Twenty scenario questions written to the Claude Certified Architect - Professional (CCAR-P) blueprint, spread across all seven domains in proportion to exam weight, each followed by the answer and an explanation that says why the winning option wins and why every distractor loses for that scenario. Each item drops you into an enterprise situation with a decision to make. Use the set as a diagnostic: score yourself against the rubric below, then follow the domain-by-domain map at the end to the guide that closes each gap. Five of the twenty are multiple-response items, matching the roughly one-quarter share on the real exam, so practice judging each option on its own merits. All twenty are new, written in the style of Preporato's full-length CCAR-P tests.
Start Here
If you have not yet mapped the exam, read the CCAR-P complete guide for the seven domains, scoring, and study path. For full-length timed practice, Preporato's CCAR-P practice tests offer 6 exams of 63 questions on the same blueprint, and the free sample questions show you the format before you commit.
CCAR-P exam update on December 7, 2026
Anthropic updates the CCAR-P exam content on December 7, 2026, and publishes the updated exam guide on November 17. If you test on December 7 or later, prepare from the updated guide. This article follows the current guide, and we will revise it once the new one is out. If you are nearly ready now, a date before December 7 lets you sit the version you studied for.
How CCAR-P questions are built
CCAR-P stems carry more than one constraint on purpose: an industry, a workload profile (volume, latency budget, cost pressure), a compliance or stakeholder condition, and a symptom, followed by a request for the best architectural response. The wrong options are real techniques that solve a neighboring problem: a Batch API answer when a user is waiting, a prompt instruction when the requirement is a security guarantee, a stronger model when retrieval is what failed. The exam asks for the best fit among options that would each work somewhere, so find the dominant constraint in the stem before you read the choices. On multiple-response items ("Select TWO"), every selection must survive on its own; one plausible-but-wrong pick costs the whole question.
Anthropic retired its official practice exam when delivery moved to Pearson VUE on June 30, 2026, and the exam guide's sample questions are the only official practice material left. The 20 questions below follow the same stem pattern.
Self-check rubric
Score one point per fully correct question (both selections right on Select TWO items). Take the set once, untimed, before reading any explanation.
Scoring your 20 answers
| Score | Reading | Next step |
|---|---|---|
| 17 to 20 | Strong. Your judgment matches the blueprint across domains. | Move to timed full-length tests and drill multiple-response speed. |
| 14 to 16 | Borderline. One or two domains are soft. | Use the domain map at the end and reread the guide for every domain you missed twice. |
| Under 14 | Study phase. The patterns are still forming. | Work through the complete guide and cheat sheet, then retake this set in a week. |
Domain 1: Solution Design & Architecture (17%)
Questions 1–3
An accounting firm wants Claude to prepare quarterly sales-tax filings from client ledger exports. Every filing follows the same sequence: normalize the ledger, map line items to tax codes, compute totals, populate the form, and flag anomalies for a reviewer. Auditors must trace each figure back to its stage. Which architecture pattern should the architect recommend?
Fixed steps plus an audit requirement are the triggers for a workflow pattern (several model calls in a developer-defined sequence with validation between them); each stage yields a checkable intermediate output. An agentic loop (A), where the model picks its own steps, adds cost and path variance to a path that never varies. One augmented call (B) hides five stages inside one opaque step, so a mapping error surfaces only in the finished form. Orchestrator-worker (C) suits independent, parallel subtasks; these stages are strictly sequential.
A university is designing an admissions-inquiry assistant. The design covers input handling across chat and email, retrieval over the course catalog, and response generation with tone guardrails. Nothing is planned after a reply is sent. A review board asks what will most weaken the solution over its first year. What should the architect identify?
A complete architecture has input, processing, output, and a feedback loop; the last turns production signals (staff corrections, escalations, low-rated answers) into evaluation data and the next iteration. Without it, quality drifts as the catalog and the questions change, and nobody measures the drift. A stronger model (C) raises the day-one ceiling and does nothing about drift. A second index (D) helps only if financial-aid queries are failing, which the scenario never says. The workflow choice (B) fits bounded question answering; an agent adds autonomy nobody asked for.
