Claude Certified Architect - Professional (CCAR-P)

An enterprise wants a Claude solution that answers HR policy questions from a curated knowledge base with citations. Latency must stay under 3 seconds and answers must be grounded. Which architecture is the most appropriate starting point?

• AA multi-agent swarm with specialized agents negotiating every answer
• BA RAG-augmented single-turn or short workflow design with retrieval, grounding instructio…
• CFine-tune a custom model on HR policies before any retrieval layer
• DAn open-ended agent with unrestricted web browsing as the only knowledge source

A team proposes an orchestrator agent that delegates research, drafting, and compliance review to worker agents for a regulated report. When is multi-agent orchestration justified over a single well-prompted Claude call with tools?

• AWhenever more than one tool is needed
• BOnly when specialized roles, independent evaluation, or parallel workstreams clearly impr…
• CWhenever the prompt exceeds 1,000 tokens
• DAlways, because multi-agent systems are the Professional standard

You must design a claims triage system that classifies severity, extracts structured fields, and routes complex cases to human adjusters. Which high-level pattern best matches this requirement set?

• APure single-shot completion with no tools or routing
• BA workflow/routing design with structured extraction, classification, and human-in-the-lo…
• CAn unrestricted autonomous agent that pays claims without human review
• DBatch fine-tuning only, with no runtime Claude calls

A product owner wants sub-500ms p95 latency for autocomplete-style suggestions, while a separate research assistant can take up to 30 seconds. How should the architect allocate Claude models and patterns?

• AUse the same multi-agent Opus stack for both to maximize quality uniformly
• BUse a low-latency model and short context path for autocomplete; reserve deeper models, t…
• CForce both workloads through max effort with maximum tokens
• DDisable streaming for autocomplete so full answers are buffered

Which statement best describes a production-ready target architecture for an enterprise Claude system?

• AA notebook demo with hardcoded API keys and no observability
• BDocumented components for ingress, auth, orchestration, model/tool access, retrieval, eva…
• CA single prompt file shared over chat with no version control
• DDirect browser access to Claude with no application boundary

An architect is comparing prompt chaining versus a single long agent loop for a multi-step document analysis pipeline with stable, sequential stages. Which choice is usually better and why?

• AAlways use a free-form agent loop because chaining is obsolete
• BPrefer prompt chaining or a structured workflow when stages are stable and independently…
• CUse multi-agent negotiation for every document stage
• DSkip intermediate stages and ask one model call for the final answer only

A retailer wants Claude to generate personalized marketing copy. Legal requires brand-safe outputs and marketing wants high volume. Which architecture decision best balances throughput and control?

• AUnrestricted agent with full CRM write access generating and publishing copy automatically
• BGeneration pipeline with policy-constrained prompts, safety checks, batch evaluation samp…
• CManual writing only with no Claude involvement
• DA multi-agent system that publishes first and audits later

When designing capacity for a Claude-backed API that spikes at 9am daily, what should the architect plan for first?

• AOnly average QPS, ignoring peaks
• BPeak concurrency, token throughput, queueing/backpressure, timeouts, and fallback behavio…
• CUnlimited retries with no backoff
• DDisabling caching to force fresh model calls for every request

A client asks for 'an agent that can do anything in our SAP landscape.' What is the architect's best first response in solution design?

• AGrant the agent admin credentials to SAP immediately
• BDecompose into scoped use cases, data boundaries, allowed tools, success metrics, and hum…
• CBuild a single mega-prompt that lists every SAP transaction
• DRefuse any SAP integration on principle

Which trade-off is most accurate when choosing between a tool-using agent and a pure RAG chatbot for internal knowledge work?

• AAgents always cost less than RAG
• BRAG is strong for grounded answers over documents; agents add value when multi-step actio…
• CRAG cannot cite sources while agents always can
• DAgents eliminate the need for evaluation

An architecture decision record (ADR) for a Claude system should primarily capture which of the following?

• AOnly the final prompt text with no rationale
• BContext, options considered, decision, consequences, and constraints (cost, latency, comp…
• CA list of employees who like Claude
• DUnstructured chat logs from the design workshop only

You need high reliability for a multi-step agent that calls external tools. Which design practice best improves reliability?

