Timed Mode Set 1 – Claude Certified Architect Professional CCAR-P

1. Question

A proposal team uses one large prompt to analyze a lengthy procurement document, compare its requirements with dozens of governance standards, calculate a pursuit score, and produce an executive recommendation. Despite repeated prompt revisions, the model still overlooks some standards and provides weak justification for its score. What is the most effective approach for improving the reliability and completeness of the output?

Generate a condensed summary of the procurement document before submitting it with the complete policy set to the original assessment prompt.
Add few-shot examples that demonstrate the preferred executive recommendation format.
Divide the workflow into sequential stages for requirement extraction, grouped standards assessment, scoring, and final synthesis, using structured intermediate outputs.
Embed a weighted compliance checklist and scoring rubric in the existing prompt, requiring a brief rationale for every policy.

2. Question

A financial services firm is developing an AI analyst to handle complex investigations where the required tools and sequence of actions cannot be determined before execution. Each intermediate result may reveal the need for additional searches, calculations, or document reviews. Which design pattern is most appropriate for this requirement?

A retrieval pipeline that gathers information from a fixed collection of approved sources before generating a response.
A routing system that assigns each request to one of several predetermined processing pipelines.
A parallel orchestration pattern that launches several specialized analysis tasks simultaneously and combines their outputs.
An autonomous agent that repeatedly plans an action, uses the appropriate tool, evaluates the result, and determines the next step.

3. Question

A manufacturing company uses specialized AI agents to inspect product specifications, verify component availability, evaluate safety requirements, and approve production orders. Every decision must be recorded in sequence, and downstream agents must not execute when an earlier agent detects a compliance issue. Which orchestration method most effectively fulfills the stated requirements?

Use a central orchestrator to invoke each agent sequentially, record every outcome, and stop the workflow when an exception occurs.
Use rule-based handoff orchestration in which each specialized AI agent transfers the workflow to the next agent and appends its result to a shared audit log.
Use state-machine orchestration that advances through agent stages based on completion signals and evaluates exceptions during a final validation step.
Use event-driven orchestration in which each specialized AI agent subscribes to workflow events and writes its decision to a centralized audit repository.

4. Question

A global company has identified several generative AI opportunities across finance, operations, customer service, and human resources. The enterprise architect must recommend one use case for the organization’s first AI pilot. Which factor should receive the highest priority when selecting the initial use case?

The use case with the highest projected user adoption based on stakeholder surveys and departmental readiness assessments.
The use case that requires the greatest volume of organizational data so the AI system can demonstrate broad enterprise knowledge.
The use case with measurable business value, such as cost reduction, efficiency gains, or SLA improvement, supported by accessible data, feasible implementation, and manageable risk.
The use case with the shortest deployment timeline, even when its expected business impact is limited.

5. Question

A generative AI service sends the same lengthy instruction set with every API request, but inserts a unique session identifier and generation time before the reusable content. Despite enabling prompt caching, nearly all requests are billed as uncached input. What is the most likely reason for the low cache hit rate?

The changing metadata appears before the reusable content, preventing the prompt prefix from matching previously cached requests.
The instruction set exceeds the maximum content size supported by prompt caching.
Only conversation messages can be cached, while system-level instructions are always processed again.
Requests arrive after the configured cache lifetime has expired, requiring the service to create a new cache entry.

6. Question

An AI application sends 18,000 tokens of technical context in each request, while an important data-handling restriction is placed between several reference sections and is sometimes ignored. How should the prompt be redesigned to make the restriction more noticeable and consistently followed?

Add an example that demonstrates how the model should respond to a request involving restricted pricing information.
Place the restriction before or after the main prompt content and keep it in a separate, clearly labeled instruction section.
Use a temperature setting of zero to force compliance with all prompt requirements.
Wrap the restriction in a descriptive XML tag while keeping it within the existing product context.

7. Question

A financial services company uses a generative AI model to create compliance reports with mandatory section titles, a fixed sequence, and standardized tables. Despite detailed formatting instructions, the generated reports still vary slightly between requests. Which prompt-engineering approach will most effectively improve formatting consistency?