A semiconductor-equipment vendor is scoping a field-service assistant that parses fault reports, searches service manuals, checks spare-parts inventory, and drafts return authorizations. The sponsor asks for "a team of agents." The architect wants evidence before choosing orchestrator-worker over one scoped agent. Which TWO conditions, if true, would justify the multi-agent design? (Select TWO.)
Decomposition is justified when the work genuinely splits: independent subtasks that can fan out and merge (B), or subtasks whose tools and context are large enough that one agent would suffer capability bloat (so many tools and so much context that tool selection degrades) (D). Strict sequential dependency (C) is a workflow signal; splitting a dependent chain across workers adds handoffs and failure points. An accuracy target (A) is met through model choice, prompting, and evaluation; decomposition alone does not move it.
Full CCAR-P set: 6 timed exams, every answer explained, $19.99 one-time.
See the practice testsDomain 2: Claude Models, Prompting & Context Engineering (13%)
Questions 4–5
An online-education platform grades about three million short-answer quiz responses per day against instructor rubrics. Nine in ten responses are clearly right or wrong; the rest need careful reasoning about partial credit. Students see grades within two seconds. The current design sends everything to the top capability tier and is far over budget. What should the architect propose?
Model routing (a fast, cheap tier handles the well-bounded majority and a deterministic rule escalates the hard slice to a capable tier) cuts cost without a quality cliff, and both paths are synchronous, so the two-second budget holds. Option A relies on self-reported confidence, which is poorly calibrated, and offers no capable path for partial credit. The Batch API (B) processes requests asynchronously at a discount for work that can wait; a student watching the screen rules it out. Chain-of-thought (asking the model to reason step by step before answering) on everything (D) adds tokens and latency to the ninety percent that needed neither.
An airline crew-scheduling assistant sends a 12,000-token prefix of union rules and tool definitions with every request, and prompt caching is enabled. Cache reads stay near zero. A trace shows the system prompt opens with the current timestamp, and the tool list is assembled in a different order on each call. What is the correct fix?
Prompt caching (storing the processed form of a stable prompt prefix so repeated calls skip reprocessing it and pay a discounted rate) works only when the prefix is identical from the first token. A leading timestamp changes it on every call, and reordered tool definitions break it before the rules begin; restructure stable-first, variable-last. Summarizing (C) trims tokens while the cache still misses every call, and it risks losing rule detail. The Batch API (A) is asynchronous and wrong for a live assistant. A larger window (D) changes nothing about repeated processing.
Full CCAR-P set: 6 timed exams, every answer explained, $19.99 one-time.
See the practice testsWatch requests 2 and 3 reuse the cached prefix while only the tail is processed fresh; the airline assistant in Question 5 never gets that reuse because its timestamp sits at the front of the prompt and its tool list reorders on every call.
Study tip: cache-breaking mistakes and batch-versus-realtime calls recur across Domains 2 and 4; the CCAR-P cost and latency optimization guide covers every lever with worked scenarios.
Quick check
A restaurant chain classifies food-safety reports into five fixed labels. A zero-shot prompt (instructions only, no examples) handles clear cases well but confuses "near miss" with "incident" whenever a report describes a hazard that was caught in time. Latency and cost per report are already tight. What is the lowest-effort change most likely to fix the confusion?
The failure is a boundary the model cannot infer from instructions alone, which is what few-shot prompting (a small set of worked examples that show the pattern) fixes, especially when one example demonstrates the confusable case; a testable criterion ("a hazard caught before exposure is a near miss") settles it at almost no token cost. Chain-of-thought (C) adds tokens and latency to every report, which the budget cannot absorb. Retrieval (B) adds a pipeline for a five-label task. A tier upgrade (D) is the expensive lever and still leaves the boundary undefined.
Full CCAR-P set: 6 timed exams, every answer explained, $19.99 one-time.
