What is prompt engineering?
Writing instructions a model follows well. Less magic than advertised, more useful than its critics claim, and bounded in what it can achieve.
Prompt engineering is the practice of writing instructions that get useful output from a language model. Its effective core is unglamorous: state the task plainly, supply the context the model cannot guess, show the format you want, and say what to do when the material is missing. What it cannot do is give a model knowledge or make output trustworthy; that takes grounding in real sources and a way to check the answer.
The practice behind the buzzword.
The name suggests a discipline with secret techniques, and the working reality is closer to careful specification. Models answer the question they read, not the one you meant, so most prompt improvement is just removing ambiguity: name the audience, fix the length, attach the material, forbid the padding. A model told "summarize this for a committee that needs the caveats" produces a different summary than one handed the bare text, and neither required wizardry. The system prompt is where a tool builder applies the same discipline permanently.
What the words control, and what they do not.
Prompting steers how the model responds, not what it knows or how reliably it knows it. Asking for sources makes the model attempt citations; it does not make those citations real, which is why instructed politeness never fixed invented references. Wording also competes with ordinary sampling variation: the same prompt yields slightly different answers across runs, more so at higher temperature. Prompting is real leverage over the shape of an answer and no leverage over its truth.
Where structure takes over.
The mature use of prompting treats it as one layer among several. In a grounded research tool the prompt asks for citations and restraint, but retrieval supplies the passages and a verification gate enforces the restraint, refusing answers the material does not support. That division of labor is the honest lesson of the prompt-engineering era: instructions shape behavior cheaply, and structure makes the behavior checkable. A tool that only prompts is asking; a tool that verifies is answering. A different question is what happens when the model's own words choose the next action, which what an AI agent is covers from the top.