CCA-F 시험 - Anthropic실제시험문제와 답 - 320문항

Question No : 1

You are implementing a validation-retry loop for a structured data extraction prompt. The extraction fails a business logic validation (e.g., line items do not sum to the total).
Which techniques represent best practices for the follow-up retry prompt? Choose 2 correct answers.

A.Return a generic error message like There were errors in your extraction. Please try again.
B.Append specific validation errors, detailing which field failed and the nature of the discrepancy.
C.Include the original document and the failed extraction in the retry prompt to guide self-correction.
D.Increase the model temperature to 1.0 so it thinks of a different answer.

정답:
Explanation:
A successful validation-retry loop requires appending specific error details (not generic messages) to the prompt and supplying both the source document and the failed extraction. This provides the exact context the model needs to correct its semantic mistake.

Question No : 2

When configuring a JSON extraction schema for document categorization where new or unexpected document types might appear, what is the recommended design pattern?

A.Use an open-ended string field with no constraints to maximize flexibility.
B.Use a rigid Enum containing only the currently known document types.
C.Include an other enum value paired with a document_type_detail string field for extensible categorization.
D.Use a PreToolUse hook to intercept and reject unknown document types.

정답:
Explanation:
To handle ambiguous or unexpected data during structured extraction, schemas should include an 'other' (or 'unclear') enum value paired with a detail string field. This prevents the model from forcing unexpected data into incorrect predefined categories.

Question No : 3

You need Claude to output raw JSON directly to the user interface of your application, without any conversational preamble or markdown formatting.
Which prompt engineering technique achieves this?

A.Add "DO NOT USE MARKDOWN" to the end of the user prompt.
B.Combine Assistant Message Prefilling ```json with and use ```as a stop sequence.
C.Set max_tokens to exactly match the length of the JSON string.
D.Provide a 10-shot example list of raw JSON outputs.

정답:
Explanation:
Assistant Message Prefilling with ```json tricks the model into thinking it has already started a code block. Providing sequence immediately terminates generation once the JSON payload completes, resulting in raw JSON without explanatory text.

Question No : 4

To prevent Claude from fabricating values when a requested field is absent from the source document, how should you engineer your JSON extraction schema?

A.Design the field with "type”: [" string, “null"] and mark it as optional/nullable.
B.Add a system prompt instruction to "never fabricate data".
C.Set the tool parameter tool_choice: " any".
D.Reduce the temperature to 0.0 to eliminate hallucinations.

정답:
Explanation:
If a field is strictly required but the data doesn't exist, the model may fabricate a value to satisfy the schema constraint. Making the field optional or nullable (e.g., allowing 'null') provides a valid escape hatch so the model can comply with the schema without hallucinating.

Question No : 5

You are designing a prompt to extract structured data from unstructured documents. You decide to use tool_use with a strict JSON schem a.
What specific guarantee does this approach provide?

A.It guarantees both structural compliance and the semantic correctness of the extracted data.
B.It eliminates JSON syntax errors and guarantees structural compliance, but does not prevent semantic errors.
C.It guarantees the model will not hallucinate missing information.
D.It eliminates the need for validation-retry loops.

정답:
Explanation:
Using tool_use with JSON schemas guarantees structural compliance (all required fields present, valid types, syntax correct) but does NOT guarantee semantic correctness (the actual values may still be inaccurate or hallucinated). You still need semantic validation rules.

Question No : 6

When splitting a generic MCP tool (e.g., analyze_document) into purpose-specific tools, which of the following represent improved tool design practices? Choose 2 correct answers.

A.Renaming the tools to have completely identical descriptions to let the LLM randomly balance the load.
B.Creating specific tools with defined input/output contracts like and sununarize_content.
C.Eliminating functional overlap by ensuring descriptions differentiate each tool's exact purpose.
D.Combining all extraction and summarization functions into a single tool with 25 optional parameters.

정답:
Explanation:
Purpose-specific tools with clear input/output contracts and distinct descriptions prevent selection ambiguity. Using generic tools or packing 25 parameters into one tool heavily degrades the model's reliability in selecting and utilizing the tool properly.

Question No : 7

Your MCP server exposes a Templated Resource with the URI pattern docs://documents/{doc_id}.
How does the MCP Python SDK handle this pattern?

A.It requires the client to pass the doc_id via an HTTP POST body.
B.It automatically parses the doc_id from the URI and passes it as a keyword argument to the underlying resource function.
C.It treats the entire URI as a static string and expects the function to manually parse it.
D.It requires the model to write a regex to extract the doc_id before passing it to the server.

정답:
Explanation:
For templated resources, the MCP Python SDK automatically parses parameters from the URI structure (like '{doc_id}') and passes them directly as keyword arguments to your Python function, avoiding manual string parsing.

Question No : 8

A backend MCP tool encounters a timeout while querying a database.
Which structured error response design is the correct approach to propagate this to the agent?