See the practice testsDomain 3: Integration (19%)
Questions 7–10
A container-port operator runs a berthing-schedule service that three Claude surfaces need: engineers in Claude Code, an internal desktop assistant for planners, and an API-based operations agent. Each team has started writing its own connector, and the service owner is fielding three different bug reports. Which integration mechanism should the architect standardize on?
Model Context Protocol (MCP) is an open standard for exposing tools and data sources to any compliant client, so one server built by the service owner serves Claude Code, the desktop assistant, and the API agent alike: build once, expose everywhere. Three direct integrations (D) is the current state and the source of the triplicate bugs. Agent-to-agent (A) fits when the far side needs its own reasoning; a schedule lookup needs none and gains a second failure domain. Embedding the schedule in prompts (C) is monolithic context that goes stale within the day.
A veterinary-clinic chain deploys an assistant that reads animal-patient records and billing across 40 clinics. Every tool call runs under one service account with read and write access to all clinics. A security review notes that a front-desk user in one clinic could, through the assistant, retrieve or edit another clinic's records. What is the correct remediation?
The gap is authorization: the assistant exercises permissions its user does not hold (a confused-deputy problem), so every tool call must carry the caller's identity and the backend must enforce least privilege (only the access the job requires) with an attributable audit trail. A prompt instruction (B) is probabilistic; a persuasive user defeats it, so it cannot enforce a security control. Rotation and a secrets manager (C) are sound hygiene that leave the over-broad shared permission intact. Encryption and read-only (A) address other threats and still let any user read any clinic.
An architecture firm's building-code assistant retrieves from long, hierarchically numbered code documents. Traces show two failure shapes: answers about a clause omit qualifiers that sit in the adjacent subsection, and queries citing an exact section number such as "R302.5.1" often retrieve unrelated text. Which TWO changes should the architect make first? (Select TWO.)
Retrieval-augmented generation (RAG) injects retrieved chunks into the prompt, so chunking (how documents are split) must match the granularity of the questions; clause-level questions over hierarchical documents need structure-aware chunks that carry headings, parent context, and overlap, so adjacent qualifiers arrive with the clause (A). Exact identifiers are where pure semantic search is weakest, and hybrid retrieval (keyword plus embedding similarity) is the standard fix (C). Thirty chunks (D) is context stuffing: more noise and cost, nothing for the identifier failure. A stronger model (B) cannot infer text that retrieval never delivered.
A cybersecurity vendor's internal operations agent has registered 350 tools across 40 services. Loading every schema consumes a third of the context window before work begins, tool-selection accuracy has fallen, and any single task uses at most five tools. New services arrive monthly. Which context strategy should the architect adopt?
Progressive discovery (loading tool definitions, schemas, and reference data on demand as the agent narrows in on a task) is the strategy that scales: each request carries only the tools it needs, selection accuracy recovers, and monthly additions stop taxing every call. Better descriptions (A) help at the margin while the token tax and attention dilution of 350 loaded schemas remain. A larger window (B) makes the monolithic approach fit and still pays for 345 irrelevant tools per request. Forty user-selected agents (D) pushes routing onto the user and fragments cross-service tasks.
Full CCAR-P set: 6 timed exams, every answer explained, $19.99 one-time.
See the practice testsStudy tip: Integration is the heaviest domain at 19%. If mechanism selection, capability bloat, or RAG chunking felt uncertain, the CCAR-P exam domains breakdown covers each Integration topic with the trade-off being tested.
Build the MCP server behind Questions 7 to 10
You design scoped tools with descriptions the model can route on and connect Claude as a client to an MCP server you built, so the mechanism choice in Question 7, the authorization gap in Question 8, and the tool bloat in Question 10 are problems you have handled in code.
- Design Tools and an MCP Server for a Claude Agent ~3hintermediateRubric-graded ProOpen project
- Build a Full MCP Server and Connect Claude as a Client ~3hintermediateRubric-graded ProOpen project
Domain 4: Evaluation, Testing & Optimization (16%)
Questions 11–13
An energy-trading desk's assistant explains settlement adjustments on a mid-tier model. Traces confirm the correct contract clauses are retrieved every time, and the prompt states its criteria explicitly. Single-step adjustments are answered correctly; anything requiring three or more dependent calculations fails. Which diagnosis and first fix should the architect record?
