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 bank wants to automate its new-account onboarding triage. Each application follows a fixed sequence of steps: extract fields from the ID document, validate them against the application form, and flag any mismatches for a human reviewer. The steps are the same every time and the output is a structured decision. Which architecture pattern best fits this workload?

   • A fully autonomous agent that decides at runtime which step to run next and when to stop
   • A prompt-chaining workflow where each step is a discrete Claude call with a defined input and output, wired together in a fixed order
   • A single monolithic prompt that asks Claude to perform the extraction, the validation, and the flagging all together in one call with no intermediate structure between the stages
   • A multi-agent swarm where separate agents negotiate the outcome of each application

2. An e-commerce support team keeps adding tools to a Claude agent so it can look up orders, issue refunds, and check inventory. Over time the agent has accumulated more than 40 tools and its accuracy at choosing the right one has dropped. Which TWO changes most directly reduce this capability-bloat problem? (Select TWO)

   • Increase max_tokens so the model has more room to consider every tool
   • Add every tool description into the system prompt as well as the tools array so the model sees each one twice
   • Raise the temperature so the model explores more tool options
   • Remove or consolidate rarely used and overlapping tools so the agent sees a smaller, well-scoped tool set
   • Group tools behind a progressive-discovery layer so only the tools relevant to the current task are exposed to the model

3. A logistics company needs to classify millions of short carrier status messages per day into a fixed set of delivery states. The task is simple and well defined, latency per message must be low, and cost per message is the dominant constraint. Which Claude model choice best fits?

   • Select the largest, most capable Claude model to maximize accuracy on every message, even though the overwhelming majority of these classification messages are completely trivial to handle
   • Select a smaller, faster Claude model (such as a Haiku-class model) because the task is simple and high-volume, and reserve larger models for genuinely hard cases
   • Use a different model chosen at random per message to average out any errors
   • Route every message through two large models and compare their outputs before deciding

4. A healthcare provider deployed a Claude feature that summarizes patient intake forms for clinicians. Before expanding it, the team wants to know whether the summaries are accurate and complete. What is the most reliable first step to evaluate quality?

   • Build a labeled evaluation set of representative intake forms with clinician-reviewed reference summaries and defined scoring criteria, then measure the system against it
   • Ship the feature to all clinicians immediately and rely on them to notice and report any problems during their normal clinical workflow whenever they happen to catch one, instead of testing it beforehand
   • Ask Claude to rate its own summaries as good or bad and treat that self-assessment as the quality score
   • Track only the average response latency and treat fast responses as correct

5. A financial-services chatbot built on Claude occasionally echoes a customer's full account number back in its replies. Which control most directly prevents sensitive identifiers from appearing in model output?

   • Increase the model temperature so responses are less repetitive
   • Add a line to the system prompt saying 'be helpful and professional'
   • Add a post-processing redaction and output-filter step that detects and masks account numbers and other PII before the response reaches the user
   • Log every model response to a central monitoring system so that any PII leaks can be found and reviewed by an analyst after the fact

6. A legal operations team wants Claude to pull the same 12 fixed fields (party names, effective date, governing law, term length) from every uploaded NDA. The field set never changes and each document is processed exactly once. Which architectural pattern is the best fit?

   • Build a multi-agent system where a planner agent inspects each contract and dynamically spawns specialized extraction sub-agents for whichever fields it decides are present
   • Use a single prompted extraction step (an augmented LLM call with structured tool_use output) applied to each document in a fixed pipeline
   • Give Claude an autonomous agentic loop with tools and let it decide how many extraction passes each contract needs before stopping
   • Fine-tune a custom model on labeled contracts so the 12 fields are predicted from the weights without any prompt

7. A platform team needs one internal Claude assistant to call a single internal deployment service that already exposes a stable, well-documented REST API used only by this one client. They ask for the lowest-effort way to let Claude invoke it. Which integration mechanism fits best?

