Free Claude CCAO-F Exam Questions (2026 Update)

Explanation

This question tests the judgment logic of the "Product and Model Selection" domain. The core principle is: model selection must match task complexity, quality requirements, latency sensitivity, and budget—not simply "choose the strongest" or "choose the cheapest."

Why C is correct:

- The task is "meeting summary," which is a routine language processing task not requiring deep reasoning.

- The requirement is "moderate volume," meaning it's neither a high-frequency batch scenario extremely sensitive to unit cost, nor a small, low-frequency scenario where cost can be ignored.

- Sonnet's positioning is "balanced": significantly stronger quality and reasoning capability than Haiku, but with much lower latency and rate-limit consumption than Opus, making it suitable for everyday work . In exam logic, this is the standard answer for "not too hard, not too trivial" tasks.

Why the other options are wrong:

- A. Haiku: The statement "only suitable when fastest turnaround is needed" is incorrect. Haiku's actual positioning is high-frequency, low-latency scenarios (such as sub-agents, customer support automation, content moderation) . For meeting summaries, Haiku would work, but the question emphasizes "balancing cost, speed, and quality"—Haiku leans toward low cost and speed but sacrifices quality margin, so it's not the optimal balance point.

- B. Opus: Opus prioritizes depth and complex reasoning, but this task doesn't require deep reasoning. Using Opus would introduce unnecessary latency and rate-limit consumption, resulting in over-provisioning and failing the "cost balance" requirement .

- D. Custom model: The Associate-level user does not need to build or train a model independently. The exam's governance and product logic consistently assumes that users select existing models rather than train their own . This option violates the role boundary.

Reference

This question belongs to Product and Model Selection (12% weight). The core test point is: matching the three model tiers—Haiku (fast/cheap), Sonnet (balanced), Opus (deep)—with task requirements, rather than selecting based on a single dimension.

Explanation

This question tests the "Product and Model Selection" domain's core decision logic: when a task has multi-turn interaction, rapid response per turn, and light reasoning, model selection should prioritize latency over reasoning depth.

Why A is correct:

Haiku is explicitly positioned by Anthropic as the fastest and lowest-latency model tier, specifically designed for high-frequency, latency-sensitive scenarios, including classification, routing, extraction, and sub-agent tasks . In official documentation, Haiku's latency is rated as "fastest," significantly ahead of Sonnet ("fast") and Opus ("medium") . For interactive workflows where every turn requires quick response and reasoning is not complex, Haiku is the optimal choice that satisfies both latency requirements and cost efficiency.

Why the other options are wrong:

- B. Choose based on alphabetical order of model names: This is an absurd distractor with no technical rationale. Model selection should be based on task characteristics (latency, reasoning depth, cost), not naming conventions.

- C. Choose any model because latency doesn't vary by model: This is factually incorrect. Anthropic officially rates different models with different latency tiers: Haiku fastest, Sonnet fast, Opus medium, Fable slower . Latency is a key dimension in model selection.

- D. Choose Opus to maximize reasoning depth despite higher latency: This directly contradicts the question requirements. The question explicitly states "per-turn reasoning is light" and "each turn must respond quickly," while Opus's positioning is deep reasoning at the cost of higher latency . Using Opus would result in over-provisioning, introducing unnecessary latency and cost.

Reference

This question belongs to Product and Model Selection (12% weight). The tested skill is: matching model tiers (Haiku/Sonnet/Opus) to task requirements based on the three-dimensional trade-off of latency, reasoning depth, and cost, rather than selecting based on a single dimension or irrelevant factors.

Explanation

This question examines the sequence judgment for model selection within the "Product and Model Selection" domain. The core principle is: before testing candidate models on a representative sample, you must first clarify the requirements baseline, then shortlist candidates based on that baseline—testing is only used to validate "whether the shortlisted candidates meet the standards," not to screen from scratch.