Include one or two few-shot examples that demonstrate the complete required report format.
Ask the model to review and reformat its initial draft in a second pass using the same written specifications
Increase the temperature to help the model interpret the formatting rules.
Add a formatting checklist that repeats the required headings, section sequence, and table structure at the end of the prompt.

8. Question

A financial services company is designing a Claude-powered claims review process with fixed steps, audit requirements, and human approval before payment. Executives need to understand why a controlled workflow is preferred over an autonomous agent and when the decision should be reconsidered. Which recommendation best communicates the architectural decision and trade-offs?

Select the controlled workflow as the permanent architecture, emphasize its predictability, and omit its limitations or reconsideration criteria to maintain stakeholder confidence.
Select the autonomous agent for greater flexibility, accept its increased control complexity, and rely on human approval to provide the required auditability and governance.
Select the controlled workflow for lower implementation risk, emphasize its immediate operational benefits, and postpone discussing adaptability until stakeholders request another formal architecture review.
Select the controlled workflow for predictable execution and auditability, acknowledge its reduced adaptability, and define future conditions that would justify adopting an autonomous agent.

9. Question

An architect is selecting chunking strategies for different document types in a RAG pipeline. Which chunking strategy is most appropriate for each scenario?

Select the correct answer from the drop-down list of options. Each correct selection is worth one point.

Technical manuals organized into chapters, sections, and subsections.

Application logs with long, continuous sequences of timestamped events.

Support transcripts containing lengthy conversations without formal section headings.

Research articles with paragraphs covering distinct topics and ideas.

3. Question

An organization is standardizing Claude Code across multiple engineering teams and must enforce security settings that developers cannot override while centrally distributing approved tools and reusable workflows. Which actions should the organization take? (Select TWO.)

Establish a managed plugin marketplace to distribute approved skills, agents, hooks, and MCP servers across teams.
Allow developers to install tools independently, then audit local configurations periodically for policy compliance.
Use server-managed settings to enforce organization-wide security controls and prevent lower-level configuration overrides.
Publish recommended settings in a shared repository and require each team to apply them during project setup.
Store approved tools in project-level configuration so individual teams can update or disable them when needed.

4. Question

A financial services company uses Claude to evaluate complex loan exception requests that require applying several eligibility rules in the correct sequence. The model occasionally reaches an incorrect decision because it skips important intermediate conditions. Which prompt engineering technique should the architect use to improve the reliability of the evaluation?

Increase the prompt temperature to encourage Claude to explore several possible reasoning paths before selecting the most likely eligibility decision.
Use zero-shot prompting with detailed decision criteria and require Claude to return a concise final determination without showing intermediate analysis.
Use chain-of-thought prompting that guides Claude through each condition before returning the final decision.
Use few-shot prompting with representative input-output examples that demonstrate how similar loan exception requests should be consistently classified.

5. Question

An insurance company is preparing a Claude-based claims assistant and has created a golden dataset containing routine, edge, and high-risk cases. The team must evaluate exact field accuracy and explanation quality while keeping the process scalable. Which evaluation approach best meets these requirements?

Use deterministic scoring for structured fields and explanations, then broaden accepted patterns as valid responses fail automated checks.
Use criterion-specific graders and calibrate their scoring with representative human-reviewed samples before running the complete evaluation.
Use a rubric-based model grader across all response criteria, then escalate borderline and inconsistent scores for manual assessment.
Use human reviewers for high-risk responses and model graders for routine cases, then compare aggregate results across system versions.

6. Question

A compliance agent uses a monolithic context strategy that loads all policies and tools into every request, even though each case requires only a small portion of the available information. An architect recommends replacing this approach with progressive discovery. Should the architect’s recommendation be adopted?

Yes. Progressive discovery can reduce unnecessary context while preserving access to relevant information when required.
Yes. Progressive discovery guarantees that Claude will always identify the correct information without retrieval failures.
No. A monolithic strategy is preferable because loading all information upfront consistently improves response accuracy.
No. Progressive discovery is appropriate only when Claude cannot access external tools or knowledge sources.

