CCAR-P
Free CCAR-P practice questions
10 questions from our CCAR-P practice tests, each with the correct answer and a written explanation. Read them with the answers shown, or switch to quiz mode and answer first.
Question 1 of 10, Solution Design & Architecture
A due-diligence platform must research 40 companies per request, each requiring independent web and filings research, then produce a comparative report. A single-agent prototype takes 50 minutes and frequently exhausts its context window. Which redesign BEST addresses both issues?
- AAn orchestrator agent that delegates each company to a subagent with its own context window and has subagents return condensed findings for synthesis Correct answer
- BA prompt-chaining workflow that researches the 40 companies strictly one after another, persisting each result to a file before starting the next one
- CA single agent that calls the Message Batches API for every research step so that the 40 companies are processed at a discount during off-peak hours
- DA single agent with a much larger context window and server-side compaction enabled so it can keep all 40 companies' raw research in one long session
Answer and explanation: The work is breadth-first and parallelizable, which is where an orchestrator-worker multi-agent design shines: subagents explore in parallel within separate context windows and return compressed results, cutting wall-clock time and protecting the lead agent's context. A bigger context window with compaction still runs serially and keeps the latency problem. Sequential chaining fixes context exhaustion but not the 50-minute runtime. The Batches API is asynchronous with results up to 24 hours later, which worsens interactive latency.
Question 2 of 10, Solution Design & Architecture
An insurer wants Claude to process first-notice-of-loss claims. Every claim follows the same five steps (extract fields, validate policy, check fraud signals, draft a summary, route to an adjuster queue), auditors require a reproducible execution path, and per-claim latency must be predictable. Which architectural pattern BEST fits these requirements?
- AAn evaluator-optimizer loop where one Claude call drafts the full claim outcome and a second call critiques and revises it until a quality bar is met
- BAn autonomous agent given all five capabilities as tools, deciding the order of steps for each claim
- CA deterministic prompt-chaining workflow in application code, with a programmatic validation gate between steps and a Claude call per step Correct answer
- DAn orchestrator-workers design where a lead Claude call dynamically plans subtasks for each claim and spawns parallel worker calls to execute them
Answer and explanation: A fixed, well-understood sequence with audit and latency-predictability requirements is the textbook case for a workflow (prompt chaining) whose path is defined in code, with gates between steps. A fully autonomous agent chooses its own path, which undermines reproducibility and predictable latency. Orchestrator-workers suits tasks whose subtasks cannot be known in advance, which is not the case here. An evaluator-optimizer loop adds a variable number of iterations and does not provide the fixed, auditable step sequence required.
Question 3 of 10, Claude Models, Prompting & Context Engineering
A multi-agent research product uses one model for planning and synthesis and runs many subagents that perform simple, high-volume source extraction. Costs are high and latency is acceptable. Which model allocation is MOST cost-effective while protecting quality?
- AThe smallest available model everywhere, including planning and synthesis, since extraction quality is the only thing that affects the report
- BA mid-tier model for the orchestrator and the most capable model for the subagents, because subagents read far more raw source material
- CA highly capable model for planning and synthesis, and a small fast model for the extraction subagents, validated by evals Correct answer
- DThe most capable model everywhere, since mixing tiers causes inconsistent formatting between agents
Answer and explanation: Matching model tier to subtask difficulty is a key cost lever: complex planning and synthesis benefit from a capable model, while simple, high-volume extraction can usually run on a small fast model, with evaluations confirming quality. Using the smallest model for planning risks poor decomposition and synthesis. Using the top model everywhere wastes spend on simple work, and formatting consistency is handled by output specifications. Putting the most expensive model on the high-volume, simple tasks inverts the allocation.
Question 4 of 10, Integration
An agent platform will expose about 400 internal tools across 30 MCP servers. Loading every tool definition into each request consumes a large share of the context window and tool-selection accuracy drops. Which approach BEST addresses this?
- ASplit the 400 tools randomly across four agents and let users pick an agent for each task
- BMerge all 400 tools into one generic execute_action tool that takes a free-text command string, so only a single definition is ever loaded
- CMark most tools for deferred loading and give Claude a tool-search capability so only relevant definitions are loaded when needed Correct answer
- DLoad all 400 definitions but enable prompt caching on the tool block, so the context-window consumption no longer affects tool accuracy
Answer and explanation: Progressive discovery, such as deferred tool loading combined with a tool-search tool, keeps the base context small and loads only the definitions relevant to the current task, improving selection accuracy and cost. A single free-text tool removes schema validation and makes behavior unpredictable and hard to secure. Caching lowers the cost of repeated tokens but the definitions still occupy context and still confuse selection. Random splitting pushes the routing problem onto users without matching tools to intent.
Question 5 of 10, Integration
A legal-research assistant answers questions over 80,000 case summaries. Users search both with exact citations like '2019 WL 4412' and with conceptual questions like 'duty to warn in product liability'. Retrieval currently uses dense embeddings only and misses many citation lookups. What is the BEST retrieval change?
- AAdd lexical BM25 search alongside the embeddings and fuse the two result lists, then rerank the merged candidates before passing them to Claude Correct answer
- BReduce chunk size to a single sentence per chunk so that each citation string dominates the embedding vector of the chunk that contains it
- CReplace the embedding model with a larger embedding model of higher dimensionality so that exact citation strings are captured more precisely
- DIncrease the number of retrieved chunks from 10 to 100 so that citation matches are more likely to appear somewhere within the retrieved set
Answer and explanation: Exact identifiers are a lexical-match problem that dense embeddings handle poorly; hybrid retrieval that adds BM25 and fuses rankings, followed by reranking, serves both exact and conceptual queries. A bigger embedding model is still a semantic representation and remains weak on arbitrary codes. Retrieving 100 chunks inflates cost and dilutes context without guaranteeing the lexical match. One-sentence chunks destroy the surrounding context needed for conceptual questions.
