Question 230 Two MCP tools both say “send message” but one posts to an internal channel and the other emails external customers. How should their interfaces change? Let the model infer channel from the recipient’s first name. Name the destination, audience, side effects, and authorization requirements clearly. Remove required recipient fields from both tools. Describe them as interchangeable notification services. Explanation Correct Answer B:External communication and internal updates have different risks and approval needs. Incorrect Answers A:A name alone does not specify a safe communication destination. C:Underspecified targets can send content to the wrong audience. D:Destination differences are material to tool choice.
Question 231 A customer asks for a price comparison but provides no region; the catalog lists public price bands for several regions. What is a reasonable partial approach? State the region assumption explicitly, show comparable public bands and invite correction if it matters. Present one region’s price as universally applicable. Complete a purchase immediately in the cheapest region. Invent a precise local tax rate and total. Explanation Correct Answer A:A reversible comparison can proceed transparently without blocking all progress. Incorrect Answers B:Regional variation is material to accuracy. C:Comparing prices does not authorize buying. D:Unknown jurisdiction details should not be manufactured.
Question 232 A team discovers its root CLAUDE.md contains production secrets alongside coding standards. What is the correct remedy? Hide the secret below a markdown comment. Remove and rotate exposed secrets, then reference secure secret provisioning from the guidance. Ignore the issue because the file is only read by Claude. Add more secret values for every environment. Explanation Correct Answer B:Version-controlled instructions should not embed live credentials. Incorrect Answers A:Comments do not prevent repository exposure. C:Repository files can be accessed and copied by other actors. D:That multiplies the exposure.
Question 233 A tool returns an HTTP 403 for a data source because the service account lacks permission. What is the correct handling? Use a different customer’s credential without approval. Retry indefinitely with the same credential. Generate plausible source content to continue. Surface the authorization failure and request the appropriate access path rather than retrying unchanged. Explanation Correct Answer D:Permission denial is not a transient content-generation issue. Incorrect Answers A:That violates isolation and authorization. B:The same denied permission will not resolve through repetition. C:Invented data conceal an access failure.
Question 234 An engineer wants a personal MCP server that should be available across several private projects. What configuration scope fits? A source file with hard-coded access tokens in each repo. A supported user-level Claude Code MCP configuration, outside each repository. A conversation instruction that the server exists without any connection. A checked-in project configuration in every unrelated repository. Explanation Correct Answer B:User-level configuration follows the individual across projects without changing team defaults. Incorrect Answers A:Embedding credentials in code is unsafe. C:Prompts cannot create a functioning MCP transport by themselves. D:Duplicating personal settings creates maintenance and secret-exposure risk.
Question 235 A tool returns a transient 503 while an agent fetches public product descriptions. Which response best balances continuity and control? Retry a bounded number of times with backoff and then surface an explicit failure. Immediately call an unrelated write tool to make progress. Delete the error and claim the description was retrieved. Retry forever with no timeout or cap. Explanation Correct Answer A:Transient errors can recover, but unlimited retries can trap the workflow. Incorrect Answers B:An unrelated side effect neither repairs retrieval nor respects the task. C:That fabricates evidence needed for the user’s decision. D:Persistent outages would consume resources without a decision boundary.
Question 236 A Claude Code CI step reports a model response as PASS even though the process exited nonzero. How should the pipeline judge success? Check process status and validate required structured output before marking the job successful. Search for the word success anywhere in stdout. Mark every run successful to avoid blocking deployment. Ignore exit codes for all LLM-driven jobs. Explanation Correct Answer A:Textual optimism cannot override a failed command or malformed result. Incorrect Answers B:The text might quote a failure or be incomplete. C:False green checks undermine CI. D:Execution failures remain meaningful.
Question 237 A reviewer corrects the model’s classification of a sensitive document. What feedback should be preserved? The corrected label, evidence, applicable rule and affected model output version. The human reviewer’s identity without the correction. Only that a human clicked disagree. An invented rule that all similar-looking documents are public. Explanation Correct Answer A:Specific feedback supports consistent future checks and traceability. Incorrect Answers B:Identity alone does not capture the correct judgment. C:A bare vote lacks the rationale. D:Overgeneralization can introduce new classification failures.
Question 238 A structured extractor keeps reversing payer and payee roles in contracts. Which few-shot design best addresses this? Include pairs with different sentence orders and explicit payer/payee field outputs. Replace every party name with the same token. Use only cases where payer always appears first. Show sample output JSON with both fields blank. Explanation Correct Answer A:Varied role-order examples teach the relationship rather than a word-position shortcut. Incorrect Answers B:Identical tokens remove the distinction being taught. C:The extractor may learn an incorrect positional heuristic. D:Empty values do not demonstrate role assignment.
Question 239 A user wants to continue a long audit from the last verified checkpoint, but the assistant keeps restarting research. What should session management enforce? Ask the user to restate every past decision. Re-run every earlier retrieval regardless of staleness. Ignore completed IDs and allocate new IDs on every run. Restore the explicit completion ledger and scope before selecting next actions. Explanation Correct Answer D:Resumption should exclude already completed research and preserve owner constraints. Incorrect Answers A:A valid checkpoint should minimize repeated user input. B:Unnecessary repetition wastes time and can change facts without reason. C:That breaks traceability and creates duplicate work.
Question 240 A pipeline requires an approval before a generated change can become a release. Where should that gate live? In a local developer note inaccessible to CI. In a message asking the model to be careful. Only after production has been deployed. In CI/CD branch protection or deployment controls, not only in an assistant prompt. Explanation Correct Answer D:A real release permission must be enforced by the automation platform. Incorrect Answers A:The platform cannot enforce a rule it never sees. B:Prompt guidance cannot guarantee release authorization. C:Late approval does not prevent an unauthorized release.