7. Question

An investment company uses Claude to review large due-diligence packages containing legal, financial, and operational documents. The current architecture occasionally produces incomplete findings and inconsistent assessments as document volume and complexity increase. Which architectural approach would most reliably improve the review process?

Specialized subagents analyze separate document categories and return their findings directly, leaving coverage validation and inconsistency resolution to downstream reviewers.
Multiple general-purpose agents independently review the complete package, after which the most comprehensive response is selected without reconciling conflicting findings.
A single autonomous agent processes the complete package with an expanded context allocation and performs additional review passes when its initial assessment appears incomplete.
An orchestrator decomposes the review, delegates tasks to specialized subagents, verifies coverage, retries missing work, and synthesizes results into a defined schema.

8. Question

A financial services company plans to use Claude to assist with reviewing complex insurance claims. An executive sponsor wants the architecture selected immediately, but the solution will affect claims specialists, compliance teams, and existing business systems. Which actions should the team take BEFORE recommending a Claude-based solution architecture? (Select TWO.)

Compare established Claude architecture patterns and shortlist the designs that best match the sponsor’s stated business objective.
Reuse the organization’s standard AI architecture and postpone detailed discovery until the implementation phase has already begun.
Analyze historical claims data to identify common request categories, processing volumes, and opportunities suitable for automation.
Conduct stakeholder workshops to define business outcomes, operational constraints, risk tolerances, and measurable acceptance criteria.
Map the current workflows, decision points, exception paths, and dependencies across existing business systems.

9. Question

A software company is validating a Claude-powered support assistant with a requirement that p95 latency for end-to-end requests remain below three seconds during peak usage. The architecture team needs evidence that the deployed system meets this performance target. How should the team measure latency?

Use latency results from a controlled preproduction test, then extrapolate that 95% of peak-production requests will finish within three seconds.
Record end-to-end request times under expected traffic, then identify the latency value that 95% of requests do not exceed.
Survey pilot users and confirm that at least 95% report that “the assistant responds very fast” during typical usage conditions.
Measure request times during a limited off-peak test, then treat the resulting 95th-percentile value as representative of peak usage.

10. Question

An architect is assigning evaluation responsibilities across a Claude-powered application. Which evaluation approach is most appropriate for each scenario?

Select the correct answer from the drop-down list of options. Each correct selection is worth one point.

Resolve cases where the generated conclusion conflicts with the policy wording or cited claim evidence:

Compare conflicting witness statements and produce a structured list of unresolved facts:

Approve a recommendation that could deny coverage or trigger regulatory reporting:

Assess whether generated customer notices follow required tone, structure, and disclosure criteria:

Generate a concise claim summary from adjuster notes, customer statements, and uploaded evidence:

3. Question

A financial services company uses Claude to review customer onboarding documents and has latency and cost dashboards for monitoring system efficiency. Although reviewers frequently correct missed fields and inaccurate classifications, the architectural design does not use these production signals to improve testing.

Which architectural approach best supports continuous quality improvement?

Use the latency and cost dashboards as the primary indicators of model quality and investigate outputs only when operational thresholds are exceeded.
Replace the existing evaluation dataset after every model update with newly generated synthetic examples that exclude production outputs.
Create a feedback loop that captures reviewer corrections and user signals, then adds a representative sample of production outputs to a labelled evaluation set.
Store all reviewer corrections in application logs for auditing while continuing to evaluate the system against the original pre-deployment test cases.

4. Question

A business sponsor expects an AI solution built with Claude to produce completely accurate results before approving production deployment. The architecture team must address this expectation while establishing controls that reflect the financial and operational impact of incorrect outputs.

Which response best supports a responsible architecture decision?

Guarantee complete accuracy by selecting the most capable model and delaying the launch until every test case produces identical results.
Set clear expectations that LLM outputs can vary, then collaborate with the sponsor to define measurable acceptance thresholds and risk-based procedures for handling errors.
Reduce the model temperature to zero and present deterministic configuration as sufficient evidence that all production outputs will be accurate.
Accept the accuracy requirement without qualification and rely on human reviewers to identify errors after the AI solution is deployed.

