Free Claude Certified Architect - Professional (CCAR-P) Practice Questions

Sample CCAR-P Practice Questions

Browse all 20 free Claude Certified Architect - Professional practice questions below.

1. A retail bank is automating new-account onboarding triage. Every application moves through the same three stages: extract fields from the ID document, validate them against the application form, and flag discrepancies for a human reviewer. Compliance requires that each stage's output be independently inspectable, and the stage sequence has not changed in four years. Which architecture best fits these constraints?

   • A prompt-chained workflow with one model call per stage and programmatic validation gates between the stages.
   • An orchestrator agent that decomposes each application into subtasks and delegates them to worker agents at run time.
   • A single model call with a comprehensive prompt covering extraction, validation, and discrepancy flagging together.
   • A routing step that classifies each application and dispatches it to one of several specialized handler prompts.

2. An e-commerce support agent has grown to more than 40 tools covering orders, refunds, inventory, and shipping. Tool-selection accuracy has fallen, input costs have risen, and prompt cache hit rates drop after each tool addition. The team must keep all capabilities available. Which TWO changes most directly address the degradation? (Select TWO)

   • Consolidate overlapping tools and rewrite each description to state exactly when it should be invoked.
   • Set tool_choice to any so the agent always commits to some tool instead of answering from stale context.
   • Move the full tool list into the system prompt so it sits in a stable position for caching.
   • Raise max_tokens on every request so the model has more room to reason before committing to a tool call.
   • Adopt the tool search tool and mark rarely used tools defer_loading so schemas load only when relevant.

3. A logistics platform classifies 30 million short carrier status messages per day into 12 fixed delivery states. Accuracy on a labeled sample is already 97% with a small model, the p95 latency budget is 500 ms, and finance has capped monthly model spend. Which model choice fits these constraints?

   • Claude Fable 5, because always-on deep reasoning maximizes per-message accuracy.
   • Claude Haiku 4.5, validated against the labeled sample before rollout.
   • Claude Sonnet 5, since its stronger reasoning adds accuracy headroom.
   • Claude Opus 4.8, to standardize on one model across all company workloads.

4. A hospital network's Claude feature summarizes patient intake forms for clinicians. Leadership wants to expand it to three more clinics but first needs evidence that the summaries are reliable. Intake forms vary widely in quality and specialty across sites. What is the strongest form of evidence to gather before expansion?

   • Scores from a held-out set of real de-identified intake forms spanning specialties, graded by clinicians against a written rubric.
   • The model's published scores on standard medical benchmarks and general reasoning suites, compared against the tiers the team considered.
   • Average clinician star ratings collected from the pilot clinic's daily use of the feature over the full quarter since launch.
   • A staged rollout to one additional clinic with close monitoring of complaints and correction requests during the first month.

5. A financial-services chatbot built on Claude occasionally echoes a customer's full account number back in replies, which the bank's data-handling policy forbids in any channel. The team needs the strongest control to stop account numbers from reaching customers. Where should that control live?

   • In a deterministic output filter that scans and masks account-number patterns before any reply is displayed.
   • In the retrieval layer, by excluding account numbers from every document indexed for the assistant's knowledge base.
   • In a fine-tuned model variant trained on historical transcripts in which account numbers were consistently masked.
   • In the system prompt, as an explicit rule that account numbers must be masked or truncated in every customer-facing reply.

6. A freight forwarder processes 40,000 customs declarations every night. For roughly 92 percent of documents, the same nine fields extract cleanly in a single structured-output call; the remainder are degraded scans that need a tariff-code lookup and a second extraction pass. Finance has capped per-document cost, and operations requires predictable nightly throughput. Which architecture best fits these constraints?

   • Route every declaration through an autonomous agent loop with lookup tools so clean and degraded documents follow one adaptive path
   • Run a fixed extraction pipeline with a confidence gate that routes only low-confidence documents into a lookup-and-reextract branch
   • Use an orchestrator agent that spawns a specialized extraction sub-agent per field so all nine fields are pulled independently in parallel
   • Process each document in one structured-output call and iterate on the prompt until degraded scans reach parity with clean-scan accuracy

7. A biotech startup's internal research assistant needs to call one endpoint of the company's lab information system to fetch assay results. No other AI client exists or is planned this quarter, the API is stable and versioned, and the two-engineer team must ship in three weeks. Which integration mechanism matches this scope and timeline?

