Prompt engineering

Prompt engineering — The practice of designing and refining model inputs — instructions, examples, and context — to elicit specific, reliable outputs from a language model.

Overview

Prompt engineering is the discipline of crafting the inputs supplied to a language model to produce desired outputs. A prompt is the complete text sent to the model at inference time: it may include a system instruction, few-shot examples (In-context learning), a Chain-of-thought scaffold, retrieved context (RAG), and the user's query. Prompt engineering shapes all of these components to improve output quality, consistency, safety, or efficiency.

Prompt engineering emerged as a distinct practice as large language models (GPT-3 and successors) demonstrated sensitivity to input phrasing. Small changes to a prompt — the order of examples, the phrasing of an instruction, the presence or absence of a chain-of-thought cue — can produce substantial changes in output quality or format.

The term is sometimes contested: critics characterize it as informal trial-and-error rather than an engineering discipline, while proponents note that systematic prompt design methods (structured templates, ablation testing, retrieval-augmented construction) constitute a reproducible engineering practice.

Core techniques

Technique	Description	Reference
Zero-shot prompting	Instruction only, no examples	—
In-context learning (few-shot)	Examples included in the prompt	GPT-3 paper
Chain-of-thought	Examples include reasoning steps	Wei et al. 2022
System prompt design	Framing persona, constraints, output format	—
RAG-augmented prompts	Retrieved documents included in context	Lewis et al. 2020
Structured output prompting	Instructing JSON / XML output for programmatic use	—