CCAR-F Exam Guide & Practice Questions

A retailer is building an autonomous Claude workflow to handle refund requests. The workflow must look up the order, check refund eligibility, and issue a refund only when the policy allows it. Requests above a set amount require a manager approval before any refund is issued. The team wants the workflow to notify the customer after the decision.

Which orchestration design best meets the requirements?

• ALet Claude repeat the entire workflow in an unrestricted loop until it decides the refund is complete.
• BUse an orchestrator with an eligibility check, a branch to request approval when required, and a refund step only after the required checks pass.
• CRun the eligibility check and refund step concurrently, then decide whether approval was needed.
• DSend every request through the same path, including approval, eligibility checks, and refund, regardless of the order details.

A publisher uses Claude to prepare a policy-change impact brief. Each brief needs separate assessments from editorial, legal, and customer support. The assessments can begin independently, and the service can run up to three at once. The final brief must reconcile their findings. If legal identifies a regulated claim, editorial must revise its assessment before synthesis, and the system must not publish the brief automatically.

Choose two orchestration choices that best meet these requirements.

Select 2 answers, then check them.

• ARun the editorial, legal, and customer-support assessments in parallel, within the service limit, because they are independent at the start.
• BBegin synthesis as soon as the first assessment finishes, then add later assessments to the completed brief.
• CIf legal identifies a regulated claim, have editorial revise its assessment before the final synthesis.
• DAllow each assessment to publish its own brief so the final synthesis step is unnecessary.
• ERun all three initial assessments sequentially, even though they do not depend on one another.

A support platform lets Claude update a customer's notification preferences. The existing tool accepts a free-form string called change and an account name, and operators report that Claude sometimes sends an unsupported preference value or omits which channel is being changed. The backend accepts only email, SMS, or push, and every update must identify a stable customer ID.

Which redesign is the best choice for making this tool call clear and valid?

• ADefine a tool with required customer_id, channel, and enabled fields, and constrain channel to email, SMS, or push.
• BKeep the free-form change field, but add a description listing examples of valid preference changes.
• CAccept customer name and an optional channel, then let the backend infer the missing customer ID and channel.
• DCreate a separate tool for each customer, with a free-form string for the requested preference.

An application team connects Claude to an internal case-management MCP server. The server exposes tools that may change between deployments, and the application currently hard-codes a tool name while forwarding Claude's requested arguments. During testing, results occasionally appear as ordinary assistant text, and Claude then repeats the same request. The team wants the bridge to follow MCP and Anthropic tool-use semantics.

Choose two. Which actions should the application team take to correct the MCP integration?

Select 2 answers, then check them.

• AUse MCP tools/list to discover available tools, and respond to tool-list change notifications so the exposed definitions can be refreshed.
• BWhen Claude returns a tool_use block, invoke the corresponding MCP tool with tools/call, then return a tool_result that uses the same tool_use ID.
• CConvert each MCP tool result into ordinary assistant text and end the turn, since Claude can infer that the operation completed.
• DKeep the hard-coded tool name and skip tool discovery, because MCP tool names are guaranteed to remain stable across deployments.
• ESend Claude's arguments directly to the MCP server as raw JSON, without using the MCP tools/call request.

A software team is adopting Claude Code in a repository used by developers on several machines. The team needs shared project settings under version control, while allowing each developer to keep personal overrides that are not committed. The repository already has a .claude directory, and the team wants to avoid relying on each developer to copy settings manually.

A software team is adopting Claude Code in a repository used by developers on several machines. The team needs shared project settings under version control, while allowing each developer to keep personal overrides that are not committed. Which arrangement best meets these requirements?

• ACommit .claude/settings.json for shared project settings and keep personal overrides in .claude/settings.local.json, which is excluded from version control.
• BStore the shared settings in each developer's ~/.claude/settings.json and leave the repository without Claude Code settings.
• CCommit .claude/settings.local.json as the shared settings file and use .claude/settings.json for personal overrides.
• DPut all settings in CLAUDE.md because Claude Code reads that file as its project settings file.

A development group is standardizing Claude Code use in a repository. Developers need the same repository-specific build and test guidance, along with the same project-level Claude Code settings. Personal configuration must remain separate, and the team wants new contributors to get the shared setup by checking out the repository. The group has decided not to depend on instructions stored only on individual workstations.

