CCAR-P exam prep guide

What the exam covers

Translating business problems into Claude-based solutions, designing end-to-end architectures (input, processing, output, feedback loops), choosing between workflow, agentic, and augmented-LLM patterns, decomposition of complex problems, and aligning designs to business value (cost, efficiency, performance SLAs).

What to prioritize

The workflow-versus-agent decision. Anthropic's own guidance is consistent: use the simplest pattern that works, and reach for autonomous agents only when the task genuinely requires open-ended decision-making. Know the standard composable patterns (prompt chaining, routing, parallelization, orchestrator-workers, evaluator-optimizer) well enough to pick the right one for a described scenario, not just define them.

Gotchas

Candidates who build agents for a living tend to over-select agentic answers. Exam scenarios often describe problems where a fixed workflow or even a single well-prompted call is the better answer, and "add more autonomy" is the distractor. Also expect questions that hinge on business constraints (an SLA, a cost ceiling) rather than technical elegance.

Starting points

Anthropic's essay "Building effective agents" is the canonical reference for these patterns. The Claude documentation's overview of agent architectures covers the same ground in reference form.

What the exam covers

Selecting Claude models based on trade-offs (capability, latency, cost), system prompt design, templates and guardrails, standard prompting techniques (zero-shot, few-shot, chain-of-thought), context window management, token optimization, and prompt reuse strategies including prompt caching, modular prompts, and Skills.

What to prioritize

Prompt caching mechanics are directly examinable: know that caching works on stable prefixes, so static content (system prompt, policy documents) goes before dynamic content, and that this reduces both latency and cost. Understand context engineering as a budget problem: what earns a place in the window, what gets retrieved on demand, what gets summarized. On model selection: reason from trade-offs, not model names.

Gotchas

The distractors here are usually destructive shortcuts: truncating documents, blindly downsizing the model, stuffing everything into few-shot examples. The correct answers preserve required context while restructuring for reuse. Also note the exam uses Anthropic's vocabulary: if Skills, system prompts, and prompt templates are vague terms to you, fix that in the docs first.

Starting points

The prompt engineering section of the Claude documentation, the prompt caching documentation, and Anthropic's engineering post on effective context engineering for AI agents.

What the exam covers

Tool and agent configuration (including spotting capability bloat), authentication and authorization gaps, choosing connection protocols (MCP, direct API or CLI, agent-to-agent), multi-agent system design and orchestration strategies, RAG pipeline design with chunking and indexing strategies, retrieval strategies matched to data shape, progressive discovery versus monolithic context, and observability at scale.

What to prioritize

Least privilege, applied to tools. The official sample question is explicit: an agent whose users only read tickets and draft replies should not have refund and delete tools at all. Removing an unneeded capability beats logging it or gating it with confirmations. Know MCP at the architecture level (what the protocol standardizes, when it beats a bespoke integration) and RAG design choices: chunk size, indexing, hybrid retrieval, and what breaks when documents refresh.

On orchestration: know when a single agent with good tools beats a multi-agent system, and when orchestrator-worker structures earn their complexity. Know the failure modes autonomy introduces (compounding errors, runaway cost, unpredictable tool use) and the controls that bound them: stop conditions, budgets, human checkpoints.

Gotchas

Compensating controls are the classic distractor: audit logging, confirmation prompts, and bigger models all appear as attractive wrong answers to authorization problems. Also expect scenarios where the correct answer is to remove agents, capabilities, or moving parts entirely.

Starting points

The Model Context Protocol documentation and specification, the tool use section of the Claude documentation, and Anthropic's engineering post on how it built its own multi-agent research system.

What the exam covers

Defining evaluation metrics (accuracy, latency, cost, safety), designing evaluation datasets and test frameworks, A/B testing, and diagnosing failures such as prompt failure, hallucination, and model mismatch. Also covers the difference between code-graded, model-graded, and human-graded evals, and when each is appropriate.

What to prioritize

Reason from evidence. Know what evidence would change a model selection decision. On evaluation design: understand the trade-offs between code-graded (fast, scalable, limited scope), model-graded (flexible, subjective risk), and human-graded (ground truth, expensive) approaches, and which scenarios each fits.

Gotchas

Diagnosis questions reward checking the boring thing first. The official sample question about a RAG system going wrong after a document refresh is instructive: the answer is stale retrieval, not model drift. If a scenario says the model version did not change, stop blaming the model.

Starting points

The Claude documentation's guide to defining success criteria and building evals, and the model overview and pricing pages for current trade-offs.

What the exam covers

Guardrails and safety controls, LLM risk and failure modes, human-in-the-loop validation, regulatory compliance (GDPR, HIPAA, FedRAMP appear by name in the blueprint), and ethical considerations: bias, fairness, transparency.

What to prioritize

Match the control to the risk. Know when human review is mandatory versus advisory, what a layered guardrail design looks like (input validation, output checks, scoped capabilities, monitoring), and roughly what GDPR, HIPAA, and FedRAMP each imply for an architecture that handles regulated data.

Gotchas

Safety questions are rarely "add more restrictions." Blanket refusals and heavy gating have costs, and the exam rewards proportionate controls. Do not confuse detective controls (logging, monitoring) with preventive ones (scoping, validation); several question patterns turn on that distinction.

Starting points

Anthropic's documentation on guardrails and reducing hallucinations, and its published usage policy.

What the exam covers

Structured discovery, stakeholder communication, communicating trade-offs, managing expectations and SLAs, documentation and implementation guidance, and the full system lifecycle: discovery, design, handoff, monitoring, iteration.

What to prioritize

This domain is easy to dismiss and worth a full 14%. If your background is heavier on building than consulting, this is likely your cheapest score improvement. Study structured discovery (how to surface requirements and constraints before designing), how to communicate architectural trade-offs to non-technical stakeholders, and how to set and manage SLAs.

Gotchas

CCAR-P is the only exam in Anthropic's current program that grades the consulting half of the role. Technical candidates who skip this material consistently report it as the source of their weakest domain scores.

Starting points

The official exam guide's objectives for domain 6 are the best study checklist here, since this material is about practice, not product features.

What the exam covers

Monitoring and observability, logging, cost and token optimization in production, team enablement with tools such as Claude Code, and rolling out model or prompt changes safely.

What to prioritize

Know what you would log, what you would alert on, and how you would roll out a model or prompt change safely. Understand the developer tooling layer: what Claude Code enables, how it fits into team workflows, and what its limits are.

Gotchas

This is the material technical candidates skip. At 7% it is lightweight, but underpreparation here costs points that are easy to recover.

Starting points

The official exam guide's objectives for domain 7 are the primary reference. The Claude Code documentation covers the developer tooling aspects.