CCAR-P mock exam

Question 1, Claude Models, Prompting & Context EngineeringMost support tickets are simple, but a few need deep reasoning. Quality, cost and latency all matter. Which model strategy do you recommend?

• ARoute by complexity, with a fast tier by default and a stronger tier for hard cases
• BLet each support agent choose a model from a dropdown on every reply
• CUse the largest model everywhere, so quality never becomes a question
• DUse the smallest model everywhere and let staff fix the harder replies

Question 2, Solution Design & ArchitectureA client wants to pull fields from invoices with one known layout, and the steps never vary. They ask for a multi-agent system. What do you recommend?

• AA single-step workflow with a schema and validation
• BAn autonomous agent that plans each document
• CA manager agent with an evaluator subagent
• DA coordinator with one subagent per field

Question 3, IntegrationAn enterprise agent has 45 tools and picks the wrong one more often as tools are added. What is the best architectural change?

• ARaise the context window and keep them all loaded
• BAdd a retry when the wrong tool is chosen
• CWrite a longer system prompt listing when to use each tool
• DScope tools by role and load the rest through progressive discovery

Question 4, IntegrationYour RAG system answers badly on documents full of tables, and the retrieved chunks show rows cut off from their headers. What do you change first?

• ARaise the temperature on generation so the model fills in the gaps
• BMove to a larger model that can infer the missing headers from context
• CIncrease the number of chunks retrieved so the headers are more likely to appear
• DUse structure-aware chunking that keeps tables and headers together

Question 5, IntegrationSeveral separate teams need their Claude apps to reach the same internal HR system, and the owning team wants to maintain the integration in one place. Which mechanism fits?

• AA nightly export pasted into the context
• BA shared system prompt describing the HR API
• CAn MCP server owned by the HR system team
• DA copy of the integration code in each app

Question 6, Evaluation, Testing & OptimizationYou need an eval set for a support assistant before launch. Which source gives you the most trustworthy results?

• AA handful of happy-path examples written by the team that built it
• BReal historical tickets labeled by reviewers, including edge cases
• CA public benchmark for general reasoning, plus a few questions from the team
• DSynthetic questions the model generates about its own product knowledge

Question 7, Evaluation, Testing & OptimizationAnswer quality fell after a model upgrade, and retrieval scores did not move. What is the sound response?

• ARerun the eval suite on both versions and pin the previous version while you tune prompts
• BRoll out to everyone and watch customer complaints to see which prompts need work
• CDouble the retrieved chunks and see whether more context brings quality back
• DRewrite all prompts from scratch immediately, then ship them with the new model

Question 8, Governance, Safety & Risk ManagementAn agent can send payment instructions that cannot be reversed. Which control suits this action best?

• AHuman approval required before the action executes
• BA disclaimer shown in the interface before the agent runs
• CDetailed logging of every payment for later review
• DA monthly sample audit of what the agent has sent

Question 9, Governance, Safety & Risk ManagementA healthcare client wants to process patient records with Claude. What should you settle before design work begins?

• AHow many tools the first agent should have and which team will own them
• BWhich model produced the highest score on a public reasoning benchmark
• CWhich prompt style the clinicians prefer for summaries of patient notes
• DWhether the deployment route and contract terms cover regulated health data

Question 10, Stakeholder Communication & Lifecycle ManagementAn executive demands "100% accuracy" from a contract review assistant. What is the right response?

• ARefuse the target outright, because LLMs are non-deterministic by design
• BPromise it for the pilot only, then renegotiate the target once results are in
• CTurn it into measurable targets, with human review where errors are costly
• DAgree to the target, and handle any misses through support after launch

Question 11, Developer Productivity & Operational EnablementTwenty engineers use Claude Code with wildly different results. What gives you the biggest team-wide gain for the least effort?

• AA shared project configuration with conventions, rules and common workflows
• BMandatory weekly prompt writing workshops for all twenty engineers
• CA rule that everyone must use the same model and the same settings
• DA dashboard ranking the engineers by token usage and merged pull requests