• AIgnore tool errors and always continue generating
• BDefine explicit tool contracts, timeouts, retries with backoff, idempotency where needed,…
• CIncrease temperature to 1.0 so the agent invents tool results
• DRemove all logging to reduce overhead

A bank wants Claude to assist relationship managers. Data residency requires EU processing only, and PII must not leave approved systems. Which architectural constraint is most critical to bake into the design?

• AUse the cheapest global endpoint regardless of region
• BSelect approved regional endpoints/deployment options, minimize PII in prompts, and keep…
• CPaste full CRM records into every prompt for maximum context
• DStore all prompts indefinitely in an unencrypted shared drive

Which is the best description of an evaluator-optimizer pattern in Claude architectures?

• ATwo models argue randomly until one times out
• BA generator produces a candidate; an evaluator scores it against criteria and the system…
• CAlways use five agents for every user message
• DReplace evaluation with higher temperature sampling

A startup wants to ship a Claude feature in two weeks. Stakeholders demand perfect multi-agent automation on day one. What should the architect recommend?

• AShip the multi-agent system without evaluation because speed matters more
• BDeliver a thinner vertical slice with clear success metrics and evals, then incrementally…
• CWait two years until a perfect platform exists
• DOutsource the entire architecture to an unreviewed vendor prompt pack

How should cost be treated in Professional Claude solution design?

• AIgnore cost until after go-live
• BModel token, tool, retrieval, and human-review costs against expected volume and set guar…
• CAlways use the largest model with maximum output tokens
• DCharge users nothing and never measure usage

For a long-running investigation workflow that may take hours, which orchestration approach is most appropriate?

• AKeep a single synchronous HTTP request open for hours without checkpoints
• BUse durable workflow orchestration with persisted state, resumable steps, tool results st…
• CRestart from scratch on every browser refresh with no state
• DStore intermediate secrets in client-side localStorage only

When should an architect typically prefer Claude Haiku over Claude Opus in a multi-step system?

• AFor the final legal sign-off on a high-stakes contract with no review
• BFor high-volume, low-latency classification, routing, or simple extraction stages where c…
• CWhenever accuracy is more important than cost, regardless of task
• DOnly when the user explicitly types the word Haiku

What is the primary purpose of putting stable role, policy, and output-format rules in the system prompt?

• ATo increase temperature automatically
• BTo provide authoritative, reusable instructions that constrain behavior across turns more…
• CTo disable tool use
• DTo bypass safety systems

An enterprise chat app repeatedly sends the same 8k-token policy manual in every request. Which context-engineering lever should be applied first?

• ADuplicate the manual three more times for emphasis
• BUse prompt caching / stable prefix strategies and retrieve only relevant sections instead…
• CRaise max_tokens for every response to match the manual size
• DRemove all user messages to free space

When is deeper thinking (a higher effort level) most justified in a Professional architecture?

• AFor every autocomplete keystroke
• BFor hard multi-step reasoning or planning tasks where quality gains outweigh latency and…
• CAs a substitute for retrieval when documents are unavailable
• DTo hide intermediate tool failures from logs

A pipeline must emit machine-readable JSON that downstream services validate against a schema. What is the best prompting/architecture approach?

• AAsk for 'JSON-ish' free text and parse with regex hope
• BUse structured outputs / tool-based schema constraints, validate responses, and define re…
• CAccept any string and cast blindly to objects
• DIncrease temperature to encourage creative field names

How should few-shot examples be chosen for a production classifier prompt?

• ARandom unrelated examples from the internet
• BRepresentative, labeled examples covering common and edge cases, versioned and tested for…
• CAs many examples as possible until the context window is full every time
• DOnly adversarial examples with no normal cases

A conversation approaches the context limit. Which strategy is most architecturally sound?

• ASilently drop the system prompt to free tokens
• BSummarize older turns, retain critical entities and tool results, keep system policies, a…
• CConcatenate all history forever without summarization
• DReset the user identity on every message

Why might an architect use Sonnet for a complex tool-using agent loop rather than Haiku?

• AHaiku cannot call tools under any configuration
• BSonnet generally offers stronger multi-step reasoning and tool-use reliability for comple…
• CSonnet is always free
• DHaiku only works offline

What is a key risk of stuffing retrieved chunks into context without ranking or budget limits?