The symptom pattern is the tell: retrieval is verified good, the prompt is already explicit, and failure concentrates on complex multi-step reasoning. That is model mismatch (a task harder than the selected tier reliably handles), and the disciplined first move is the identical prompt on a more capable tier for that slice, then a decision on routing only the hard cases. Re-ranking (A) improves retrieval precision, which is already fine. A prompt rewrite (D) fits failures that are uniform across inputs. Citations (C) address fabricated content; here the model has the right text and fails to reason over it.
An HR-technology vendor is designing evaluation for a job-description generator before launch. Every output must be valid JSON with six required fields, must avoid biased or exclusionary language, and should read naturally to recruiters. Releases will ship weekly and volume is high. Which evaluation design fits these requirements?
Mixed methodology assigns each requirement to the cheapest grader that measures it reliably: schema and required fields are objective, so code checks catch them deterministically; language quality and bias are qualitative, so an LLM-as-judge (a separate model call scoring output against a written rubric) scales to weekly releases; a periodic human sample calibrates the judge and reviews its bias findings. Full human review (C) cannot keep pace with volume. A judge grading schema validity (B) applies a probabilistic grader to a deterministic property. Exact match (A) fails valid descriptions that differ in wording, which is nearly all of them.
A legal publisher's summarization pipeline runs at two and a half times its inference budget. Seventy percent of traffic is overnight bulk work with no one waiting; thirty percent is interactive with healthy margin on p95 latency (the time within which 95 percent of requests complete). Every request carries a 9,000-token stable instruction prefix. Which TWO levers should the architect pull first? (Select TWO.)
The Message Batches API processes asynchronous requests at a discounted rate for work that can wait, which describes the seventy percent exactly (D). Prompt caching turns the 9,000-token prefix from a per-request cost into a cheap cache read across both traffic classes (A). Together they attack the two largest cost drivers without touching quality. A verification pass (B) adds a model call to every request; it may raise quality and it certainly raises spend. Upgrading everything (C) multiplies cost on the assumption that retries drive the overrun, which the scenario never says.
Full CCAR-P set: 6 timed exams, every answer explained, $19.99 one-time.
See the practice testsTap each grader to see its cost against its fidelity; Question 12 is this chart applied to one product, with code checks on the schema, an LLM judge on language quality, and a human sample that calibrates the judge.
Which CCAR-P domains are you weakest in?
Answer 20 exam-style questions and see your score against the pass mark in each domain, with a plan for what to study first. Free, no sign-up, about 38 minutes.
Ready for full exams? See the CCAR-P practice tests and course
Domain 5: Governance, Safety & Risk Management (14%)
Questions 14–15
An automotive dealership group's assistant books service appointments at high volume, and it can also issue goodwill credits and send binding price quotes to customers. Finance wants control over the money-moving actions without destroying the efficiency case for routine bookings. Where should the architect place human-in-the-loop review (a human decision point inside the automated flow)?
Human-in-the-loop placement is risk-based: a human belongs before irreversible or financially binding actions (credits, quotes), and high-volume low-risk actions (bookings) get sampled review so the efficiency case survives. Approving everything (B) is the all-or-nothing trap; it burns reviewer time on routine bookings and slows the throughput the system was bought for. A prompt rule with no gate (D) treats probabilistic instructions as a financial control, and a talked-into credit still goes out. Weekly review (A) discovers a bad quote after the customer holds it; binding actions need the control before execution.
A telehealth startup will use Claude to draft visit summaries for patients in Germany and in the United States. Summaries contain diagnoses and medication history, and the service runs in a single US cloud region. Legal asks the architect which regulatory regimes shape the architecture and what controls follow. What should the architect answer?
Match the framework to data type and jurisdiction, and expect overlaps. US patients' diagnoses are protected health information (PHI) under HIPAA (the US health-data privacy law), which requires a business associate agreement (BAA, the contract binding a vendor that handles PHI) plus safeguards such as audit logging. German patients bring GDPR (the EU data-protection regulation) with data minimization, erasure rights, and a residency question the single US region must answer. HIPAA alone (B) ignores the EU patients; GDPR alone (D) ignores US PHI. FedRAMP (C) governs US federal cloud authorization and has no bearing on a private telehealth company.
