Study guide · free · updated August 2026
The complete Claude Certified Architect — Professional study guide
Everything you need to prepare for the Claude Certified Architect — Professional exam, in one page: what the 38 blueprint objectives actually ask of you, how the 19% integration domain should reshape your revision timetable, a four-week plan, and what to do in the last 48 hours. It is free, it takes about fifteen minutes to read, and every section links to practice questions on the same topic.
In short
- Study by weight, not by curiosity. Integration (19%), solution design (17%) and evaluation (16%) are over half the exam between them.
- Most items are judgement, not recall. Two options are usually both defensible; the constraint stated in the stem is what decides between them.
- Sit a cold mock early. A score from week one is worth more than a week of reading, because it tells you which of the seven domains is actually your problem.
- Aim well clear of 720, not at it. A pool you have seen before flatters you; a live form will not.
- Experience beats hours. If you have shipped a retrieval pipeline and an agent loop, you are revising vocabulary. If you have not, you are learning the subject, and that takes longer than any plan says.
Who this exam is for
It is aimed at people who design systems built on Claude rather than people who use Claude. The distinction runs through the whole blueprint: you are rarely asked what a feature does, and very often asked which of two reasonable designs fits a stated constraint. Domains 5 and 6, governance and stakeholder communication, are 28% of the exam between them and contain almost no code at all.
That shapes who finds it hard. Strong engineers who have never had to justify a latency budget to a finance team, or explain why an SLA cannot promise per-response correctness, tend to lose marks in D6 rather than D3. Consultants with the opposite background lose them in retrieval and integration. It is worth being honest with yourself about which one you are before you plan anything.
The shape of the exam
A scored form is 63 questions in 120 minutes, and the pass mark is 720 on a scale that runs from 100 to 1000. The scale is the part people misread: 720 is not 72% of the items. Raw performance is converted onto a fixed range, so the number of questions you need right is not fixed either. The format article works through the mechanics, including the three question types and why multiple-response items are the most expensive place to be careless.
What is on the Claude Certified Architect — Professional exam
The exam is organised into 7 weighted domains covering 38 objectives in total. The weight is the share of a scored form each domain takes, so it is also the share of your revision it deserves. Integration is the heaviest at 19%; developer productivity is the lightest at 7%.
| Domain | Weight | Objectives | On a mock | In the pool |
|---|---|---|---|---|
| D1Solution Design & Architecture | 17% | 6 | 11 | 34 |
| D2Models, Prompting & Context Engineering | 13% | 5 | 8 | 26 |
| D3Integration | 19% | 8 | 12 | 38 |
| D4Evaluation, Testing & Optimization | 16% | 6 | 10 | 32 |
| D5Governance, Safety & Risk Management | 14% | 5 | 9 | 28 |
| D6Stakeholder Communication & Lifecycle | 14% | 5 | 9 | 28 |
| D7Developer Productivity & Enablement | 7% | 3 | 4 | 14 |
weights from the published exam blueprint · pool counts from this site’s question bank
A four-week plan that respects the weights
This assumes six to eight hours a week and some production experience. Compress it if you have more of either; stretch week two if you have less.
- 1
Week 1 · find out where you actually are
Sit a timed mock cold, before revising anything. It will feel premature and that is the point: the per-domain breakdown is the only unbiased map of your weaknesses you will ever get, and it is worth more than the marks it costs your ego. Then read this guide's domain sections and the objective notes for anything the report flagged.
- 2
Week 2 · the heavy three
Integration, solution design and evaluation, in that order. That is 52% of the exam. Drill each by domain, then use the objective drills for any row you keep missing. Read the explanation on every item you get wrong, and on every item you got right by elimination. Those are the ones that will catch you on a form with better distractors.
- 3
Week 3 · the soft-looking three
Governance, stakeholder communication and prompting: 41% together, and the domains technical candidates skip because they look like common sense. They are not. Run the 120-question review across a few sittings to cover the ground properly, and pay attention to who the stem says you are talking to.