• AIt always improves accuracy without trade-offs
• BNoise, distraction, higher cost/latency, and potential contradiction among chunks can deg…
• CIt disables the model completely
• DIt guarantees GDPR compliance

Which prompting practice best reduces prompt-injection impact in a tool-using enterprise assistant?

• APut untrusted user content and tool outputs in clearly delimited sections and instruct th…
• BConcatenate user text into the system prompt as authority
• CDisable all system instructions
• DTrust tool outputs more than system policy by default

A team moving to Claude Sonnet 5.5 used temperature 0 for extraction and temperature 0.8 for creative marketing variants on an older model. What should the architect do?

• AKeep one enterprise-wide temperature for every route
• BRemove the sampling parameters (current models reject non-default values with a 400 error…
• CRely on temperature changes instead of evaluation
• DRaise temperature on extraction to improve factual accuracy

What is the best use of stop sequences or explicit output delimiters in production prompts?

• ATo make the model ignore safety systems
• BTo bound generation into parseable regions and reduce trailing chatter that breaks parsers
• CTo increase token billing artificially
• DTo replace authentication

An architect needs Claude to follow a multi-constraint style guide (tone, banned phrases, length, locale). Which approach is most robust?

• AMention constraints casually once in a user message only
• BEncode constraints clearly in system/developer instructions, provide examples, validate o…
• CRely solely on the model remembering verbal feedback from last month
• DHide constraints in tool names only

When is Model Context Protocol (MCP) generally a better fit than ad-hoc one-off HTTP tool wrappers?

• AWhen you need a standardized way to expose tools, resources, and prompts to Claude client…
• BWhen you want to avoid all authentication forever
• CWhen the only goal is to increase prompt size
• DWhen you must prevent any tool use

A backend service needs server-to-server Claude calls with strict service identity and no interactive user. Which integration style fits best?

• ABrowser-only session cookies shared among servers
• BServer-side Claude API integration using service credentials stored in a secret manager,…
• CEmbedding API keys in mobile apps
• DEmailing prompts to a shared inbox for manual processing

Which statement correctly contrasts MCP tools vs MCP resources?

• ATools are side-effecting or computational actions; resources are readable contextual data…
• BResources execute writes; tools are read-only by definition always
• CThey are identical concepts with different marketing names
• DResources replace authentication for tools

Designing a production RAG pipeline, which order of stages is most appropriate?

• AGenerate answer first, then retrieve documents that match the answer
• BIngest/index → retrieve relevant chunks for the query → ground generation on chunks → cit…
• CTrain a new foundation model weekly instead of retrieving
• DSkip indexing and paste the entire corporate drive into every prompt

How should tool authentication be designed when Claude agents call internal APIs?

• APass the user's plaintext password in the prompt
• BUse delegated, scoped tokens or service identities with least privilege, short TTL, and a…
• CShare one global admin token across all tools and tenants
• DDisable auth for internal tools because the agent is trusted

An integration must fan out the same user question to a search tool, a CRM lookup, and a knowledge base, then synthesize. Which pattern fits?

• AStrict single-tool serial only, never parallel
• BParallel tool calls / parallelization with aggregation and conflict handling in the synth…
• CIgnore CRM and knowledge base to save tokens always
• DRun tools after sending the final answer to the user

What is a sound approach to chunking documents for enterprise RAG?

• AOne giant chunk per company forever
• BChunk with overlap/structure awareness (headings, semantic boundaries), store metadata (s…
• CRandom 10-character chunks only
• DChunk only images and discard text

A Claude agent must create a support ticket. Which tool design is best?

• AA vague tool named do_stuff with a free-form string argument and no schema
• BA clearly named create_support_ticket tool with typed parameters, validation, idempotency…
• CAllow the model to SSH into production and run SQL freely
• DHardcode ticket creation in the client with no model tool

Which failure mode is most associated with poor RAG hybrid search design?

• APerfect recall with zero cost always
• BMissing exact identifiers (ticket IDs, SKUs) when using semantic search alone without key…
• CGuaranteed prevention of all hallucinations without generation
• DAutomatic compliance certification

How should webhooks or event-driven systems integrate with Claude processing?

• ABlock the webhook sender for minutes while the full agent loop completes synchronously
• BAcknowledge quickly, enqueue durable work, process asynchronously with retries/dead-lette…
• CDrop events if the model is busy
• DProcess only during manual business hours with no queue