   • Stand up a dedicated MCP server that wraps the deployment API so that any future Claude client across the company could later discover and reuse it
   • Build an agent-to-agent protocol so the deployment service is exposed as its own autonomous agent that the assistant negotiates with
   • Define a tool that calls the existing REST endpoint directly through Messages API tool_use, since one client needs one integration
   • Paste the full OpenAPI spec into the system prompt and have Claude describe each call so a human can run it manually

8. An insurance triage prompt sends a large, stable prefix (classification guidelines, taxonomy, and few-shot examples) plus a short variable claim description on every request, at very high volume. Results do not have to be returned synchronously. Which TWO changes reduce cost and latency without hurting accuracy? (Select TWO)

   • Enable prompt caching on the stable guideline-and-example prefix so those repeated tokens are not reprocessed on every call
   • Raise the sampling temperature so Claude generates fewer output tokens per classification
   • Move the variable claim text to the very front of the prompt, ahead of the stable guidelines
   • Submit the high-volume classifications through the Batch API since results are not needed synchronously
   • Delete the few-shot examples entirely to shrink the prompt, accepting the lower accuracy

9. A retail team has written a new prompt version for generating product descriptions and wants to know whether it is genuinely better than the current one before rolling it out to all catalog items. What is the soundest way to decide?

   • Score both prompt versions on the same held-out, expert-labeled evaluation set using a fixed rubric, then compare the results
   • Deploy the new prompt to 100 percent of traffic and watch aggregate customer engagement over a week, even though many unrelated factors could move those numbers during the same period
   • Ask Claude to read its old and new descriptions and state which set it thinks is higher quality
   • Compare the average output token count of each version and prefer the one that produces longer descriptions

10. A contract-review assistant drafts clause-level risk summaries that attorneys rely on when advising clients. Which control most directly manages the risk that a hallucinated or misstated clause is acted upon?

    • Increase max_tokens so each risk summary can be longer and more complete
    • Set temperature to zero so hallucinations are eliminated from the summaries
    • Add a standard disclaimer stating that the summaries may contain inaccuracies and that readers should independently verify every clause before relying on them
    • Require attorney review of each summary against the cited source clauses before it is used with a client

11. A telecom network operations center receives thousands of raw alarm events per hour and wants Claude to group related alarms and draft a first-response summary for each incident so on-call engineers read fewer logs. For a first deployment that must show value quickly, which approach fits best?

    • Start with a single augmented LLM call that takes a batch of correlated alarms and returns a structured incident summary, then move to agentic patterns only if one call proves insufficient
    • Build a fully autonomous multi-agent system where each agent owns a network domain and negotiates incident ownership before any summary is produced
    • Fine-tune a model on historical alarm logs so incident summaries are generated without providing alarm context at inference time
    • Replace the existing alarm correlation engine entirely and have Claude continuously poll raw device telemetry on a fixed schedule, reconstructing correlation logic from scratch inside the prompt

12. An edtech tutoring app must classify each incoming student message into one of five intents. Volume is very high, each classification is simple, and the product requires low per-message latency and cost. Which model choice fits best?

    • Use Claude Opus 4.8 for every classification request so that the single highest-capability model handles even the simplest routing decision regardless of the volume involved
    • Use Claude Haiku 4.5, the fastest and most cost-effective model, since the task is simple and latency and cost dominate
    • Use Claude Fable 5 with maximum effort to guarantee the intent labels are never wrong
    • Alternate requests between two large models to average out latency spikes across the fleet

13. You are designing a retrieval pipeline for a manufacturing knowledge base of equipment manuals, where technicians ask targeted questions such as the torque spec for a specific bolt. Which TWO design choices best match this data shape and query pattern? (Select TWO)

    • Chunk the manuals along their existing section and table boundaries so a retrieved chunk contains a complete, self-contained procedure or spec
    • Embed each entire manual as a single large vector so every query retrieves the whole document at once
    • Store structured spec tables as discrete records and retrieve them by exact key lookup rather than by semantic similarity alone
    • Increase chunk overlap to about 90 percent so nearly every chunk repeats most of its neighbors
    • Skip indexing entirely and paste all manuals into the system prompt on every request

14. A team launching a Claude-based resume screening assistant is asked what to measure before rollout. Which metric set most directly reflects whether the assistant is fit for production?

