CCA-F 练习题库· Prompt Engineering & Structured Output
A developer generates a 300-line module using Claude Code in one session, then asks the same session to review the code for bugs. The review returns "looks good" with minor style suggestions but misses two logic errors that a colleague catches in manual review. What is the most likely cause, and what architectural change addresses it?
- AThe context window was too large; the fix is to limit code generation to smaller chunks so the review has less to process.
- BThe review prompt was too vague; adding more explicit review criteria to the same session would catch the missed errors.
- CThe reviewing instance retains reasoning context from generation, making it less likely to question its own decisions. Use a second independent Claude instance without the generation context to perform the review.
- D生成时用的模型 tier 比处理评审请求的那个更强，所以评审方根本没有足够的推理深度去发现那两处逻辑错误；把两次调用都固定到同一个更大的 tier 就能抹平这个质量差。
正确答案：C
解析
The most likely cause is self-review bias: because the reviewing instance shares the same session and reasoning context that generated the module, it is anchored to its own prior decisions and unlikely to question them, so it misses logic errors a fresh human reviewer catches.
The architectural fix is to run the review in a second, independent Claude instance (or subagent) that has no generation context, giving an unbiased pass over the code — this is the documented separate-verifier / clean-context pattern.
- Amisattributes the failure to context-window size (300 lines is trivial and chunking doesn't remove the blindspot).
- Bclearer criteria may help some categories but leaves the self-consistency bias intact within the same session.
- Dinvents a model-tier mismatch not present in the scenario.
延伸阅读
本题为本站自有原创练习题，非任何官方考试内容；解析对照 Anthropic 公开文档撰写， 如与最新文档不符请以官方为准。想在限时环境下检验水平，可参加 模拟考（题目与本页题库不重叠）。