Prompt engineering is the deliberate design and refinement of prompts to influence a language model’s responses. Wording, tone, structure and context provide signals that affect which learned patterns become salient during generation. As a result, prompts that ask for the same task can produce different outputs. section 6.3.2 section 6.3.3
Even details that appear irrelevant to the task can matter. Studies have found changes in predictions or accuracy after variations in punctuation, spacing, capitalization, label formatting, separators and example order. These effects differ across model–task combinations; newer models may be more robust, and rigid scoring can exaggerate apparent sensitivity. section 6.3.2 section 6.3.3
Refining a prompt by trial and error is not necessarily a reliable route to better performance. Without correct answers or another objective reference point, a user may mistake a convincing response for an improvement. In a classification experiment, participants sometimes believed their revisions helped even as accuracy declined. A formulation that performs poorly for one model may also behave differently with another. section 6.3.3
Context can also shift the balance between a model’s competing objectives. For example, framing a request as a fictional cybersecurity training scenario introduces defensive cues that may increase compliance where a direct request would be refused. A jailbreak deliberately manipulates such cues to seek outputs normally restricted by safety guidelines; it does not necessarily disable a rule or safeguard. section 6.3.4
Prompt changes should therefore be judged against reliable feedback where possible, not just the apparent quality of a response. Sensitivity is real, but its magnitude depends on the model, the task and how performance is measured. section 6.3.3 section 6.3.2


















































