Claude Certification Scenario Questions: CCAR-F Exam Guide

Every item on the CCAR-F exam is a scenario. That is not marketing language; it is a structural fact stated in the official exam guide. If you are preparing for the Claude Certified Architect, Foundations exam and you have been drilling definition-recall flashcards, you are optimising for the wrong signal. This guide explains how claude certification scenario questions work, what cognitive moves they demand, and how to build the judgment that earns a scaled score of 720 or above.

What exactly is a scenario-based question?

A scenario-based question presents a realistic production situation, names the constraints, and asks you to choose the best architectural or engineering response. It does not ask "what is a tool description?" It asks something closer to: "A coordinator agent is routing requests to three subagents and 12% of calls are landing on the wrong specialist. Which single change most directly fixes the routing?" The answer requires you to reason from cause to fix, not to recite a definition.

Per Anthropic's exam guide, every item on the CCAR-F is scenario-based and tests practical judgment, not recall. That framing has a direct implication for study: you need to practise making decisions under constraints, not just reading concept summaries.

How are the 60 items distributed across domains?

The exam draws 60 items from five domains. Understanding the weight of each domain tells you where scenario volume concentrates.

Domain 1 alone accounts for roughly 16 of the 60 items. That means the largest single cluster of scenario questions will test your ability to design, debug, and reason about multi-agent systems. Domains 3 and 4 are tied at 20% each, making them jointly the second-largest surface area. Context management, at 15%, is the smallest domain but historically surfaces some of the trickiest reliability scenarios.

The exam also draws 4 scenarios at random from a bank of 6 at each sitting, so the specific narrative framing of your sitting will differ from a colleague's, even if the underlying task statements are the same.

What cognitive moves do scenario questions actually test?

We have analysed the 30 task statements across all five domains and identified four recurring judgment patterns that appear in the majority of scenario questions.

Root-cause tracing. The scenario describes a symptom (wrong tool selected, loop terminates early, output loses attribution) and asks you to identify the proximate cause before choosing a fix. Jumping to a solution without diagnosing the cause is the most common trap. For example, a scenario about agentic loop anti-patterns will typically describe observable behaviour and ask which architectural decision produced it.

Proportionate fixes. The exam consistently rewards the smallest change that solves the stated problem. If a scenario describes a tool routing issue caused by an ambiguous description, the correct answer is almost always to revise the description, not to restructure the entire tool catalogue. This principle appears explicitly in the exam's scoring philosophy: proportionate fixes over over-engineered solutions.

Deterministic over probabilistic when stakes are high. When a scenario involves irreversible actions, financial transactions, or compliance requirements, the exam rewards programmatic enforcement over prompt-based guidance. Understanding when to use hooks vs prompts is a concrete expression of this principle.

Structured handoffs and escalation. Many scenarios involve a decision point where the agent cannot proceed safely. The correct answer requires knowing when to pause, what information to surface to a human, and how to structure that handoff so the human can act without re-reading the entire conversation history.

How does the scenario bank affect your preparation?

Each sitting draws 4 scenarios at random from a bank of 6. This matters for two reasons. First, you cannot predict which narrative framing you will see, so you must be fluent across all five domains rather than betting on a subset. Second, the underlying task statements are fixed even when the surface story changes. A scenario about a financial services compliance agent and a scenario about a healthcare triage agent may both be testing the same task statement from Domain 1: knowing when programmatic enforcement is required versus when a prompt-level constraint is sufficient.

The practical implication is that your preparation should focus on task statements, not on memorising specific scenario narratives. When you encounter a practice scenario, ask yourself: which task statement is this testing? What is the decision the examiners want me to make? That metacognitive habit transfers across scenario variants in a way that narrative memorisation does not.

What does a well-constructed scenario question look like?

Below is an illustrative example that mirrors the style and cognitive demand of real CCAR-F items. This is not a leaked question; it is a constructed example based on the published task statements.

Scenario: A coordinator agent manages three specialist subagents: a data-retrieval
agent, a calculation agent, and a report-formatting agent. The coordinator uses
Claude's tool_choice parameter set to "auto". Over 500 production runs, the
calculation agent is invoked on 8% of requests that should have gone to the
data-retrieval agent. Logs show the two agents have tool descriptions that both
mention "financial data".

Which single change most directly reduces the misrouting rate?

A. Set tool_choice to "any" to force the model to always select a tool.
B. Rewrite the two conflicting tool descriptions to make their scopes mutually exclusive.
C. Add a pre-routing classification step that uses a separate Claude call to label
   each request before passing it to the coordinator.
D. Increase the coordinator's system prompt length to include more examples of
   correct routing decisions.

The correct answer is B. The root cause is overlapping tool descriptions, and the proportionate fix is to resolve the overlap at the description level. Option C is a valid architectural pattern but introduces latency and complexity that the scenario does not justify. Option D adds prompt length without addressing the structural ambiguity. Option A changes selection behaviour without fixing the underlying signal problem.

This example illustrates diagnosing tool misrouting: the exam expects you to trace the symptom to its cause and apply the lowest-cost fix.

Which domain produces the hardest scenario questions?

Based on the task statement complexity and the number of interacting concepts, Domain 1 (Agentic Architecture & Orchestration) produces the scenarios with the most moving parts. A single Domain 1 scenario may require you to reason simultaneously about coordinator responsibilities, subagent context isolation, error propagation, and session management.