5. Question

An engineering department is standardizing AI-assisted coding tools after successful pilots led by early adopters. Different teams currently maintain separate instructions, tool permissions, and project settings, causing inconsistent behavior and difficult configuration updates.

Which configuration approach best supports organization-wide consistency?

Allow each team to maintain independent local settings and coordinate changes through periodic meetings.
Ask early adopters to publish recommended configurations that developers may copy and modify as needed.
Place all project standards in personal configuration files so developers can customize required controls for each assignment.
Define shared, version-controlled project standards that apply to every team member’s tooling, while limiting personal configuration to individual preferences.

6. Question

A compliance team uses a RAG system to answer questions from regulatory manuals organized into articles, clauses, definitions, and referenced provisions. The current token-based chunks often split related sections, causing retrieval to omit definitions or cross-referenced requirements.

Which chunking approach best improves retrieval accuracy?

Create uniform token-sized chunks and increase the overlap until most referenced text appears more than once.
Use structure-aware chunking that follows clause and section boundaries, with metadata preserving links to definitions and cross-references.
Store each regulatory manual as one large chunk so every definition and reference remains in the same retrieval unit.
Generate smaller token-based chunks and rank them independently without retaining section identifiers or document relationships.

7. Question

A public agency deploys an AI assistant on its public website to answer questions about permits and eligibility. The team plans to measure user satisfaction but must also ensure that visitors understand the nature and limits of the service, especially when responses could affect consequential decisions.

Which approach best supports responsible user interaction?

Clearly disclose the service as AI-powered, outline its limitations, and enable escalation to a human for decisions with significant consequences.
Present the AI assistant as a standard support representative and disclose its automated nature only when a user reports an incorrect response.
Display a general disclaimer in the website footer and require users to verify all responses independently without offering human support.
Measure response accuracy and user satisfaction, then remove disclosure notices when the AI assistant reaches the required performance target.

8. Question

An engineering firm uses an AI assistant to answer questions from a large repository of technical standards, but sending the entire repository with every request increases token costs and causes the model to overlook relevant passages. Which strategy would most effectively improve answer reliability while reducing request costs?

Create a condensed summary of each engineering standard and include the full set of summaries with every query.
Divide the complete repository into several sequential prompts and preserve the model’s summaries between requests.
Retrieve the most relevant passages for each query and provide only those passages to the model.
Remove headings, formatting, and spacing from the repository before adding it to the prompt.

9. Question

An engineering firm uses an AI assistant to answer questions from a large repository of technical standards, but sending the entire repository with every request increases token costs and causes the model to overlook relevant passages. Which strategy would most effectively improve answer reliability while reducing request costs?

Create a condensed summary of each engineering standard and include the full set of summaries with every query.
Divide the complete repository into several sequential prompts and preserve the model’s summaries between requests.
Retrieve the most relevant passages for each query and provide only those passages to the model.
Remove headings, formatting, and spacing from the repository before adding it to the prompt.

10. Question

An enterprise AI assistant exposes dozens of tools from multiple departments, causing slower responses and frequent selection of similar but irrelevant tools. Usage logs also show that several registered tools are redundant or have not been called by any production request. What should be implemented first to improve tool-selection accuracy and efficiency?

Expand every tool description with additional examples and detailed usage instructions.
Group the tools under standardized naming conventions and shared metadata tags for easier comparison.
Introduce a scoring layer that ranks all registered tools by semantic similarity to each incoming request.
Remove unused or duplicate tools and implement dynamic discovery that exposes only context-relevant tools.

11. Question

A virtual assistant retrieves payroll records through a shared API credential that can access every worker’s profile. The assistant is instructed in its system prompt to display only the authenticated requester’s information. What is the primary security weakness in this design?

Access restrictions rely on model instructions instead of backend authorization controls, allowing prompt injection or model failure to expose other workers’ records
Shared API credentials are prohibited whenever a language model processes personal information.
The API credential requires write permissions to create a complete record of employee requests.
Each worker should provide personal HR credentials directly to the virtual assistant.

