Prompt Engineering

Prompt engineering studies how instructions, examples, context, roles, constraints, and multimodal cues can steer pretrained models, sometimes without changing their parameters. Common approaches include zero- and few-shot templates, structured reasoning and task decomposition, retrieval and tool-use context, prompt search or routing, and learned prompts for language, vision, and graph models. These methods are applied to tasks such as classification, generation, reasoning, robotics, and education, with evaluation increasingly testing sensitivity across models, datasets, languages, and repeated runs. A central tension is that task-specific prompts can improve accuracy and control, but gains are inconsistent and may trade off against cost, clarity, robustness, fairness, or security; controlled comparisons remain essential.