- 4
Week 4 · sit forms, not chapters
Two or three more timed mocks, spaced, with the mistakes deck cleared between them. You are no longer learning material; you are practising pace and the discipline of flagging and moving on. Stop revising the day before.
If you have less than four weeks, cut week 3 rather than week 1. Sitting a mock without knowing where you stand is the step nothing else substitutes for. And a 21-question quiz takes ten minutes if a full form does not fit today.
How to use practice questions properly
The single most common way to waste a question bank is to grind it for the score. A practice item is worth what you learn from the rationale, and the score is only a measurement instrument, which is why the modes on this site are kept separate.
| Mode | Length | Use it for | Predicts your score |
|---|---|---|---|
| Timed mock | 63 items, 120 min | Measuring where you stand under real pressure | Yes, the only one that does |
| Guided test | 63 items, untimed | Learning the reasoning behind each key | No; it shows you the answer |
| Extended review | 120 items, untimed | Covering every objective at least once | No |
| Quick quiz | 21 items | A ten-minute temperature check | No |
| Domain and objective drills | Varies | Repairing one specific weakness | No |
| Mistakes deck | Whatever you have missed | Making a fix stick. Two correct in a row retires an item | No |
The reason for that last column is worth stating plainly: a score from a mode that showed you the answer predicts nothing. This is not strictness for its own sake. It is the only way the projection on the dashboard means anything.
Every exam objective, with practice questions
The score report you get after a real sitting grades you objective by objective, not just domain by domain. Every objective below has its own drill drawn from the question pool, so a report that says you were weak on one row maps to somewhere specific to go.
D1Solution Design & Architecture
- 1.1Translate business problems into Claude-based AI solutions(6)
- 1.2Design end-to-end architectures (input → processing → output → feedback loops)(6)
- 1.3Select appropriate architectural patterns (workflow, agentic, augmented LLM)(6)
- 1.4Design multi-agent systems and orchestration strategies(5)
- 1.5Apply decomposition techniques for complex problem solving(5)
- 1.6Align solutions to business value pillars (efficiency, transformation, productivity, cost, performance SLAs)(6)
D2Models, Prompting & Context Engineering
- 2.1Select appropriate Claude models based on trade-offs(6)
- 2.2Design system prompts, templates, and guardrails(5)
- 2.3Apply prompt engineering techniques (zero-shot, few-shot, chain-of-thought)(5)
- 2.4Optimize context windows and manage token usage(5)
- 2.5Implement prompt reuse strategies (caching, modular prompts, Skills)(5)
D3Integration
- 3.1Evaluate tool/agent configuration for capability bloat(5)
- 3.2Analyze authentication and authorization requirements to identify security gaps(5)
- 3.3Evaluate accuracy-latency trade-offs and justify configuration decisions(5)
- 3.4Analyze observability challenges and select monitoring strategies at scale(5)
- 3.5Design a RAG pipeline with appropriate chunking and indexing strategies(5)
- 3.6Apply retrieval strategies matched to data shape and query pattern(5)
- 3.7Evaluate connection protocols and select the appropriate integration mechanism (MCP, API/CLI, agent-to-agent)(4)
- 3.8Evaluate progressive discovery vs. monolithic context strategy(4)
D4Evaluation, Testing & Optimization
- 4.1Define evaluation metrics (accuracy, latency, cost, safety, security)(6)
- 4.2Design evaluation datasets and test frameworks using mixed methodologies(5)
- 4.3Conduct A/B testing and iterative improvements(5)
- 4.4Diagnose system issues (prompt failure, hallucinations, model mismatch)(5)
- 4.5Optimize token usage, latency, and cost-performance trade-offs(5)
- 4.6Monitor system performance using logging and observability tools(6)
D5Governance, Safety & Risk Management
- 5.1Implement guardrails and safety controls(6)
- 5.2Identify risks, limitations, and failure modes of LLM systems(6)
- 5.3Apply human-in-the-loop validation strategies(5)
- 5.4Ensure compliance with regulations (e.g., GDPR, HIPAA, FedRAMP)(6)
- 5.5Address ethical AI considerations (bias, fairness, transparency)(5)
D6Stakeholder Communication & Lifecycle
- 6.1Conduct structured discovery and requirement gathering(6)
- 6.2Communicate architectural decisions and trade-offs(6)
- 6.3Manage stakeholder feedback loops and expectation alignment (including SLAs)(6)
- 6.4Document architectures and provide implementation guidance(5)
- 6.5Support lifecycle phases (discovery, design, handoff, monitoring, iteration)(5)
The last 48 hours
- Stop learning new material. Anything you meet for the first time the day before is more likely to displace something you knew than to be recalled.