12. Question

A company uses one enterprise Claude-based assistant to process employee benefits questions, infrastructure incidents, and expense approvals. As more tools and instructions are added, the assistant increasingly selects tools from the wrong business function, and the company now plans to support vendor purchasing requests. Which architectural change will most effectively improve tool-selection accuracy while supporting future expansion?

Introduce a routing or supervisor layer that delegates requests to specialized domain agents with limited tools and focused instructions.
Add the purchasing tools to the existing assistant and rely on more detailed tool descriptions to prevent incorrect selections.
Expand the shared system prompt with additional examples covering every possible interaction across all business functions.
Require the existing assistant to call several tools for each request and compare their outputs before responding.

13. Question

A financial services company is designing an AI-powered claims process in which every submission must follow the same predefined validation sequence and produce a traceable record of each processing step. Which characteristics make a fixed workflow more suitable than an autonomous agent? (Select TWO.)

The system must interpret previously unseen document structures without predefined handling rules.
The execution sequence is established in advance and remains consistent for every claim.
The solution must independently explore alternative methods for resolving ambiguous cases.
Each processing stage must be reproducible and available for compliance auditing.
The model must dynamically choose recovery actions when external tools behave unexpectedly.

14. Question

A global consulting firm is deciding whether an AI-driven corporate investigation should use multiple agents because the work includes separate legal, financial, and cybersecurity analyses that can be completed concurrently. Which task characteristics most strongly support using a multi-agent architecture? (Select TWO.)

The investigation will process a large number of user requests after deployment.
The investigation depends on a shared evidence repository that every analysis workflow can access.
Each investigation domain requires specialized instructions, tools, and context that would be difficult for one agent to manage effectively.
The workflow requires a central coordinator to consolidate findings and maintain a consistent reporting format.
Independent investigation workstreams can execute simultaneously to shorten the total completion time.

15. Question

An industrial inventory assistant relies on semantic vector retrieval to search a large component catalog. Descriptive queries return relevant results, but searches containing exact stock codes frequently retrieve similar-looking yet incorrect components. Which retrieval approach will most effectively improve the accuracy of both query types?

Replace vector retrieval with keyword-only search for every catalog query.
Generate embeddings with a higher-dimensional semantic model.
Normalize all stock codes into natural language descriptions before indexing.
Combine exact or keyword matching with semantic search and adjust their ranking weights based on the query.

16. Question

A RAG-based generative AI support application updates its knowledge repository each night, but some responses still cite outdated procedures that were replaced during the latest synchronization. Which actions would most effectively prevent the application from retrieving superseded information? (Select TWO.)

Increase the maximum number of retrieved chunks so the model receives a broader set of documents.
Review a small sample of generated responses at the end of each month.
Implement an automated post-refresh indexing workflow that verifies all current documents were embedded completely and consistently.
Store document lifecycle and version attributes in the retrieval index, and apply filters that exclude inactive or obsolete records.
Reduce the model’s temperature to make its generated responses more deterministic.

17. Question

A product team requests that an AI assistant deliver complete answers in less than one second, although performance testing shows that retrieval, tool execution, and reasoning require at least four seconds. Which of the following should the engineering team recommend?

Define the SLA using only time to first token and treat it as the latency for a complete response.
Disable the retrieval and reasoning components to satisfy the requested response time.
Base the production SLA on the model provider’s published benchmark latency rather than end-to-end application testing.
Share the measured latency breakdown, negotiate a realistic service-level agreement (SLA), and improve perceived responsiveness through streaming and status updates.

18. Question

An enterprise support assistant has begun producing inconsistent responses over the past several days, even though its model release, prompt configuration, and latency measurements have remained stable. Which data sources should be prioritized to identify the cause? (Select TWO.)

Detailed execution traces from low-quality sessions showing retrieved content, tool requests, and tool responses.
Weekly user-feedback scores segmented by application version and customer group.
A comparison of average response length and token consumption between recent and historical sessions.
A comparison of recent user inputs against the expected distribution of topics, formats, and languages.
API availability reports showing endpoint uptime and outage history.

19. Question

A company operating a production RAG assistant needs an automated signal that can reveal deteriorating response quality before the issue becomes apparent through user behavior or feedback. Which metric would provide the most effective early warning?

