An engineer assembles the port-operations prompt by concatenating the system instructions, the retrieved circular text and the planner's question into one undifferentiated block. The model occasionally follows a sentence inside a circular that reads like an instruction. What prompt-construction practice most directly reduces this?
Select an answer to reveal the explanation.
Short Explanation
When you combine rules and retrieved text into one wall of text, the model can confuse document words with instructions. Delimit the retrieved passages, label them as reference data, and keep your instructions in the system role; this structure is your first line of defense.
Full Explanation
Language models do not have a hardware boundary between instructions and content; the separation is something the prompt has to convey. Explicit delimiters around retrieved passages, a label identifying them as reference material to use as evidence rather than obey, and keeping actual instructions in the system role give the model a consistent structural signal about which text carries authority. This is a standard defense against instruction-like content arriving from a corpus, but it is not a guarantee: treat retrieved text as untrusted and validate outputs. Shortening passages may reduce exposure statistically, but it leaves the vulnerability intact and can cost retrieval quality. Escaping punctuation misunderstands the mechanism: there is no parser to escape, because the model reads natural language, and mangling punctuation can degrade comprehension. Putting retrieved passages before the system instructions does not establish a category distinction; the model still sees a stream of text, and position alone cannot mark a passage as data. For an exam, distinguish structural separation from attempts to reduce or disguise suspicious text. In practice, check that the prompt clearly labels retrieved passages as data, keeps governing instructions in the system role, and applies output guardrails.