Prompt engineering is the discipline of writing the input you give a language model - the instructions, context, and examples - so that it produces the output you want. Because a modern model’s behavior is steered heavily by what it is asked, small changes in wording, structure, or the inclusion of a few worked examples can swing the quality of a result dramatically, without touching the model itself.
The technical foundation comes from the 2020 GPT-3 paper, “Language Models are Few-Shot Learners” by Tom Brown and colleagues at OpenAI. They showed that a large model could perform a wide range of tasks with “tasks and few-shot demonstrations specified purely via text interaction with the model,” and “without any gradient updates or fine-tuning.” In other words, you could teach the model what you wanted simply by describing the task and showing a few examples in the prompt - the behavior known as in-context learning. That discovery turned the prompt itself into the primary control surface, which is exactly what prompt engineering operates on.
The discipline has since been codified by the major labs in their own documentation. Anthropic’s prompt-engineering guide, for example, organizes the work around techniques it names directly - “clarity and examples to XML structuring, role prompting, thinking, and prompt chaining.” These are not folklore but officially recommended techniques, which is part of why prompt engineering is treated as a real skill rather than guesswork.
Why business readers should care: prompt engineering is the cheapest, fastest lever for improving AI output - no training, no new infrastructure, just better instructions. For most business use cases, a well-constructed prompt closes more of the gap to “production quality” than people expect, and it is the natural first step before reaching for more expensive options like fine-tuning.