A development group is standardizing Claude Code use in a repository. Developers need the same repository-specific build and test guidance, along with the same project-level Claude Code settings. Personal configuration must remain separate, and the team wants new contributors to get the shared setup by checking out the repository. Choose two.

Select 2 answers, then check them.

• AAdd repository-specific build and test guidance to a root-level CLAUDE.md file and track it in version control.
• BPlace project-level Claude Code settings in .claude/settings.json and track that file in version control.
• CPut the repository guidance only in each developer's ~/.claude/CLAUDE.md file.
• DStore the shared project settings in .claude/settings.local.json and commit that file.
• EAdd the guidance only to each developer's shell profile so it is available in the terminal.
• FKeep a separate copy of the guidance outside the repository and ask each contributor to maintain it.

A support team uses Claude to extract a customer name, order number, and issue summary from incoming messages. The current prompt asks for a concise answer, but the application fails when Claude adds commentary, omits a key, or writes a phrase such as unknown where the application expects a null value. The team wants to keep the original message available as the only source for extraction.

Which prompt revision is most likely to produce consistent, machine-readable results?

• AAsk Claude to be concise and return valid JSON, but leave the fields and treatment of missing information unspecified.
• BState the task, place the source text inside clear delimiters, specify each required key and value type, define missing values as null, and require JSON with no surrounding text.
• CAsk Claude to explain its reasoning before returning JSON, and place the source text and instructions together in one paragraph.
• DProvide one example of the desired response and ask Claude to imitate it, without specifying required fields or how to handle missing information.

A claims operations team asks Claude to assign each report to one of three existing categories, delayed payment, missing document, or disputed charge. The current prompt lists the names but gives no definitions, and reports involving both a late payment and a missing invoice receive different labels across runs. The downstream report accepts only those three categories and cannot process explanatory prose in place of a label.

Choose two prompt changes that are most likely to reduce inconsistent category assignments.

Select 2 answers, then check them.

• ADefine what qualifies for each category, including a rule for cases that appear to fit more than one category.
• BAsk Claude to use its general judgment because rigid category definitions may limit its interpretation.
• CAdd labeled examples that include borderline cases and show the expected category for each one.
• DRemove the category list and ask Claude to create the most suitable category for each report.
• ERequire a detailed free-form analysis before Claude selects a category, without specifying how to format the final result.

A legal operations team uses Claude to review a large set of vendor agreements over several hours. The application sends the accumulated conversation with each request, and the team is nearing the model’s context limit. Reviewers must retain the latest scope restrictions, agreed definitions, and open issues, while still being able to check claims against the original agreements.

Which change best addresses the risk that important requirements will be lost as the conversation approaches Claude’s context limit?

• APeriodically replace older turns with a concise checkpoint containing active requirements, decisions, unresolved questions, and references to supporting source material.
• BContinue sending the full conversation until Claude can no longer accept the request, then start a new conversation with only the latest user message.
• CKeep every assistant response in the request, but remove the original source documents because the responses already summarize them.
• DAsk Claude to remember the complete conversation for the next request without including any prior messages or stored state.
• EIncrease the sampling temperature so Claude is more likely to recall details that appeared earlier.

A product support team uses Claude across separate API requests to help agents investigate complex customer cases. Case facts can change after a specialist confirms a correction, and long transcripts sometimes contain earlier, conflicting values. The team has limited context budget and needs agents to receive consistent guidance over an extended investigation without treating an old transcript statement as current.

A. Maintain a versioned, authoritative project-state record and include the current relevant portion in each request. B. Test the application on representative long-session cases and compare its outputs with reviewed expected results. Choose two.

Select 2 answers, then check them.

• AMaintain a versioned, authoritative project-state record and include the current relevant portion in each request.
• BTest the application on representative long-session cases and compare its outputs with reviewed expected results.
• CKeep all past messages in every request, even when doing so exceeds the team’s context budget.
• DRely on Claude to retain facts from earlier API requests without resending them or storing them elsewhere.
• EResolve conflicting project facts by selecting whichever value appears most often in the conversation.
• FUse a higher temperature for later requests so Claude can recover details that have become inconsistent.