Question 6 of 10, Evaluation, Testing & Optimization
A RAG assistant sometimes states policy details that do not appear in any retrieved document. Retrieval logs show the correct documents were retrieved. Which mitigation MOST directly reduces these hallucinations?
- ARaise the temperature so the model's answers vary between runs, letting users regenerate until they get a grounded answer
- BInstruct Claude to answer only from the provided documents, to say when the answer is not present, and to cite supporting quotes Correct answer
- CIncrease the number of retrieved documents from five to fifty, so that every possible policy detail is available somewhere in context
- DSwitch to a smaller model that has less background knowledge, so it is less likely to add details from its training data
Answer and explanation: When retrieval is correct but answers contain unsupported details, the fix is in grounding: restricting answers to the provided context, explicitly permitting 'not found', and requiring citations or quotes reduce fabrication and make unsupported claims visible. Retrieving fifty documents dilutes context and adds cost without addressing grounding. Smaller models are not more faithful in general. Higher temperature increases variability and pushes verification onto users.
Question 7 of 10, Evaluation, Testing & Optimization
A team must decide whether a new prompt version for a support-reply generator is better than the current one. Which evaluation approach is MOST appropriate before rollout?
- AHave the prompt author review ten outputs from the new version and approve it if those replies read noticeably better than recent production replies
- BRun both versions on the same representative test set with edge cases, scored by a rubric grader calibrated to human labels Correct answer
- CShip the new prompt to all users and compare the average customer satisfaction score for the week before and the week after the change
- DAsk Claude to compare the two prompt texts directly and state which one is likely to produce better replies, then deploy the recommended version
Answer and explanation: A sound comparison holds the dataset constant, covers representative and edge cases, and uses a scalable grader whose judgments have been checked against human labels. Ten hand-picked outputs reviewed by the author are too few and biased. A before/after production comparison is confounded by seasonality and traffic mix and exposes all users to an untested change. Judging prompt text without running it on data measures nothing about actual outputs.
Question 8 of 10, Governance, Safety & Risk Management. Choose 2.
A consumer app wants to block jailbreak attempts and harmful requests before they reach its main Claude-powered assistant, with low added cost. Which TWO measures are MOST appropriate? (Choose two.)
- AEnforce limits in application code, such as allowed topics, maximum lengths and per-user rate limits, independent of model behavior Correct answer
- BRaise the main assistant's temperature so its responses vary and attackers cannot reproduce a successful jailbreak consistently
- CDisable logging of user inputs so that attackers cannot learn from the system's past behavior and refine their jailbreak attempts
- DScreen inputs with a lightweight harmlessness classifier, such as a small fast Claude model with a focused screening prompt Correct answer
- ERely on the main assistant's system prompt alone, because one well-written instruction reliably prevents all known jailbreaks
Answer and explanation: A small, fast model used as a pre-screen catches many harmful inputs cheaply, and deterministic application controls such as topic scoping, length limits and rate limits do not depend on model judgment; layering both gives defense in depth. A single system prompt is not a complete defense. Higher temperature does not prevent jailbreaks and degrades quality. Disabling logging blinds the defenders rather than the attackers and hinders incident response.
Question 9 of 10, Stakeholder Communication & Lifecycle Management
At the start of an engagement, a retail client says it wants 'an AI agent for customer service'. What should the architect do FIRST in discovery?
- ASelect the multi-agent architecture and the cloud platform first so that the engineering team can start provisioning infrastructure in parallel
- BDraft a 50-page architecture document covering every possible customer-service capability so that no future requirement is missed later on
- CBuild a working demo with the most capable Claude model and full tool access so stakeholders can see what is possible before requirements are set
- DMap the specific customer journeys, volumes, costs and constraints, then agree measurable success criteria with stakeholders Correct answer
Answer and explanation: Structured discovery starts by translating a vague goal into concrete use cases, baselines, constraints and measurable success criteria that stakeholders agree to. A demo before requirements anchors expectations on capabilities rather than needs. Choosing architecture and platform before understanding the problem risks rework. An exhaustive document without prioritized use cases delays value and does not define success.
Question 10 of 10, Developer Productivity & Operational Enablement
A platform team wants every engineer's Claude Code sessions to block reading .env files and secrets directories, and individual developers must not be able to override this. Where should the rule be configured?
- AIn each repository's CLAUDE.md file, with a sentence instructing Claude never to open .env files or anything under the secrets directories
- BIn each developer's user settings file in their home directory, populated by an onboarding script that runs once on first install
- CIn the project's shared settings file committed to the repository, so that the deny rule travels with the code to every checkout
- DIn deny permission rules within managed (enterprise) settings deployed by IT, which take precedence over user and project settings Correct answer
Answer and explanation: Managed settings deployed by administrators have the highest precedence and cannot be overridden by users, and deny rules block the matching tool access. CLAUDE.md is guidance loaded into context, not an enforced permission boundary. User settings can be edited by the developer. Project settings are shared but can be overridden by higher-precedence settings and only apply to that repository.