   • Deploy a standalone MCP server wrapping the lab system so the capability becomes discoverable by future Claude clients across the company
   • Front the lab system with its own agent and have the assistant delegate assay queries to it over an agent-to-agent interface
   • Build a thin internal gateway service that translates the assistant's natural-language requests into structured lab-system API calls
   • Define a single tool against the existing REST endpoint and invoke it through Messages API tool use from the assistant

8. An insurer classifies incoming claims with a prompt carrying about 6,000 tokens of stable guidelines, taxonomy, and worked examples, followed by a short variable claim description. Volume is roughly 900,000 classifications per week, results feed a nightly job rather than a live user, and accuracy must not regress. Which TWO changes cut cost the most while preserving accuracy? (Select TWO)

   • Reorder the prompt so the variable claim description comes first and the stable guidelines follow at the end of each request
   • Split the taxonomy across several smaller sequential calls so each individual request carries a lighter context load
   • Mark the stable guideline-and-example block with a cache breakpoint so repeated requests read it at the reduced cached-token rate
   • Submit the classifications through the Message Batches API, which halves token pricing for asynchronous workloads
   • Move the workload to a larger model tier so the worked examples can be dropped from the prompt while holding accuracy

9. A home-goods retailer has written a new prompt version for its product-description generator. The team holds 400 expert-labeled examples with a scoring rubric, and leadership wants a go or no-go decision this week, before any customer traffic sees the new version. What is the soundest way to decide between the two prompts?

   • Have the generating model review each pair of outputs and record which version it judges stronger for every item
   • Release the new version behind a feature flag to full traffic and compare this week's engagement metrics with last week's
   • Launch a four-week production A/B test and use the labeled set only as a tie-breaker if the test proves inconclusive
   • Score both prompt versions against the held-out labeled set using the fixed rubric and compare the results side by side

10. A wealth-management firm uses Claude to draft investment-suitability memos that advisers act on directly with clients. Compliance is worried about a single failure mode: a fabricated figure in a memo driving a client decision. Which control most directly reduces that risk?

    • Require the model to attach a source citation beside every figure it includes in each generated memo
    • Require adviser sign-off checking each figure against the cited source records before client use
    • Set temperature to zero so each memo is generated deterministically from the same underlying inputs
    • Add a standing notice to memos stating that figures should be independently verified before reliance

11. A parcel logistics company receives roughly 4,000 delivery-exception events per hour and must sort each one into a fixed set of nine categories so application code can route it to the right resolution queue. Every decision must be reproducible for audit, per-event latency must stay under three seconds, and the per-event cost ceiling is strict. Which architecture best fits these constraints?

    • A three-step prompt chain that first summarizes the event, then enriches it with account history, then assigns a category
    • An autonomous agent loop that inspects each incoming event, selects investigation tools at runtime, and decides the routing path on its own
    • One structured classification call per event against the fixed taxonomy, with routing decisions executed by application code
    • An orchestrator model that fans each event out to specialized worker models and merges their category votes

12. A fintech support platform classifies about two million short customer messages per day into six intent labels. Pilot testing showed all current Claude models meet the quality bar for this task. The team needs sub-second responses and the lowest workable per-message cost. Which model choice best fits?

    • Claude Opus 4.8, since the flagship tier gives the most headroom if message complexity grows later
    • Claude Opus 4.8 through the Batch API, since the 50 percent discount offsets the higher tier pricing
    • Claude Haiku 4.5, since it is the fastest tier and its pricing suits very high request volumes
    • Claude Sonnet 5, since the mid tier is the standard default for production workloads of any size

13. A legal-tech firm is building retrieval over 80,000 commercial contracts. Users ask clause-level questions and answers must quote the exact clause language. The current prototype embeds each contract as one vector and retrieval quality is poor: answers cite the right contract but the wrong clause. Which TWO changes would most directly improve clause-level retrieval? (Select TWO)

    • Re-chunk documents along clause and section boundaries with modest overlap, storing section metadata with each chunk
    • Drop retrieval and load each full contract into the 1M-token context window at question time
    • Combine semantic search with keyword matching and apply a reranking step over the top candidate chunks
    • Fine-tune a model on the contract corpus so it can answer clause questions from learned weights
    • Raise the retriever's top-k from 5 to 100 so the correct clause is more likely to be somewhere in context

14. An insurance carrier is six weeks from launching a Claude-based claims-intake assistant. Leadership asks the team to demonstrate the system is ready. Nothing has been measured yet beyond ad-hoc demos that looked convincing. What should the team establish first to ground the launch decision?

    • A structured user-satisfaction survey program scheduled to begin with the first production cohort
    • A load test demonstrating the deployment holds its latency and throughput targets at twice the projected launch traffic volume
    • Task-level success criteria plus a graded evaluation set built from representative real claims interactions
    • A comparison of the chosen model against alternatives on published general reasoning benchmarks

15. A social gaming platform uses Claude to moderate chat and can mute players automatically. Volume is around 50,000 flags per day, the trust-and-safety team has six reviewers, a regulator requires an appeal path, and false mutes are a known churn driver. How should automated action be governed?

