Question 280
An engineering postmortem must be summarized for customer-support staff. Which prompt detail most directly improves usefulness?
- Ask for the longest possible technical derivation.
- Use a generic instruction to sound impressive.
- Specify audience knowledge, customer impact, workarounds, and escalation needs.
- Preserve every internal acronym without definition.
Explanation
Correct Answer
C:Audience and action needs guide which technical facts to retain and how to explain them. Support staff need information that changes how they assist customers. Internal implementation details belong only where they explain impact, workarounds, or escalation decisions for that audience.
Incorrect Answers
A:Maximum technical depth may obscure the practical information support staff need.
B:An impression-focused instruction does not identify the information needed by support staff.
D:Unexplained terminology can make the summary less accessible to its intended readers.
Question 281
A classifier's predictions change substantially when the same labeled examples are reordered. What should the prompt evaluation examine?
- Compare file sizes of the prompt variants while skipping classification on held-out inputs.
- Reorder both the examples and the held-out inputs while changing label definitions, then attribute any score change solely to example order.
- Performance across controlled example-order variants on representative inputs.
- Only the one ordering that produced the highest development score.
- A claim that examples have no influence on classification.
Explanation
Correct Answer
C:Varying only the order reveals whether the decision rule is robust or depends on incidental presentation. Keep labels, input cases, and other settings fixed during the comparison so ordering is the variable under examination. Robustness matters when future prompt maintenance may change example placement.
Incorrect Answers
A:Equal size does not establish equal model behavior. The observed defect concerns predictions under different example orders.
B:Simultaneous changes confound the experiment. Order sensitivity needs a controlled comparison in which the decision rule and evaluation inputs remain consistent.
D:Selecting a favorable ordering alone can conceal fragility on other arrangements and inputs.
E:The observed behavior directly shows that their presentation affects results.
Question 282
A prompt performs well on formal support tickets but poorly on short, misspelled messages common in production. What should the examples include?
- Only the model's own confident predictions without review.
- Use informal production messages but retain labels from unrelated formal examples.
- Representative informal messages with correct labels and the same decision rules.
- Only longer and more grammatical versions of the existing examples.
Explanation
Correct Answer
C:Examples from the actual input style expose the model to the variation it must handle. Informal phrasing is a distribution characteristic, not necessarily a different intent taxonomy. Correctly labeled examples can preserve the business rule while showing how it appears in real messages.
Incorrect Answers
A:Unverified predictions can reinforce the same errors instead of providing reliable demonstrations.
B:Representative wording cannot compensate for labels that do not match the actual messages.
D:Polished variants retain the mismatch with the informal messages causing production errors.
Question 283
Two nearly identical examples demonstrate the same intent but are assigned conflicting labels without a rule explaining the difference. What should the designer do first?
- Review the annotations and correct inconsistent examples against the intended taxonomy.
- Treat the disagreement as evidence that every prediction is correct.
- Add more contradictory copies so both labels are equally frequent.
- Increase output length so the model can defend either label.
Explanation
Correct Answer
A:Contradictory demonstrations can teach unstable behavior even when the model follows the prompt carefully. The same semantic situation should not receive incompatible labels unless an explicit contextual distinction explains why. Annotation cleanup improves the target the model is being asked to imitate.
Incorrect Answers
B:An incoherent example set cannot justify abandoning the task's expected labels.
C:Balancing erroneous examples does not establish a coherent decision rule.
D:A longer rationale does not repair inconsistent supervision.
Question 284
A prompt inserts user text inside XML-like delimiters, but the text can contain the closing delimiter itself. What should the application do?
- Encode or safely serialize the inserted content and preserve a clear distinction between data and instructions.
- Choose a rarer delimiter and assume no escaping or validation is ever needed.
- Treat any text after the first closing delimiter as a new system instruction.
- Assume delimiters cannot occur in ordinary user input.
Explanation
Correct Answer
A:Safe serialization prevents source text from accidentally breaking the intended prompt structure. Structural delimiters describe boundaries only when inserted data cannot masquerade as those boundaries. Encoding also avoids altering the actual source content just to make it fit a prompt template.
Incorrect Answers
B:A less common marker can still occur in unrestricted source text and break prompt boundaries.
C:User-controlled delimiters should not grant the following content higher authority.
D:Unrestricted text can contain the same characters as the application's chosen delimiters.
Question 285
A rendered extraction prompt still contains '{{customer_text}}' instead of the customer's document. What should the pipeline do?
- Let the model invent a likely customer document.
- Retry the same rendered prompt with a larger output limit.
- Fail template validation and supply the actual document before calling the model.
- Treat the placeholder itself as the customer's complete statement.
