Prompt Engineering

No jargon, no math

Since a model continues the text you give it, prompt engineering means writing the text worth continuing. Working prompts tend to share four moves. First, the task in one sentence: what goes in, what must come out. Second, the role: who the model is and who it works for, because a quality engineer and a marketer frame the same answer differently. Third, examples, the so-called few-shot: two or three input-to-output pairs teach format and judgement faster than any paragraph of rules. Fourth, constraints: the output format, length limits, and explicit instructions for missing data, such as 'if a price is absent, say missing_price, never guess'. Then you iterate: run real cases, find where it fails, adjust one variable at a time. Where it breaks: piling instructions until they contradict each other, examples that drift away from real cases, and trusting a prompt that only ever saw three cheerful test inputs.