prompting

Prompting is the practice of crafting inputs, or prompts, to guide the behavior of a language model or other AI system. A prompt encodes the task, context, and desired format, and the model's output is conditioned on that input. Prompt design aims to elicit useful, accurate, and aligned behavior without explicit retraining, by leveraging the model's learned patterns and knowledge.

Common prompting paradigms include zero-shot prompting, where the model is asked to perform a task without examples; few-shot prompting, which includes a small set of demonstrations; and instruction prompting, where a concise directive states the goal. In-context learning places examples directly in the prompt to steer the response. System or role prompts set global behavior, while user prompts request specific outputs.

Techniques such as chain-of-thought prompting encourage the model to reveal reasoning steps, while structured prompts specify required formats such as lists, code blocks, or data tables. Prompt templates and prompt libraries enable reuse and standardization, and prompt chaining combines multiple steps or subtasks into a sequence.

Prompting intersects with broader issues of reliability and safety. Small changes in wording can lead to substantially different results, affecting accuracy and bias. Risks include hallucinations, unsafe content, and prompt injection, in which adversaries craft prompts or inputs to bypass safeguards or extract sensitive information.

In practice, prompting underpins many applications, including chat assistants, writing aids, code generators, data extraction, and tutoring. It is often used alongside fine-tuning, retrieval augmented generation, or model instruction tuning to improve robustness and controllability.