Question 260 An AI assistant recommends deleting a production customer table and rates its own confidence at 98%. What should govern execution? Explicit authorized human approval and a validated recovery plan for the destructive operation. The presence of valid SQL syntax. The confidence percentage alone. A second model merely repeating the recommendation. Explanation Correct Answer A:Self-reported confidence cannot substitute for change authority. Incorrect Answers B:A valid command can still destroy valuable data. C:Model certainty is not business permission. D:Agreement is not an approval by the owner.
Question 261 A research coordinator has specialist web, code, and data tools. The current task only involves a provided spreadsheet. Which optimization is best? Expose spreadsheet analysis tools and avoid unrelated web or code mutation capabilities. Remove spreadsheet access and ask the model to guess its rows. Run web search and deployment tools before opening the spreadsheet. Grant administrative database access for completeness. Explanation Correct Answer A:Smaller relevant tool sets improve routing and privilege separation. Incorrect Answers B:Without the file the agent lacks the requested evidence. C:Extra unrelated tools add latency and risk. D:The request does not require database mutation.
Question 262 A team migrates from a deprecated structured-output request field to a newer Claude API configuration. What is the safest approach? Replace structured output with arbitrary markdown everywhere. Use the current documented output_config.format interface and regression-test schema behavior. Assume the output format setting cannot change across versions. Keep an old request shape indefinitely without checking failures. Explanation Correct Answer B:Version-aware migration avoids relying on deprecated API fields. Incorrect Answers A:That regresses machine-readable reliability. C:API evolution requires documented verification. D:Deprecation can lead to rejected requests.
Question 263 A research system divides a global policy change into many country reports. Which TWO design controls make decomposition effective? Let workers select overlapping countries with no shared assignment list. Specify required geographic and topical coverage before assigning workers. Plan a final pass to reconcile cross-border conflicts and terminology. Use the number of output pages as the only completeness metric. Ask the final writer to invent a result for each missing country. Explanation Correct Answers B:A defined coverage matrix prevents regions or themes from being silently omitted. C:Local country conclusions may conflict and require integrated adjudication. Incorrect Answers A:Uncoordinated work can produce duplicates and leave gaps. D:Length cannot prove required jurisdictional coverage. E:No amount of synthesis replaces the missing evidence.
Question 264 A new engineer clones a project and asks Claude Code to run tests, but the test command is absent from memory. Where should the team record the canonical command? In project guidance such as CLAUDE.md with any required working-directory notes. In a shell alias that runs only on the author’s machine. In a screenshot inside an unrelated marketing folder. Only in the last developer’s private chat session. Explanation Correct Answer A:Stable build instructions prevent guesses and are available to everyone. Incorrect Answers B:Other users cannot depend on an unshared local alias. C:The command is not reliably discoverable from there. D:A cloned repository does not inherit that private history.
Question 265 A lookup_customer tool description says “Returns useful customer data.” The model confuses email and customer ID inputs. What improvement most directly addresses this? Specify accepted identifier types, validation rules, and which field is authoritative. Make the response always say customer found. Increase model creativity. Remove the input schema entirely. Explanation Correct Answer A:Precise input semantics prevent wrong-customer lookups and tool misrouting. Incorrect Answers B:False success cannot correct mismatched identifiers. C:Sampling does not define which identifier the tool expects. D:Without a schema the ambiguity becomes worse.
Question 266 An assistant is examining a recurring failure in a large test log. What should it keep in the working context? An invented trace inferred from the test name. Only the final word FAILED without context. Every successful test log line and color escape sequence. The failing test identifiers, first meaningful error, relevant stack frames and corresponding source locations. Explanation Correct Answer D:Condensed evidence retains the actionable failure path. Incorrect Answers A:The actual stack and error must ground the diagnosis. B:That lacks the cause and location. C:Voluminous success output obscures the failure.
Question 267 A source-comparison assistant treats a 2% change as significant despite a requirement to flag only changes above 10%. What is the prompt defect? All percentages should be treated as equivalent. The model needs a more creative writing style. A change report should never compare values. The numeric materiality threshold is not explicit or correctly applied. Explanation Correct Answer D:Stating and enforcing the threshold aligns findings with business criteria. Incorrect Answers A:That removes the materiality distinction. B:Style does not fix threshold logic. C:The task specifically requires numeric comparisons.
Question 268 A retail tool returns prices in cents but the agent reports them as dollars. Which integration mechanism should normalize this before model reasoning? A PostToolUse transformation that converts the tool result into a documented unit. A final HTML formatter that changes only the display currency symbol. A Stop hook that deletes all prices. A prompt telling the model to assume all values use dollars. Explanation Correct Answer A:A tool-result hook can standardize heterogeneous outputs before they enter subsequent reasoning. Incorrect Answers B:Formatting after reasoning cannot correct an earlier 100x arithmetic error. C:Removing values does not normalize or preserve the required cost evidence. D:The actual tool contract returns cents, so that instruction causes errors.
Question 269 A team has a rarely used performance-test checklist and a universal “run unit tests” policy. Where should each live? Store neither and ask the model to guess current CI requirements. Load the full performance checklist for every trivial file read. Put stable universal policy in project memory and the specialized checklist in a skill. Remove the universal test policy and hope it is remembered. Explanation Correct Answer C:Different frequency and scope favor different instruction mechanisms. Incorrect Answers A:Local engineering policy cannot be safely inferred. B:Unnecessary context slows ordinary tasks. D:Required common behavior should be explicit.
Question 270 An agent is fixing an API authorization regression in a monorepo. Which TWO context choices are most useful? Read the endpoint’s route, middleware and service path rather than every package. Ignore repository-specific conventions and tests. Inspect the relevant permission tests and recent diff for the failing behavior. Use a similarly named file from an unrelated project. Load every generated dependency into the prompt. Explanation Correct Answers A:The targeted call chain exposes enforcement boundaries. C:Tests and changed lines ground a reproducible fix. Incorrect Answers B:That risks incompatible edits. D:Cross-project context is not evidence for this repository. E:That consumes context with low-value data.