    • Only the average number of tokens per request, since token usage is the clearest signal of quality
    • Only the number of resumes processed per hour, because throughput is what stakeholders notice
    • Accuracy against a labeled screening set, plus latency, cost per screen, and a fairness check across demographic groups
    • Only end-user satisfaction surveys collected six months after launch, treated as the single gate for whether the system is working correctly

15. A gaming company deploys Claude to auto-moderate chat and can issue temporary mutes automatically. Which control most directly limits the harm from a wrong automated action?

    • Raise the model temperature so moderation decisions vary more between similar messages, which the team believes will avoid systematic bias
    • Log every moderation decision to an analytics dashboard reviewed monthly
    • Add more few-shot examples of toxic messages to the system prompt
    • Require human review before permanent bans while allowing short automated mutes that expire on their own

16. A state benefits agency is building a Claude application that reads a citizen's uploaded income documents, extracts a fixed set of eligibility fields, validates them against published policy thresholds, and routes the case to a human caseworker. The sequence of steps never varies. Which architecture best fits this problem?

    • Deploy a fully autonomous agent that decides which action to take at each step at runtime, discovering the sequence of operations dynamically through open-ended tool use
    • Build a fixed workflow that chains discrete prompted steps (extract, validate against thresholds, route) with your code controlling the order
    • Send the documents, policy thresholds, and routing rules in one large prompt and ask Claude to return the final decision in a single call
    • Fine-tune a model on past eligibility decisions so the policy logic is encoded in the model weights

17. A pharmaceutical company runs a Claude pipeline that summarizes adverse-event narratives from clinical trial reports. Before launch, the team wants a single evaluation approach that catches both factually wrong summaries and summaries that omit a required safety field. Which evaluation design best meets this need?

    • Track only average output token count per summary as a proxy for completeness
    • Rely on user thumbs-up and thumbs-down collected after launch as the primary quality signal
    • Combine a code-based check that each required safety field is present with an LLM-as-judge or human review of factual accuracy against the source narrative
    • Measure end-to-end latency and cost per summary and set a pass threshold on both, assuming faster and cheaper outputs are more accurate

18. A real-estate platform is building a RAG system so Claude can answer buyer questions from a corpus of property listings, each with structured attributes (price, bedrooms, location) and a long free-text description. Which TWO design choices most improve retrieval quality for this mixed data shape? (Select TWO)

    • Embed each entire listing as one large chunk so no context is ever split
    • Store structured attributes as metadata and filter on them before or alongside semantic search over the descriptions
    • Increase the embedding model's temperature to diversify the vectors it produces
    • Chunk long descriptions into semantically coherent passages sized to the retriever, rather than by a fixed character count that splits mid-sentence
    • Replace retrieval entirely by pasting the full listing corpus into the system prompt on every query and letting the model scan it

19. A travel booking company needs Claude to classify incoming customer chat messages into one of eight intent categories. The task is simple and high-volume, and low latency and low cost matter most. Which model choice is most appropriate as a starting point?

    • Start with a small, fast model such as Claude Haiku and only escalate if accuracy is insufficient
    • Start with the largest, most capable model available to guarantee accuracy, and accept the higher per-message cost and latency as the baseline
    • Fine-tune a mid-size model on labeled intents before measuring whether a base model already suffices
    • Use a different model for every category so each classifier is specialized

20. A security operations center deploys a Claude agent that triages alerts and can call tools to isolate a host from the network. Leadership is worried about the agent isolating a critical production server on a false positive. Which control most directly addresses this risk?

    • Lower the model temperature so the agent's decisions are more deterministic
    • Add more few-shot examples of well-written triage summaries to the system prompt
    • Switch to a larger, more capable model so that the agent classifies alerts more accurately and therefore isolates hosts less often by mistake
    • Require human approval before the isolate-host action executes on any server tagged critical