Full CCAR-P set: 6 timed exams, every answer explained, $19.99 one-time.
See the practice testsStudy tip: guardrail layers, human-in-the-loop placement, and the GDPR, HIPAA, and FedRAMP triggers are covered end to end in the CCAR-P governance, guardrails, and compliance guide.
Quick check
A property-management assistant summarizes incoming tenant emails and can file maintenance orders and pay vendors through tools. A red-team exercise shows that a crafted email can instruct the assistant to issue a vendor payment. Leadership wants the exposure closed before launch. Which TWO controls should the architect implement? (Select TWO.)
This is prompt injection (malicious instructions smuggled through content the model reads, here tenant emails), and durable defenses are architectural. Isolating the component that reads untrusted content from the component holding payment tools (A) removes the path from attacker text to money. Deterministic tool permissioning plus an approval gate on payments (C) means even a manipulated model cannot complete the action. A prompt instruction (D) is a probabilistic layer worth having and useless as the sole control; the red team just talked around it. A stronger model (B) resists more attempts and still leaves an attacker unlimited tries against a system with no hard stop.
Full CCAR-P set: 6 timed exams, every answer explained, $19.99 one-time.
See the practice testsDomain 6: Stakeholder Communication & Lifecycle Management (14%)
Questions 17–19
A regional airline's operations executive opens an engagement by saying the ops center "needs an AI copilot like the one at the conference," confirms the budget is approved, and asks for a proof of concept in three weeks. No success criteria, data inventory, or constraints have been discussed. What should the architect do first?
Discovery precedes design in the lifecycle (discovery, design, handoff, monitoring, iteration), and its outputs are exactly what the sponsor skipped: a problem statement with measurable success criteria, a constraint inventory (latency, cost, compliance), a data and access map, and a stakeholder map with decision owners. Building first (A) is the skip-discovery trap; a demo answers a question nobody has defined and anchors expectations in the wrong place. Model and integration choices (B) depend on constraints still unknown. Copying the conference architecture (D) imports another company's problem framing wholesale.
A mining company's procurement director tells the architect that the new contract-summarization assistant must "never miss a clause" and "answer within a second," and asks for both in the statement of work. The corpus is large and clause language varies by supplier. Which TWO responses reflect professional expectation-setting? (Select TWO.)
LLM systems are probabilistic, so the professional move converts absolutes into targets the architecture can meet and monitoring can enforce: an accuracy floor measured on a representative evaluation set, an explicit error budget, and a defined path when a miss occurs (B), plus latency stated as percentiles validated during design (A). Accepting the guarantees (C) commits to perfection, and a verification pass adds latency to a one-second ask. Promising the fastest tier (D) trades away the accuracy the director cares most about and still cannot guarantee a wall-clock number under load.
A crop-advisory assistant has been live for a year. The original architect has left, and a new technical lead proposes replacing retrieval with a fine-tuned model, arguing that nobody can explain why retrieval was chosen. Meeting notes and chat threads exist but settle nothing. What should the architect institute so this does not recur?
An architecture decision record (ADR) is a short, version-controlled document capturing one consequential decision: the context at the time, the options considered, the choice made, and its consequences. It answers exactly the question the new lead is asking and preserves the reasoning through staff turnover. A master design document (C) describes what exists and rarely why an alternative lost, and it goes stale as one monolith. Code comments (D) explain mechanics at the line level, the wrong altitude for a pattern decision. Archived minutes (B) are the notes the team already has: unstructured and mixed with everything else.
Full CCAR-P set: 6 timed exams, every answer explained, $19.99 one-time.
See the practice testsDomain 7: Developer Productivity & Operational Enablement (7%)
Quick check
A video-streaming company is rolling Claude Code out to 120 engineers. Repository conventions, build and test commands, and a shared MCP server for the internal metrics service currently live in one senior engineer's personal configuration, and nobody else's sessions behave the same way. Security also forbids credentials in the repository. What should the architect direct?