A rise in customer-reported incidents related to inaccurate responses.
An increase in the assistant’s monthly token-processing costs.
A downward trend in retrieved-context relevance combined with a higher percentage of requests that cannot produce a source-grounded response.
A quarter-over-quarter decline in the number of users interacting with the assistant.

20. Question

A generative AI application summarizes content from external knowledge sources, but during testing, a retrieved file containing embedded commands causes the model to disregard its configured security rules. Which design principle should be applied to prevent this behavior?

Restrict the knowledge base to documents created by authenticated employees.
Encrypt all model instructions before storing them in the application configuration.
Apply a retrieval confidence threshold so only highly relevant document passages can influence the generated response.
Treat retrieved content as untrusted data, isolate it from trusted instructions, and prevent it from receiving instruction-level authority.

21. Question

A financial institution uses a Claude-powered assistant to prepare replies to customers disputing fees and account charges. Compliance requires safeguards that prevent the AI from communicating unauthorized refunds, credits, or other financial promises. Which of the following will meet the requirement? (Select TWO.)

Provide the model with additional examples of previous customer disputes.
Expand employee training on effective prompt-writing techniques.
Require an authorized employee to review and approve every AI-drafted reply before delivery.
Store AI-generated responses for periodic compliance audits.
Apply guardrails that prevent the assistant from making financial commitments beyond approved policies.

22. Question

A company is deploying a RAG-enabled assistant that can retrieve external knowledge-base articles and invoke internal tools, and the security team wants to minimize the impact of malicious instructions hidden in retrieved content. Which of the following would best minimize this security exposure? (Select TWO.)

Increase the retrieval relevance threshold so the agent processes only highly matched document chunks.
Clearly mark retrieved material as untrusted content and instruct the model to use it only as supporting context rather than as instructions to execute.
Configure deterministic generation to prevent the model from following injected prompts.
Apply least-privilege permissions to the assistant’s tools and limit access to sensitive actions and data.
Require the assistant to include source citations for claims derived from retrieved content.

23. Question

An organization is scaling AI coding agents across teams while maintaining consistency and operational safety. Which approaches should it implement? (Select TWO.)

Allow each team to configure AI agents independently without shared governance.
Grant AI agents full access to repositories and deployment environments to maximize productivity.
Store approved AI instructions, configurations, and reusable development workflows in version control for teams to share.
Apply execution restrictions and approval controls to sensitive agent actions, such as protected-branch changes or destructive commands.
Require teams to replace existing development tools whenever a newer AI solution becomes available.

24. Question

An AI assistant repeatedly chooses the wrong function because several available tools handle similar tasks and roles are difficult to distinguish. Which design changes would most effectively improve tool selection? (Select TWO.)

Clearly define each tool’s intended use, required inputs, and situations where it should not be used.
Create additional specialized tools for uncommon variations of each task.
Arrange tools into categories based on the type of task they perform.
Replace the tools with one universal function that interprets unrestricted natural-language instructions.
Merge or eliminate tools that perform substantially similar functions.

25. Question

A platform team is designing an AI assistant that can access hundreds of specialized tools and knowledge sources, but each user task typically requires only a few, and the model has a limited context window. Which design considerations most strongly support loading tools and information only when they are needed? (Select TWO.)

Most requests rely on the same small set of tools every time.
Only a fraction of the available resources are relevant to each individual task.
The entire task must be completed without any intermediate retrieval or selection step.
Preloading all available resources would consume excessive context and may reduce model performance.
All supporting information is compact enough to remain permanently in the prompt.

26. Question

An architect is preparing to move an AI initiative from discovery into solution design and needs to confirm that the project is ready for the next phase. Which items should be established before design begins? (Select TWO.)

Defined business outcomes with measurable evaluation targets and minimum acceptable performance levels.
A finalized prompt that will be deployed without further iteration.
Documented findings on data readiness, accessibility, quality, security, and regulatory requirements.
A complete end-user training plan and support documentation.
A finalized deployment topology specifying all production infrastructure components.