Why A is correct:

Option A requires defining cquality standards, latency tolerance, and expected request volume—these three are the prerequisite inputs for model selection. Without these benchmarks, you have no criteria to judge whether a model's output is "good enough," and you cannot evaluate whether latency and cost are acceptable. Anthropic's official model selection guide similarly lists Capabilities, Speed, and Cost as the first step in evaluation . This step must be completed before testing, otherwise the testing results cannot be interpreted.

Why C is correct:

Option C is to cselect candidate models based on the requirements profile and known capabilities. After clarifying the requirements (A), the next logical step is to shortlist based on each model tier's public positioning (Haiku fast/cheap, Sonnet balanced, Opus deep reasoning) . Testing itself consumes time and resources, so you should not test all models; instead, you should first use known information to narrow the scope before conducting representative sample testing on the shortlisted candidates.

Why the other options are wrong:

- B. Retest the selected model when a new version is released: This is a -deployment maintenance action, not a pre-testing step. Retesting upon new version release is a long-term operational practice, unrelated to "requirements steps before testing candidate models."

- D. Promote the selected model to production traffic: This is the final step after testing and selection are complete; it cannot possibly occur before testing.

- E. Sign off on the production rollout plan with platform and security teams: This is also a post-selection, pre-launch governance action. Although it is a necessary step in the productionization process, it is not a prerequisite step for "testing candidate models on a representative sample."

Reference

This question belongs to Product and Model Selection (12% weight). The tested skill is: understanding the sequential logic of the model selection process— define requirements → shortlist candidates → test and validate → deploy, rather than conflating testing, deployment, and maintenance steps.

Explanation

This question tests the core judgment of "Product and Model Selection" domain: matching capability tier to task weight. The principle is—use the lightest tier that meets the quality bar, not the heaviest tier, nor the lightest tier that cannot meet the bar.

Why C is correct:

Option C uses a higher-capability model for complex multi-step legal reasoning under review. Legal reasoning involves multi-step, high-risk, and requires review, which is precisely the core strength of higher-tier models like Opus—deep reasoning, nuanced synthesis, and complex multi-step problem handling. Anthropic's official legal solutions page also confirms that Opus-series models lead in legal benchmarks, particularly in citation accuracy and long-document comprehension. For high-risk tasks where "getting it wrong is costly," investing in capability tier is appropriate.

Why E is correct:

Option E uses a lighter, faster model for routine FAQ-style replies under tight latency budgets. This is a textbook use case for Haiku: high-frequency, well-defined, latency-sensitive, and not requiring deep reasoning. Anthropic officially positions Haiku as "fastest" and suitable for classification, extraction, routing, and other high-capacity, latency-sensitive tasks. FAQ replies have highly standardized content; the lighter model is fully sufficient and avoids wasting time and quota with higher-tier models.

Why the other options are wrong:

- A. Selecting a higher-capability model solely because it is the most expensive: This is "choosing by price" rather than "choosing by need." The most expensive does not mean the most suitable; using Opus for simple tasks only increases latency and consumption without bringing quality gains.

- B. Using a higher-capability model for single-line keyword extraction at high volume: This is a typical over-provisioning. Keyword extraction is a simple, well-defined task that Haiku can handle efficiently. Using a higher-tier model at high volume wastes cost and slows down response time.

- D. Using a lighter, faster model for nuanced multi-document strategic synthesis: This is under-provisioning. Cross-document synthesis and nuanced judgment require deep reasoning and context integration capability, which is precisely the scenario for higher-tier models like Opus. The light model may produce seemingly plausible but actually incomplete or biased synthesis.

Reference

This question belongs to "Product and Model Selection" (12% weight). The tested skill is: based on the three-dimensional trade-offs of task complexity, risk level, and latency requirements, avoiding the two extremes of "worshipping the strongest model" and "cutting costs everywhere," and selecting the tier that truly matches the task.

Explanation

This question tests the generalization of the cost, speed, and quality trade-off in the "Product and Model Selection" domain.

