Prompt engineering
Prompt engineering is writing and structuring the instructions you give a language model so the output is reliable enough to ship — not a one-off chat that happened to work.
A production prompt is a spec: role, constraints, the shape of the answer, examples of good and bad, and what to do when the input is rubbish. You version it, you test it against a set of real cases, and you change one thing at a time. That is closer to programming than to witty phrasing.
It sits under every serious agent and RAG setup. Retrieval decides what the model sees; the prompt decides what it is allowed to do with it. A clever model with a vague prompt still invents. A smaller model with a tight prompt often does not.
The fad version is a prompt pack you paste once. The working version is evals: a list of inputs, an expected behaviour, and a fail when the model drifts after a provider update. We do not ship an AI step without that list.
Book a call