- Clear the mistakes deck. It is the highest-yield hour available to you, because every item in it is one you have personally demonstrated you get wrong.
- Re-read the domain weights. Knowing that D7 is four questions stops you spending real minutes on it during the exam.
- Rehearse the pace, not the content. Just under two minutes an item, and a hard scenario is allowed to cost four as long as a recall item costs twenty seconds.
- Sleep. Genuinely the highest-leverage intervention left, and the one everybody skips.
During the exam
- 01Read the constraint before the options. Most stems bury one sentence that decides the answer: a latency budget, a compliance regime, a named audience, a data shape. Find it first and the four plausible options collapse to one.
- 02Answer the question asked. The best design in the abstract is a distractor if the stem asked for the cheapest one that meets an SLA.
- 03Flag and move. Two minutes is the budget. If an item is still open at three, flag it and go. The marks you lose by not reaching question 60 are certain; the ones you might gain by staying are not.
- 04Never leave a blank. There is no guessing penalty, so an unanswered item is strictly worse than a guess.
- 05On the second pass, trust the flags, not the doubts. Re-reading an item you were confident about is how a right answer becomes a wrong one.
If you failed
The score report grades every objective separately, which makes a retake far more plannable than most certifications. Take the rows it marked weak, open the matching objective drill from the index above, and work only those. You do not need to re-cover the domain. You need to fix the two or three rows that cost you the margin, and then sit fresh forms until the projection holds above the cut twice running.
Frequently asked questions
How long does it take to prepare for the Claude Certified Architect exam?
Four weeks at six to eight hours a week is a reasonable plan for someone who has built production systems on an LLM. Without that background, expect longer, because the exam assumes you have made these trade-offs rather than read about them. The most reliable way to find out is to sit a timed mock cold in your first week and read the gap between your scaled score and the 720 cut.
What is the hardest domain?
Integration is the heaviest at 19% and the most technically demanding: retrieval design, authentication across an agentic system, protocol choice and observability. But the domain that most often costs experienced engineers their margin is stakeholder communication and lifecycle, at 14%, because it looks like common sense and is graded on specifics.
Are practice exams enough to pass?
They are enough to tell you whether you would pass, and enough to fix specific gaps. They are not a substitute for having designed a system with these constraints, because most items test judgement between two defensible options rather than recall of a fact. Use them to find the gaps, then close the gaps with reading and, ideally, with building something.
Do I need to write code for this exam?
No. It is an architecture exam. You need to understand what a chunking strategy or a caching boundary does and when it is the right call, not to implement one under time pressure.
Is this study guide free?
Yes. This guide, all 200 practice questions and every mode on the site are free, with no account and no sign-up. Nothing is held back behind a paid tier.
Reading is the smaller half. 63 questions, 120 minutes, scored against the 720 cut. Free, and nothing leaves your browser.
Sit a timed mockWho wrote this
Junaid Muhammad is a Senior Forward Deployed Engineer and holds the Claude Certified Architect, Professional. This guide is the material he assembled while preparing for the exam, rewritten after passing it. It is independent and unofficial: the questions are original and written against the published blueprint, not recalled from a real sitting, and the site has no affiliation with or endorsement by Anthropic. For booking, policies and the official exam guide, go to Anthropic directly.