Why B is correct:

Anthropic's official model documentation explicitly lists the positioning of the three tiers: Haiku is "fastest" with the lowest price (starting at $0.10/$0.50 per million tokens); Sonnet is "fast" with moderate price ($2/$10); Opus is "medium" speed with the highest price ($4/$20). In terms of usage limits, Anthropic Academy's model selection guide also ranks the tiers by resource consumption from light to heavy: Haiku lightest, Sonnet moderate, Opus heaviest. Therefore, "stronger capability = more resources + slower speed; lighter = faster + cheaper" is the correct generalization.

Why the other options are wrong:

- A. More capable models consume less usage limit because of more efficient reasoning: Incorrect. More capable models consume more resources, not less. Anthropic's guidance explicitly warns that using Opus for work that lighter models can handle will "consume usage without bringing gains and slow you down".

- C. Lighter models consume more resources because they require multiple passes to achieve the same quality: Incorrect. Lighter models are designed for single-pass, high-throughput scenarios and do not require "multiple passes to catch up with stronger models." This option reverses the actual resource consumption relationship.

- D. Speed and resource consumption are unrelated to capability; the difference lies only in reasoning depth: Incorrect. Speed, price, and capability are positively correlated—Anthropic officially marks Haiku as fastest and cheapest, Opus as slower and more expensive. The difference is not "only in reasoning depth."

Reference

This question belongs to "Product and Model Selection" (12% weight). The tested skill is: accurately describing the actual relationship of the speed-cost-capability triangle across the Haiku/Sonnet/Opus tiers, avoiding myths such as "stronger models are more efficient" or "lighter models secretly consume more."

Explanation

This question tests the core security configuration principle in the Configuration and Knowledge Management domain: when contexts should not be mixed, use independent Projects for isolation.

Why A is correct:

Claude's Projects are designed as self-contained workspaces, each with its own chat history and knowledge base . More importantly, Anthropic's memory feature employs project-scoped memory with strict isolation—memory within each Project is distinct and does not automatically flow into other Projects . When using Team or Enterprise plans, these Project boundaries are explicitly described as "safety guardrails" for managing complex projects while avoiding mixing unrelated details . Therefore, configuring separate memory, knowledge sources, and instructions for each client engagement, and verifying that cross-contamination does not occur, is the configuration that most directly addresses the requirement.

Why the other options are wrong:

- B. Using a single shared Project and assuming Claude will automatically keep the two clients' contexts separate: This is the typical configuration that leads to context bleed. Within the same Project, knowledge and memory are shared across all conversations ; Claude cannot "guess" which content belongs to which client based on the client's identity. Assuming isolation without configuration is a high-risk error.

- C. Sharing the same memory context to save setup effort: This explicitly violates the isolation requirement. Memory sharing means that one client's details may be referenced in another client's conversations, which is precisely the prohibited operation in scenarios involving sensitive client information. Although setup is simplified, the cost is data leakage.

- D. Skipping Project configuration and using regular chats for both engagements: Regular chats do not have scoped memory and knowledge configuration; while chats do not automatically share memory with each other (unless account-level memory is enabled), they lack structured isolation—you cannot attach independent knowledge bases and instructions to each engagement, and you cannot verify isolation at the Project level. This approach does not provide a reliable configuration guarantee, and is therefore not the best choice.

Reference

This question belongs to Configuration and Knowledge Management (12% weight). The tested skill is: using the scoped nature of Projects to achieve context isolation, understanding that memory and knowledge within a Project are limited to that Project, and that cross-client isolation must be ensured through explicit configuration rather than relying on the model's "automatic understanding."

Explanation

This question tests the maintenance and update signals in the "Configuration and Knowledge Management" domain. The core judgment logic is: when a Project's configuration (instructions or knowledge files) becomes outdated, the most direct symptom is that outputs cite content that is no longer current.

Why C is correct:

The quality of a Project's configuration depends on its most outdated file . When a Project's knowledge base contains superseded policies, retired terminology, or obsolete guidance, Claude will retrieve and reference these outdated materials when answering questions, causing outputs to "increasingly reference terminology, sources, or guidance that no longer reflects current practice." This is precisely the signal that the configuration needs immediate updating—you need to remove outdated files, replace them with the current versions, and update instructions that have drifted .