Team-wide behavior belongs in version-controlled, project-level configuration (a project CLAUDE.md file for conventions and commands, plus project-scoped MCP server definitions) so every engineer's session inherits it on checkout, with secrets referenced through environment variables so nothing sensitive is committed. Copying personal settings (C) propagates a snapshot that drifts immediately and never updates. Committing credentials (D) buys convenience by violating the security requirement outright. A README (A) documents conventions for humans and does nothing to configure the tool, so sessions still diverge.
Full CCAR-P set: 6 timed exams, every answer explained, $19.99 one-time.
See the practice testsStudy tip: the traps in this set (flagship-everywhere, agentic-by-default, prompt-guarantee, context-stuffing, skip-discovery) are catalogued with fixes in Common CCAR-P exam mistakes to avoid.
If you missed these, read this
Tally your misses by domain, then go straight to the guide that covers the judgment you dropped.
- Domain 1, Questions 1 to 3 (pattern selection, feedback loops, decomposition): Multi-agent orchestration patterns for CCAR-P.
- Domain 2, Questions 4 to 6 (routing, caching, prompting technique): Cost and latency optimization: caching, batch, routing and the Domain 2 rules in the CCAR-P cheat sheet.
- Domain 3, Questions 7 to 10 (MCP, authorization, RAG, progressive discovery): the Integration section of the exam domains breakdown and Deploying Claude in the enterprise for identity and network design.
- Domain 4, Questions 11 to 13 (diagnosis, mixed-methodology evaluation, cost levers): CCAR-P evaluation strategy: eval datasets, metrics, and diagnosis.
- Domain 5, Questions 14 to 16 (HITL, compliance regimes, injection defenses): Governance, guardrails, and compliance.
- Domain 6, Questions 17 to 19 (discovery, SLA framing, ADRs): Domain 6 of the complete guide and the trap list in How to pass CCAR-P on your first attempt.
- Domain 7, Question 20 (team configuration): the Domain 7 section of the cheat sheet.
Rebuild the Domain 1 and 4 judgment as graded projects
The workflow-patterns task rebuilds the Domain 1 choice between an augmented call, a workflow, and an agentic loop on one job, and the decision-prompt task has you prove a prompt change on an eval set, which is the Domain 4 judgment behind Questions 11 to 13.
- Build the Canonical Claude Agent Workflow Patterns ~4hintermediateRubric-graded ProOpen project
- Engineer a Claude Decision Prompt and Prove It with Eval ~3hintermediateRubric-graded ProOpen project
Twenty questions is a diagnostic; the exam presents 63 in 120 minutes, and pacing under a clock is a separate skill from judgment. Preporato's CCAR-P practice tests provide six full-length exams on the same 7-domain blueprint with an explanation for every option, and the free sample questions is the zero-cost way to see the format. Both are included in Preporato Pro; see pricing for the Pro plan and the single-cert bundle.
Key Takeaways
0/12 completedNext steps
Score the set, note the domains where you dropped points, and read the linked guides for those domains before your next attempt. Then move to timed practice: Preporato's CCAR-P practice tests mirror the 63-question, 120-minute format, and the CCAR-P exam format guide explains pacing, scaled scoring, and test-day logistics. If you are still deciding between the two Claude architect tiers, CCA-F vs CCAR-P settles that first.
Sources:
- Anthropic Partner Academy: Claude Certified Architect - Professional certification
- Claude Developer Platform docs: Prompt caching
- Claude Developer Platform docs: Batch processing
- Claude Code docs: Settings and configuration scopes
- Model Context Protocol specification and overview
- U.S. Department of Health and Human Services: HIPAA
What to do next
Not sure where you stand yet?
The free CCAR-P readiness check takes about 38 minutes and gives you a study plan.
![CCAR-P Practice Questions with Explanations: 20 Scenario Items [2026]](/blog/ccar-p-practice-questions-with-explanations-2026.webp)
