam ↓
Use the simplest architecture the problem needs.
The ladder is single augmented call → fixed workflow → autonomous agent → multi-agent system, and every step up has to earn its place. Fixed workflows win when the steps are known, identical every time and individually auditable. Agents win when the path only shows up mid-task. Multiple agents are worth it only when sub-tasks need genuinely different tools or context, or can run in parallel. A supervisor is the answer when you need guaranteed ordering, an audit trail and the ability to halt. Tool-selection accuracy dropping as capabilities pile up is the signal to split an agent behind a router. (#1.1, 1.2, 1.4, 1.8, 1.9, 1.10, 1.11, 1.12)
Diagnose before you treat.
"Use a bigger model", "lower the temperature", "make the prompt longer" and "write the rule in capitals" are almost always distractors. The right answer names the actual failure and fixes it where it lives. Skipped items on a long task are a decomposition problem. A rule buried mid-prompt is a placement problem. Wrong tool picks come from overlapping tools. Failures with the code, model and prompt unchanged mean look at traces and input drift first. Rising cost means analyze where the spend goes before pulling any lever. (#1.3, 2.4, 3.2, 3.10, 3.13, 4.5, 4.9, 7.3, 7.11)
Security and safety live in the system, never in the prompt.
Authorization belongs in the access-control layer with per-user credentials, not in an instruction to "only return this user's data". Retrieved content is untrusted data, never instructions. Defense in depth means delimiting untrusted input and giving tools least privilege. Match the control to the risk: preventative guardrails for things that must never happen, a human approval gate only where an action is consequential or irreversible, monitoring and audit for trend detection. Compliance obligations such as GDPR and health data shape the architecture before any data flows. (#3.3, 3.15, 3.16, 5.1, 5.2, 5.3, 5.5, 5.7, 5.8, 5.9, 5.10, 5.11, 5.12, 7.6)
Give the model only what it needs, well structured.
Retrieve the relevant sections instead of stuffing the context window. Chunk by document structure, not fixed sizes. Use hybrid keyword and semantic search when users type exact identifiers like part numbers, and version metadata so superseded documents never get served. Keep the static prefix first and stable so prompt caching can hit (a timestamp at the top breaks it). Load specialized instructions and tools on demand. Put critical rules at the start or end of the prompt, and show the output format with an example instead of describing it. (#2.2, 2.4, 2.5, 2.6, 2.7, 2.9, 3.4, 3.5, 3.9, 3.11, 7.9)
Judge everything against business-defined success, with numbers.
"Good enough to launch" means task-specific metrics with agreed thresholds, not anecdotes or public benchmark scores. Build eval sets from real queries plus deliberate edge cases, score subjective quality at scale with an LLM judge calibrated against human ratings, and move offline gains through a limited A/B test and model upgrades through the full suite plus a canary rollout. Pick the smallest model that meets the target, and quantify trade-offs against the SLA. With stakeholders, turn "100% accurate" into measurable acceptance criteria, run discovery before building ("we need a chatbot" is a solution, not a problem), and report against the success criteria the sponsor agreed. (#1.5, 1.13, 2.1, 2.10, 3.6, 4.1, 4.2, 4.3, 4.4, 4.6, 4.11, 4.12, 4.13, 6.1, 6.2, 6.3, 6.6, 6.10, 6.11)
When two answers both look good, pick the one that fixes the root cause with the smallest mechanism that can be measured or enforced. The distractor adds capacity, adds instructions, or hopes. — TWD
How the practice set is built
Anthropic's exam guide for Claude Certified Architect – Professional comes with three sample questions, and there's no official practice exam yet. This fills that gap with a 89-question bank. Each full practice exam draws 63 questions, the live exam's item count, with each domain's share set by the blueprint weights, so a retake gives you a different mix.
Every question is also tagged with one of the 32 objectives Anthropic's score report uses, and your results break down by objective the same way. The bank was expanded after real exam feedback showed where the first version was thin: Claude Code configuration for teams, guardrail design, observability at scale, evaluation metrics, authorization, and stakeholder and lifecycle work now have dedicated questions.
| Domain | Weight | Per exam | In the bank |
|---|---|---|---|
| 1 · Solution Design & Architecture | 17% | 11 | 13 |
| 2 · Claude Models, Prompting & Context Engineering | 13% | 8 | 10 |
| 3 · Integration | 19% | 12 | 17 |
| 4 · Evaluation, Testing & Optimization | 16% | 10 | 13 |
| 5 · Governance, Safety & Risk Management | 14% | 9 | 12 |
| 6 · Stakeholder Communication & Lifecycle Management | 14% | 9 | 13 |
| 7 · Developer Productivity & Operational Enablement | 7% | 4 | 11 |
The three question formats
- Multiple choice. One correct answer out of four. The most common format. The distractors are plausible, so expect to choose between a good answer and the best one, not between right and obviously wrong.
- Multiple response. Two or more correct answers. The question tells you how many to pick, and you only get the mark if you pick exactly the right set.
- Scenario matching. Several short scenarios, each with a drop-down of the same small option set. Classify every one. Options can be used more than once, so don't assume a one-to-one mapping. You get the mark only when every scenario is right.
How to use it
- Take the full exam cold, timed. 120 minutes for 63 questions, the same limit as the live exam. No peeking at explanations until you submit.
- Check the objective breakdown. It lists the 32 exam objectives weakest first. "Drill your weak objectives" builds a study session from every question in the bank that tests them.
- Read the breakdown, not just the score. Aim for roughly 75% or better in every domain. One weak domain can sink an otherwise good result.
- Drill the weak domains in study mode. You get the answer and the reasoning after each question, including why each wrong option falls short. That second part is where most of the learning is.
- Retake after you've built something. Your progress saves in this browser, so you can stop and resume, and your last ten attempts stay on the start screen.
These are original practice questions written against Anthropic's publicly available exam guide and its blueprint objectives. They are not actual exam content and aren't drawn from the live item bank. This page isn't affiliated with or endorsed by Anthropic. The score here is a raw percentage, not the exam's scaled score, and no practice set guarantees a pass. Use it alongside hands-on building and the official documentation.
Coming from the Foundations exam? The judgments it tests (workflow vs. agent, coordinator–subagent design, tool design, hooks vs. prompts) are the base Domain 1 and 3 build on. The Foundations study guide and labs cover them, and its exam lineup section maps where the labs fall short for CCAR-P (RAG, evaluation, observability, governance, stakeholder work).