Domain 5 (Context Management & Reliability) produces scenarios that are deceptively simple in framing but require precise knowledge of failure modes. A scenario might describe an agent that performs well for the first 20 turns of a session and then begins producing inconsistent outputs. The correct diagnosis is context degradation, and the correct fix involves summary injection for fresh sessions rather than increasing the model's context window or switching models.

Domain 4 (Prompt Engineering & Structured Output) scenarios frequently test your ability to distinguish between when few-shot examples are the highest-leverage intervention and when schema design is the real bottleneck. These scenarios require you to hold both the prompt layer and the output validation layer in mind simultaneously.

How should you practise scenario questions effectively?

Effective scenario practice has three phases that most candidates skip.

Phase 1: Diagnosis before answer selection. Before reading the answer options, write one sentence describing the root cause of the problem in the scenario. This forces you to commit to a diagnosis before the answer options anchor your thinking. Candidates who skip this step are more susceptible to plausible-sounding distractors.

Phase 2: Elimination with reasoning. For each wrong answer, articulate why it is wrong, not just that it is wrong. "Option C is wrong because it introduces an additional Claude call that adds latency and cost without addressing the description ambiguity that caused the misrouting" is useful. "Option C is wrong because it's too complex" is not.

Phase 3: Domain tagging. After each practice question, tag it to a specific task statement. Our concept library at /concepts maps 174 atomic concepts to the 30 task statements across all five domains. If you find yourself consistently missing questions tagged to a particular task statement, that is a precise signal about where to focus study time, not a vague sense that "Domain 1 is hard."

The AI Skill Certs adaptive engine uses Bayesian Knowledge Tracing with a 0.90 mastery threshold, which means it will keep surfacing scenario questions in a given task statement area until your response pattern demonstrates reliable mastery, not just a lucky correct answer.

What are the most common mistakes candidates make on scenario questions?

We see four failure patterns repeatedly in practice exam data.

Choosing the most sophisticated answer. The exam rewards proportionate fixes. Candidates with strong engineering backgrounds often gravitate toward architecturally elegant but over-engineered solutions. If the scenario describes a small, well-defined problem, the correct answer is usually a small, well-defined fix.

Ignoring the stakes signal. Scenarios that mention irreversible actions, regulatory requirements, or financial transactions are signalling that deterministic enforcement is required. Choosing a prompt-based solution in a high-stakes scenario is almost always wrong, regardless of how well-crafted the prompt is.

Conflating session management options. The exam distinguishes carefully between resuming a session, forking a session, and starting fresh. Each has a specific use case. When to resume vs fork vs fresh start is a task statement that appears in multiple scenario framings, and candidates who treat these as interchangeable will miss those items.

Missing the handoff trigger. Many Domain 1 scenarios include a decision point where the correct answer is to escalate to a human rather than continue autonomously. Candidates who are focused on agentic solutions sometimes overlook the structured handoff option entirely. The exam treats knowing when not to proceed autonomously as a core architectural competency.

How do the five domains connect in multi-domain scenarios?

Some of the more demanding scenario questions on the CCAR-F span two domains. A scenario might describe a Claude Code workflow (Domain 3) that uses an MCP server (Domain 2) and ask about the correct configuration scoping decision. Or it might describe a prompt engineering problem (Domain 4) that is actually caused by context degradation (Domain 5).

The three-level configuration hierarchy in Domain 3 is a good example of a concept that appears in cross-domain scenarios: a question might describe a team-level CLAUDE.md setting that conflicts with a project-level setting, and ask which takes precedence and what the correct resolution is. Getting this right requires knowing both the configuration mechanics and the version control implications.

For Domain 2, tool descriptions as a selection mechanism connects directly to Domain 1 coordinator logic. The coordinator's routing quality is a function of tool description quality; these two domains are not independent in production systems, and the exam reflects that dependency.

What is the scoring context for scenario questions?

The CCAR-F is scored on a scale of 100 to 1000, with a passing score of 720. Anthropic does not publish the raw-to-scaled conversion, so we will not state an exact question count as the pass mark. What we can say is that the score report gives you percent-correct by domain, which means you will know after your sitting exactly which domain scenario clusters cost you points.

The exam costs $125 USD per attempt. Given that each sitting is a fresh draw from the scenario bank, candidates who fail and re-sit will encounter different scenario framings even if the underlying task statements are the same. That is an argument for deep task-statement fluency over surface-level scenario memorisation.

The credential is valid for 12 months from the date it is awarded, which means the investment in scenario-level judgment has a defined renewal horizon. As of 3 June 2026, more than 10,000 individuals have earned certifications across the Claude Partner Network programme.

How should you use the concept library to prepare for scenario questions?

The concept library at /concepts covers 174 atomic concepts mapped to the five CCAR-F domains and 30 task statements. The most effective use of the library for scenario preparation is not to read every concept page sequentially. It is to use it as a diagnostic reference: when you miss a practice scenario, identify the concept it was testing, read that concept page, then immediately attempt two or three more scenarios tagged to the same task statement.

For Domain 1, the agentic architecture section covers the full range of coordinator, subagent, and orchestration concepts that appear in the heaviest scenario cluster. For Domain 4, the prompt engineering section covers structured output schema design, few-shot construction, and the conditions under which each technique is the highest-leverage intervention.

The goal is not concept coverage. The goal is scenario fluency: the ability to read a production situation, identify the decision the scenario is testing, and select the response that a senior Claude architect would recognise as correct.