Question 300
A prompt gives two examples that contradict a new instruction about refusing unsupported claims. The model keeps copying the examples. What is the most targeted correction?
- Add a stronger adjective to the instruction while leaving the conflicting examples unchanged.
- Move the conflicting examples into a longer user message without changing their outcomes.
- Revise the few-shot examples to demonstrate the new evidence and abstention policy consistently.
- Make the example answers longer so the model has more tokens to copy.
Explanation
Correct Answer
C:Examples condition the model’s expected answer pattern as well as prose instructions. Contradictory demonstrations undermine the intended rule; align both positive and negative cases with the current acceptance criteria.
Incorrect Answers
A:Rhetorical emphasis does not remove the contradictory training signal in the supplied context. The model may still imitate examples that reward unsupported conclusions.
B:Relocation alone does not resolve conflicting behavior specifications. The assistant still observes demonstrations of unsupported claims and may reproduce that failure mode.
D:Extra verbosity can reinforce the wrong pattern rather than correct it. The issue is inconsistent examples and decision criteria, not insufficient example length.
Question 301
A request asks for a complex report but leaves audience and decision criteria unspecified. What should the prompt engineer do?
- Supply the intended reader, purpose, evidence standard, and constraints.
- Specify which constraints have priority and which outputs are prohibited.
- As an alternative, explicitly request only the schema-defined fields and pair this with runtime enforcement.
- Assess each revision against a fixed regression suite and documented failures.
Explanation
Correct Answer
A:Context about the report's use guides which details matter and avoids generic content that misses the business goal.
Incorrect Answers
B:No wording trick makes incompatible goals jointly satisfiable; a coherent instruction hierarchy is prerequisite to evaluation; when the goal is task audience specification, optimizing contradictory prompt constraints misses that operational boundary.
C:Format instructions improve compliance, but production parsers still need structured-output controls or validation; the task audience specification criterion differs from output-only prompt guidance; the gap persists.
D:Prompt improvements should be measured on representative examples rather than judged by a single demonstration; because iterative prompt regression testing is its concern, task audience specification remains outside its effect here.
Question 302
An assistant must answer in valid fields but sometimes adds a leading sentence. Which prompt-only change is helpful but not sufficient?
- Explicitly request only the schema-defined fields and pair this with runtime enforcement.
- Assign exact arithmetic to code and verify the result independently.
- Document the system's operating constraints and available implementation choices.
- Document how absent values must be represented in the required schema.
Explanation
Correct Answer
A:Format instructions improve compliance, but production parsers still need structured-output controls or validation.
Incorrect Answers
B:Prompt wording can affect reasoning but cannot replace deterministic computation or validation for precise totals; this implements correct arithmetic enforcement rather than the demanded output-only prompt guidance.
C:A recommendation is discriminating only when the model knows the architecture and criteria it must optimize; where output-only prompt guidance matters, a control for operational context in prompting leaves that test unmet.
D:Distinguishing missing data from an omitted key supports consistent downstream parsing and auditability; the consequence is improved missing-field representation, not verified output-only prompt guidance.
Question 303
A model produces generic advice when the task requires a precise engineering recommendation. What role of context should be improved?
- Describe the system's operating constraints and available implementation choices.
- As an alternative, limit response scope, request evidence-backed claims, and allow abstention on missing facts.
- Require a generic yes-or-no answer without describing deployment constraints.
- Provide only the desired response style and leave implementation requirements implicit.
Explanation
Correct Answer
A:A recommendation is discriminating only when the model knows the architecture and criteria it must optimize. Supply the current stack, nonnegotiable constraints, traffic characteristics, and criteria by which the recommendation will be judged.
Incorrect Answers
B:Clear scope and uncertainty policy discourage unsupported elaboration without pretending to guarantee truth; this may help concise grounded response policy but cannot by itself deliver operational context in prompting.
C:A binary output format cannot create missing facts about the architecture or its operating environment. Useful engineering guidance depends on specifying relevant constraints and alternatives.
D:Stylistic formatting can improve consistency, but it does not convey system limitations or competing technical options. The model would still lack what it needs to recommend a credible design.
