Anthropic CCAR-F 問題集 練習問題・模擬試験｜日本語版＆英語版対応

サンプル問題 1： When designing an automated code review system for a large repository, you notice that developers are experiencing 'alert fatigue' because the AI flags too many non-issues and stylistic preferences. What is the most effective prompt engineering technique to reduce these false positives?

• A: Instruct the model to be thorough and flag everything suspicious to ensure no bugs are missed.
• B: Define explicit, measurable review criteria, such as flag functions exceeding 50 lines of code, instead of relying on vague instructions.
• C: Set the model's temperature to 0.0 to ensure deterministic output and eliminate all false positives.
• D: Provide at least 15 few-shot examples covering every possible coding style in the repository.

サンプル問題 2： You are building a classifier to categorize customer reviews, but the model struggles with ambiguous scenarios like sarcasm and mixed sentiments. You decide to use few-shot prompting. What is the recommended best practice for this scenario?

• A: Provide exactly 1 example of a standard positive review to establish the JSON format.
• B: Provide 8 to 10 examples to ensure every possible sentiment category is covered comprehensively.
• C: Provide 2 to 4 examples, ensuring that at least one example specifically addresses an ambiguous edge case like sarcasm.
• D: Avoid few-shot prompting and instead use a detailed system prompt explaining the exact definition of sarcasm.

サンプル問題 3： A pull request modifies 14 files across a stock tracking module. When analyzing all files in a single pass, Claude provides superficial comments, misses obvious bugs, and produces contradictory feedback. How should you redesign the multi-pass review architecture to resolve this?

• A: Switch to a higher-tier model with a larger context window so all 14 files receive adequate attention in one pass.
• B: Require developers to manually split large PRs into smaller submissions of 3-4 files before triggering the automated review.
• C: Run three independent single-pass reviews on the full PR and only flag issues that achieve consensus across at least two runs.
• D: Split the review into focused passes: analyze each file individually for local issues, followed by a separate integration-focused pass for cross-file data flow.