    • Automate all mutes at a single tuned global threshold and audit a weekly sample of muting decisions for quality drift
    • Auto-mute only above a validated confidence threshold, route borderline flags to reviewers, and log rationale for appeals
    • Require a trust-and-safety reviewer to confirm every proposed mute before it takes effect, keeping the model in an advisory role
    • Tighten the moderation prompt so fewer flags fire, and keep automatic muting on for whatever still gets flagged

16. A freight forwarder is automating customs declarations. Every shipment follows the same pipeline: extract commodity codes from invoices, validate them against the current tariff schedule, compute duties, and queue the filing for a licensed broker's sign-off. Regulators require a reviewable record of each intermediate result, and the per-shipment processing budget is tight. Which architecture fits these constraints best?

    • An autonomous agent that plans its own path per shipment, choosing which extraction and validation actions to run based on what it discovers in the documents
    • A code-orchestrated chain of focused prompts, one per pipeline step, with each intermediate output validated and persisted before the next step runs
    • A single comprehensive prompt that performs extraction, tariff validation, and duty computation in one model call to minimize orchestration code
    • A routing layer that classifies each shipment on arrival and dispatches it to one of several differently tuned prompt variants

17. A medical-device maker summarizes 20,000 field incident reports a month with Claude. Each summary must preserve four mandatory fields (device model, failure mode, patient impact, action taken) and must not introduce claims absent from the source report. The team can fund expert human labeling for only a few hundred examples. Which evaluation design fits these constraints?

    • Score each summary's word-overlap (ROUGE) against its source report, tune the threshold on the labeled examples, and automatically fail any summary that falls below it
    • Have clinical reviewers examine every summary before release, since regulated safety content warrants expert human judgment on each individual output
    • Use an LLM judge with a detailed grading rubric as the single quality gate, since one grader can score field presence and faithfulness together in a single pass
    • Enforce the four mandatory fields with deterministic code checks, and score faithfulness with an LLM judge whose rubric is calibrated against the human-labeled sample

18. An auto-parts marketplace is building retrieval over 800,000 listings. Each listing pairs structured fields (part number, price, compatible vehicle years) with long free-text condition and fitment notes. Buyers query by exact part number about as often as by descriptive phrase, and numeric constraints must be honored exactly. Which TWO retrieval design choices best serve this mixed workload? (Select TWO)

    • Embed each complete listing as a single vector so the structured fields and the free-text notes always travel together through retrieval and into the context
    • Fold the structured fields into the description text before embedding so one similarity search can satisfy price and year constraints
    • Combine lexical keyword matching with semantic search so exact part-number queries and descriptive-phrase queries both retrieve reliably
    • Raise top-k from 10 to 100 and pass every candidate into the model's context so ranking weaknesses are absorbed by the model's own reading of the set
    • Index the structured fields as filterable metadata so part-number, price, and year constraints are applied exactly rather than approximated by embeddings

19. A travel booking company classifies two million inbound chat messages a month into eight intent categories. The p95 latency target is under one second, finance has capped monthly model spend, and most messages are short and unambiguous. An internal labeled test set exists. Which model strategy fits this profile?

    • Standardize on the most capable Opus-tier model so classification quality is settled from day one, absorbing the higher per-message cost and latency
    • Start with Claude Haiku, the fastest and lowest-cost tier, validate accuracy on the labeled test set, and escalate tiers only if it misses the bar
    • Send each message to both a small and a large model and accept the label where they agree, re-running disagreements with a tie-breaking prompt
    • Fine-tune a mid-tier model on the labeled intents before measuring what a prompted base model already achieves on the existing test set

20. A security operations center runs a Claude agent that triages endpoint alerts and can isolate machines from the network. Contract SLAs require containment of commodity malware on employee workstations within 60 seconds around the clock. Leadership's specific worry is a false positive isolating a revenue-critical production server, and the analyst team is small. Which control design resolves this tension?

    • Restrict the agent to read-only enrichment and recommendation duties, with analysts executing every containment step manually from its written findings
    • Auto-execute isolation only when the model's stated confidence clears a high global threshold, holding lower-confidence verdicts for review
    • Tier the permission policy by asset class: auto-execute isolation on workstations to meet the SLA, and gate any server tagged critical behind analyst approval
    • Require an analyst to approve every isolation action across the entire fleet, workstation or server alike, so no machine can be disconnected on a bad classification