Why the other options are wrong:

- A. The Project's chat volume has steadily declined: Chat volume decline may reflect changes in team usage habits, project lifecycle, or seasonality, and is not a reliable signal that "the configuration needs updating." The configuration itself may still be accurate and effective; a decline in usage does not mean the content is outdated.

- B. Outputs have grown longer on average: Changes in output length are unrelated to configuration timeliness. Longer outputs may be due to changes in the complexity of the questions asked, or adjustments in instructions, and cannot indicate that knowledge or guidance has become outdated.

- D. Members have started using ad-hoc chats outside the Project: If members are bypassing the Project for ad-hoc chats, this may indicate that the Project configuration does not meet their needs (such as insufficient convenience or lack of trust), but this is more likely a usage habit or tool selection issue than a direct signal that the configuration content has become outdated. The most typical and direct signal that configuration needs updating is still "outputs cite outdated content."

Reference

This question belongs to Configuration and Knowledge Management (12% weight). The tested skill is: identifying when a Project's configuration needs maintenance—outdated knowledge is the primary cause of outdated outputs, and the signal to watch for is outputs citing obsolete content, rather than usage volume or output length fluctuations.

Explanation

This question tests the configuration correction capability in the "Configuration and Knowledge Management" domain. The core judgment criterion is: when the output exhibits multiple specific deviations, the fix must be targeted, explicit, and actionable—directly writing each requirement into the instructions, rather than vaguely "letting the model figure it out on its own."

Why C is correct:

The current three problems are: (1) drafts are too brief; (2) goodwill credits are offered in policy-prohibited situations; (3) contractions are used that violate brand guidelines. Option C addresses each one:

- "No contractions" explicitly prohibits specific language features that violate brand voice, rather than vaguely saying "match the brand voice guide";

- "Minimum length" directly addresses the "too brief" problem;

- "Only when the attached eligibility policy permits" transforms the credit decision from "freely offered" to a condition-triggered action, with clear conditional logic.

Research on instructions indicates that specific constraints that "tell Claude what to check and what to refuse" produce inspectable outputs, while vague descriptions cannot be wrong—precisely because they do nothing. Option C is exactly the transformation from vague to specific.

Why the other options are wrong:

- A. Replace "empathetic and casual" with "match the brand voice guide" and remove the credit instruction, letting length be determined by each draft: "Match the brand voice guide" still fails to tell Claude which specific rules to follow (such as no contractions); removing the credit instruction means the model may still improvise, or not offer credits at all when it should. Letting length "be determined by each draft" directly abandons control over the "too brief" problem.

- B. Move policies to the knowledge base and rely solely on documents to govern, without changing custom instructions: Knowledge base documents are reference materials, not behavioral constraints. Simply uploading the brand guide and credit policy does not mean Claude will strictly follow them every time—especially when the instructions still say "casual" and "offer credit." The knowledge base provides information, while the instructions provide rules; the two are not interchangeable.

- D. Keep existing instructions unchanged and just add a sentence telling Claude to consult the attached policies: This is the most likely to mislead. The existing "casual" instruction conflicts with "no contractions"; the existing "offer goodwill credits" conflicts with "only recommend when permitted." Adding a "go consult the documents" sentence creates direct contradictions between instructions and documents, and the model may selectively follow whichever is more convenient. Fixing the source of the conflict (the instructions) is the correct approach.

Reference

This question belongs to Configuration and Knowledge Management (12% weight). The tested skill is: diagnosing the root causes of multiple output deviations and correcting them with explicit, non-conflicting, condition-clear instructions, rather than relying on knowledge base documents or vague references to "fix" behavioral problems.

Explanation

This question tests the sequence of Project configuration in the "Configuration and Knowledge Management" domain. The core principle is: before adding anything to a Project, you must first clarify what it is for and what it needs to accomplish—purpose is the prerequisite for all subsequent configuration decisions.