A.Return a generic {"error": "Operation failed"} to secure internal infrastructure details.
B.Return a valid empty JSON array [ ] to avoid crashing the agentic loop.
C.Return structured error metadata including an isError: true flag, the error category, and an isRetryable’: true boolean.
D.Terminate the MCP server process completely.

정답:
Explanation:
Proper error propagation in MCP requires returning structured error metadata with 'isError: true', the error category (e.g., timeout), and an 'isRetryable' boolean. Returning generic errors hides context, and silently returning empty arrays is a critical anti-pattern because the agent will hallucinate that no data exists.

Question No : 9

You are designing an MCP tool to look up customer orders.
Which approach to tool descriptions is recommended to ensure reliable tool selection by Claude?

A.Keep the description under 10 words to save context window tokens.
B.Include the input format, example queries, edge cases, and boundaries explaining when to use it versus similar tools.
C.Use the description solely to list the required JSON properties.
D.Leave the description blank and rely on Claude's semantic understanding of the tool's name.

정답:
Explanation:
Tool descriptions act as the primary mechanism LLMs use for tool selection. A well-designed tool description must be detailed and include input formats, example queries, edge cases, and explicit boundary instructions explaining when to use it versus similar tools to prevent misrouting.

Question No : 10

You are designing an MCP interface for a backend system.
When should you choose to implement an MCP Resource instead of an MCP Tool?

A.When you need the model to execute side-effects, such as updating a database record.
B.When you want to expose a static content catalog or documentation that can be fetched to provide context without requiring exploratory tool calls.
C.When you need to provide pre-built conversation templates.
D.When you need to execute bash commands on the local machine.

정답:
Explanation:
MCP Resources are used to expose content catalogs (like documentation, static data, or schemas) allowing clients to fetch data directly to inject into prompts. Tools are used when you want the model to perform actions, execute logic, or mutate state.

Question No : 11

According to best practices, which of the following scenarios should trigger a developer to use Plan Mode in Claude Code? Choose 2 correct answers.

A.When implementing multi-file architectural changes.
B.When adding a single console. log statement to a known file.
C.When evaluating multiple valid approaches for a new feature implementation.
D.When fixing a minor typo in project documentation.

정답:
Explanation:
Plan mode is specifically designed for complex tasks that involve multiple valid approaches, architectural decisions, and multi-file modifications. Simple, well-defined fixes like fixing typos or adding single log statements should use Direct Execution to avoid unnecessary overhead.

Question No : 12

While working in Claude Code, you press Shift + Tab to enter Plan Mode.
Which subagent is typically associated with isolating the verbose discovery phase to prevent context window exhaustion during planning?

A.The Review subagent
B.The Write subagent
C.The Explore subagent
D.The Plan-Executor subagent

정답:
Explanation:
The Explore subagent is used for isolating verbose discovery output and returning focused summaries to preserve the main conversation's context window during the research and planning phase.

Question No : 13

Which of the following tasks is the MOST appropriate use case for Direct Execution rather than Plan Mode?

A.Changing the state management library from Redux to Zu stand across a large React application.
B.Designing an implementation approach for a new API integration with unknown dependencies.
C.Fixing a specific bug with a clear stack trace in a single TypeScript file.
D.Restructuring the database schema and updating all corresponding repository models.

정답:
Explanation:
Direct execution is appropriate for simple, well-scoped changes such as a single-file bug fix with a clear stack trace. Plan mode adds unnecessary overhead for simple fixes, but is essential for complex architectural changes or multi-file migrations.

Question No : 14

A developer needs to perform a monolith-to-microservices migration spanning 45+ files.
Why is Plan Mode recommended over Direct Execution for this task?

A.It uses the Message Batches API to reduce token costs by 50 percent during execution.
B.It enables safe codebase exploration and design to determine architectural approaches before making costly file modifications.
C.It automatically forks the session into a persistent subagent.
D.It bypasses the token limit context window by executing across multiple continuous sessions automatically.

정답:
Explanation:
Plan mode is designed for complex tasks involving large-scale changes, multiple valid approaches, and architectural decisions. It enables safe codebase exploration and design before committing to changes, which prevents costly rework compared to Direct Execution.

Question No : 15

What is the primary difference in tool usage between Plan Mode and Direct Execution in Claude Code?

A.Plan Mode uses Write and Edit heavily to scaffold boilerplate code.
B.Plan Mode restricts Claude to using only read-only tools to analyze the codebase before executing.
C.Plan Mode completely disables all tool usage until the user explicitly permits it.
D.Plan Mode automatically switches to a higher-tier model with larger tool capabilities.

정답:
Explanation:
Plan mode restricts Claude to using only read-only tools to explore the codebase, map dependencies, and design an implementation strategy before committing to costly file modifications. Direct execution allows Claude to use modification tools like Write and Edit immediately.

Anthropic CCA-F 시험