A prompt is the instruction and context handed to a model at inference time. Because large models perform in-context learning (Brown et al., 2020), the prompt — not retraining — is the primary way most users steer behaviour. The systematic study of this is surveyed by Liu et al. (2023).
Effective prompts state the task clearly, supply relevant context or examples, and specify the desired format. Small wording changes can meaningfully shift results.