Question 304
A multilingual support assistant must answer a Spanish-speaking buyer, preserve registered product names, and quote prices exactly as stated in USD rather than inventing conversions. Which TWO prompt requirements directly address those separate risks? Choose TWO.
- Require original currency codes and amounts to be preserved unless an authorized rate source and conversion request are supplied.
- Include an approved terminology glossary and explicitly keep protected product names untranslated.
- Silently round every USD price into a presumed regional currency to make the answer feel local.
- Remove the customer language from context and let Claude pick whichever language is most probable.
- Ask Claude to translate every proper noun and trademark to match the customer’s language.
Explanation
Correct Answers
A:Currency conversion is not a stylistic translation. Keeping the stated USD amount and code avoids inventing exchange rates; any requested conversion needs a trusted rate and an explicit calculation policy.
B:A localization glossary identifies terms that are not ordinary language to translate. An explicit no-translation rule for registered names helps preserve product identifiers while allowing the rest of the answer to match the buyer’s language.
Incorrect Answers
C:Unrequested conversions can misstate quoted prices and fabricate exchange rates. Localized prose should not change amounts or currencies without an authorized conversion source and a clear user instruction.
D:Omitting the known audience language makes response localization less reliable. The prompt should state the requested language rather than leave it as an implicit guess.
E:An instruction to rewrite all proper nouns conflicts with the registered-name preservation requirement. Product names often function as stable identifiers rather than natural-language text.
Question 305
A prompt asks a model to evaluate a policy with contradictory requirements. What should the developer resolve before optimizing wording?
- Add a clear no-evidence abstention path and test that branch.
- Characterize the system's operating constraints and available implementation choices.
- Clarify which constraints have priority and which outputs are prohibited.
- Hand off exact arithmetic to code and verify the result independently.
Explanation
Correct Answer
C:No wording trick makes incompatible goals jointly satisfiable; a coherent instruction hierarchy is prerequisite to evaluation.
Incorrect Answers
A:Without an explicit failure outcome, the model may fill gaps with plausible text; the client must support an unanswered status; this implements explicit insufficient evidence path rather than the demanded contradictory prompt constraints.
B:A recommendation is discriminating only when the model knows the architecture and criteria it must optimize; it follows operational context in prompting constraints but leaves contradictory prompt constraints unspecified.
D:Prompt wording can affect reasoning but cannot replace deterministic computation or validation for precise totals; when the goal is contradictory prompt constraints, optimizing correct arithmetic enforcement misses that operational boundary.
Question 306
A user-submitted web page says to ignore system rules. What should the prompt design reinforce?
- For this workflow, delimit quoted email bodies as source content under the actual user instruction.
- Label the page as untrusted evidence, with instructions provided separately.
- Route exact arithmetic to code and verify the result independently.
- Use diverse examples that cover distinct edge conditions and failure modes.
Explanation
Correct Answer
B:Clear data boundaries reduce instruction confusion; actual tool authorization still requires enforcement outside the prompt.
Incorrect Answers
A:Clear containment of quoted text prevents the model from assuming the email sender can issue instructions to the assistant; the effect concerns nested instruction boundaries; the success condition is untrusted content labeling.
C:Prompt wording can affect reasoning but cannot replace deterministic computation or validation for precise totals; the scenario's controlling requirement is untrusted content labeling, not the related correct arithmetic enforcement concern.
D:Variety communicates generalizable rules; repeated trivial cases add cost without clarifying boundaries; testing untrusted content labeling would still fail after improving multi-criterion integration.
Question 307
A model mistakes placeholder braces in a template for actual customer data. What should integration code do?
- Render and validate template variables before making the API call.
- For this workflow, limit response scope, request evidence-backed claims, and allow abstention on missing facts.
- State how absent values must be represented in the required schema.
- Introduce a few carefully selected examples that distinguish ambiguous cases.
Explanation
Correct Answer
A:Unresolved placeholders or broken interpolation are input-construction defects, not model reasoning failures. Test template rendering before API submission and escape user-controlled fields rather than relying on the model to interpret placeholders.