Explanation
Correct Answer
C:Checking unresolved placeholders catches the assembly error before an invalid request is processed. Template assembly is an application concern that can be checked before inference. Calling the model with an unresolved variable cannot recover the absent customer's source document.
Incorrect Answers
A:Invented input cannot support an extraction task for the missing source.
B:More output capacity cannot supply the absent source content.
D:A template variable is not the intended document and would produce meaningless extraction results.
Question 286
A product document never mentions offline operation. The extractor must return supported capabilities. What should it conclude about offline support?
- That offline operation is definitely unsupported.
- That the product cannot be summarized at all.
- That offline support is not established by the supplied evidence.
- That every modern product supports offline operation.
Explanation
Correct Answer
C:An unmentioned feature should remain unknown unless the task provides a rule for interpreting omission. A capability may be undocumented rather than absent from the product. The output should distinguish that evidence gap from an explicit statement that the feature is unavailable.
Incorrect Answers
A:Silence in the document does not establish a negative capability claim.
B:The missing capability fact does not prevent extracting other supported information.
D:General assumptions cannot establish this product's documented capabilities.
Question 287
A routing task permits exactly one label, but some messages equally meet two label definitions. Which prompt addition is most useful?
- An explicit business-priority tie-breaker and examples where it applies.
- A statement that ambiguous inputs never occur.
- A request to output both labels despite the single-label contract.
- A rule to choose the alphabetically first label in every case.
Explanation
Correct Answer
A:A defined tie policy makes ambiguous single-label decisions consistent with the workflow's needs. The tie-breaker should reflect actual handling priorities, such as which team must own a combined case. Without that rule, a forced single answer may be arbitrary even when both labels are understood.
Incorrect Answers
B:Denying an observed input class leaves the decision problem unresolved.
C:Returning multiple labels violates the downstream routing requirement.
D:Alphabetical order is arbitrary unless it matches an explicitly intended business priority.
Question 288
A customer message requests both a refund and an address correction. The workflow can perform both, but the prompt forces one intent. What should change?
- Duplicate the whole message under one intent without recording the other.
- Discard whichever request appears second.
- Use a multi-intent output contract and define how independent intents are represented.
- Label every multi-intent message as unrelated.
Explanation
Correct Answer
C:The output should preserve both actionable requests when the workflow supports processing them. Each intent needs enough associated details to support its own action path. A single overall category is unsuitable when it would discard an independently requested change.
Incorrect Answers
A:Repetition does not preserve the missing action as a separately processable intent.
B:Text order does not justify losing an independent user request.
D:The message clearly contains supported intents and should not be rejected as irrelevant.
Question 289
Two classification labels, 'billing problem' and 'payment issue,' overlap without a documented distinction. What should the designer do before optimizing the prompt?
- Measure success only on messages that clearly fit neither label.
- Keep the overlapping labels but evaluate only exact agreement with the model's previous outputs, treating those predictions as the definition of the categories.
- Require the model to explain why identical definitions are different.
- Increase the number of examples while assigning overlapping cases arbitrarily.
- Clarify or merge overlapping categories and relabel affected evaluation examples.
Explanation
Correct Answer
E:A coherent label definition is necessary for meaningful training examples and evaluation. Prompt optimization cannot create a business distinction that the category owners have not defined. Updating reference labels keeps evaluation consistent with the clarified meanings.
Incorrect Answers
A:Avoiding the overlap hides the problem rather than defining the target behavior.
B:Self-consistency does not establish a coherent business taxonomy. Previous arbitrary decisions can remain consistently wrong or ambiguous.
C:A model cannot recover a business distinction that the taxonomy never specifies.
D:Arbitrary examples preserve an incoherent taxonomy and introduce label noise.
Question 290
A source says, 'Aster acquired Birch. The buyer will retain its existing finance system.' The extraction field is the organization retaining its finance system. Which output follows the stated relationship?
- Birch, because it is the last organization named before the second sentence.
- Aster, because the buyer is the acquiring organization.
- No organization, because the second sentence does not repeat a proper name.
- Both organizations, because an acquisition combines their operations.
Explanation
Correct Answer
B:The second sentence refers to the role established in the first. Resolving that relation identifies Aster even though Birch is the closer organization name in the text.
Incorrect Answers
A:Nearest-name matching ignores the acquisition roles. Birch is the acquired organization, while the source attributes retention to the buyer.
C:Extraction can resolve an explicit, unambiguous reference across sentences. Requiring a repeated proper name would discard information supported by the source's stated relationship.
D:That is an inference about integration plans. The sentence assigns the action to one role and does not say that both systems will be retained.