Why A is correct:

According to Anthropic's official documentation, the first step in creating a Project is to give it a name and description, but Claude itself cannot access this name and description information . This means that the "purpose" must be clearly articulated by the creator before subsequent configuration, so that it can be translated into actual instructions and knowledge selection. Only after clarifying "what recurring tasks this Project supports" can you determine which instructions to write, which reference materials to upload, and who to invite. Without defining the purpose, all subsequent steps lack a criterion for judgment.

Why the other options are wrong:

- B. Test the configured Project with representative requests: Testing is a verification step after configuration is complete, not the first step. Without clarifying the purpose and completing configuration, there is nothing to test. Testing should occur after instructions and knowledge have been added.

- C. Add reference materials and instructions: This is an important configuration step, but it must come after clarifying the purpose. Only after knowing "what the Project is for" can you judge which documents belong in the knowledge base and what behavioral rules the instructions should contain. Anthropic's practice guidelines also confirm that instructions should carry "standing rules," while knowledge carries "reference documents to work from"—the choice of which documents to upload depends entirely on the Project's purpose .

- D. Invite collaborators and confirm permissions: Sharing and permission settings are collaboration-level operations after the Project has been configured. Inviting others to a Project with unclear purpose and incomplete configuration will only expose them to a half-finished workspace.

Reference

This question belongs to Configuration and Knowledge Management (12% weight). The tested skill is: understanding the logical sequence of Project configuration—define purpose → configure instructions and knowledge → test and verify → share and collaborate. Purpose is the prerequisite for all configuration decisions; without a clear purpose, there is no basis for selecting instructions and knowledge.

Explanation

This question examines the collaborative configuration of Skills and Project elements within the "Configuration and Knowledge Management" domain. The core judgment criterion is: Skills provide "how to do it" procedural capabilities, Projects provide "what you know" contextual materials, and the two are complementary and should not be mutually exclusive.

Why B is correct:

According to Anthropic's official explanation, Projects and Skills have clearly differentiated positioning: Projects deliver background knowledge and context (documents, reference materials, persistent instructions), while Skills deliver procedural knowledge and executable capabilities (methods, frameworks, code), loaded on-demand when relevant. The official documentation explicitly recommends using the two together: for example, in a "Product Development" Project, upload product specifications and user research, while combining Skills for writing technical documentation and analyzing user feedback data—the Project tells Claude "what you need to know," and the Skill tells Claude "how to do it".

Option B does exactly this: add the Skill (providing procedural capabilities), upload knowledge sources (providing reference context), and write instructions that reference both—ensuring that Claude both follows the Skill's methodology and bases its work on the Project's specific materials when executing the course design workflow. This is the officially recommended combined approach.

Why the other options are wrong:

- A. Write instructions that contradict the Skill's documented purpose: This creates direct conflicts. When the instructions say one thing and the Skill says another, Claude's behavior becomes unpredictable. Instructions should reinforce and invoke the Skill, not contradict it.

- C. Upload knowledge sources but disable the Skill: This eliminates the packaged procedural capability the workflow needs. Knowledge sources tell Claude "what materials are available," but they cannot replace the Skill's "how to execute the course design process." Disabling the Skill means each conversation must rebuild the methodology from scratch, which is precisely the inefficiency Skills are designed to solve.

- D. Add the Skill but omit knowledge sources and custom instructions: This strips the Project of its contextual foundation. The Skill provides a general methodology, but the specific course materials, brand standards, and workflow rules for course design exist in the knowledge sources and instructions. Missing these, the Skill can only execute "generic course design," not the "course design for this specific recurring workflow."

Reference

This question belongs to Configuration and Knowledge Management (12% weight). The tested skill is: understanding the complementary relationship between Skills and Project elements—Skills = procedural capability, Project = contextual knowledge, with instructions serving as the link that invokes both, rather than viewing them as substitutes or competitors.