Incorrect Answers
B:Clear scope and uncertainty policy discourage unsupported elaboration without pretending to guarantee truth; this implements concise grounded response policy rather than the demanded prompt variable interpolation.
C:Distinguishing missing data from an omitted key supports consistent downstream parsing and auditability; testing prompt variable interpolation would still fail after improving missing-field representation.
D:Demonstrations expose the decision boundary better than adding general adjectives to the rule; where prompt variable interpolation matters, a control for borderline few-shot examples leaves that test unmet.
Question 308
A team wants a short, factual answer rather than a chain of speculative details. What should the instruction specify?
- Limit response scope, request evidence-backed claims, and allow abstention on missing facts.
- Build a parameterized, versioned prompt template.
- Partition governing instructions from quoted source material with clear boundaries.
- Mark the page as untrusted evidence, with instructions provided separately.
Explanation
Correct Answer
A:Clear scope and uncertainty policy discourage unsupported elaboration without pretending to guarantee truth. Allow the model to state that evidence is insufficient and require citations for claims that must be traceable.
Incorrect Answers
B:Template variables isolate runtime inputs from shared instructions, making edits testable and consistent across use cases; for this request, concise grounded response policy is the missing piece, not reusable prompt template.
C:Tagged or distinctly placed input reduces confusion about what is content to analyze versus instructions to obey; because prompt data delimiting is its concern, concise grounded response policy remains outside its effect here.
D:Clear data boundaries reduce instruction confusion; actual tool authorization still requires enforcement outside the prompt; the effect concerns untrusted content labeling; the success condition is concise grounded response policy.
Question 309
An assistant must compare competing proposals using cost, latency, and security. How should the task be framed?
- Provide a clear no-evidence abstention path and test that branch.
- Provide a few carefully selected examples that distinguish ambiguous cases.
- Set out how absent values must be represented in the required schema.
- Define which constraints have priority and which outputs are prohibited.
- Provide explicit evaluation criteria and a comparison output structure.
Explanation
Correct Answer
E:A criterion-by-criterion assessment reduces arbitrary preference and makes the recommendation auditable. Use the same criteria and evidence standard for each proposal so a reader can compare its trade-offs rather than just its style.
Incorrect Answers
A:Without an explicit failure outcome, the model may fill gaps with plausible text; the client must support an unanswered status; the scenario's controlling requirement is rubric-guided comparison, not the related explicit insufficient evidence path concern.
B:Demonstrations expose the decision boundary better than adding general adjectives to the rule; the rubric-guided comparison criterion differs from borderline few-shot examples; the gap persists.
C:Distinguishing missing data from an omitted key supports consistent downstream parsing and auditability; it follows missing-field representation constraints but leaves rubric-guided comparison unspecified.
D:No wording trick makes incompatible goals jointly satisfiable; a coherent instruction hierarchy is prerequisite to evaluation; testing rubric-guided comparison would still fail after improving contradictory prompt constraints.
Question 310
A prompt works on ordinary billing questions but fails on deliberate requests to waive approval rules. Which new evaluation set best tests that weakness?
- Increase temperature to see whether risky requests sometimes get more creative replies.
- Repeat the successful billing examples and count how often wording stays identical.
- Judge only whether the final answer sounds reassuring to a reviewer.
- Add adversarial and boundary-case prompts that try to override approval constraints, and score the resulting actions.
Explanation
Correct Answer
D:A regression suite made only of happy-path examples cannot reveal policy bypass attempts. Include realistic role-spoofing, ambiguous authority and high-impact action cases, and assert both responses and tool-side effects.
Incorrect Answers
A:Changing sampling may vary responses but does not establish that the policy is enforced. Security assurance requires representative hostile cases and explicit pass/fail criteria.
B:Stylistic agreement on easy inputs provides little evidence about authorization robustness. The failure occurs on adversarial intent, so the test set must exercise that boundary.
C:A fluent refusal or approval statement can diverge from actual tool activity. Evaluations must inspect authoritative execution behavior for high-impact operations.