Prompt engineering

Prompt engineering is the discipline of writing the text that shapes how a large language model responds. That text includes the system prompt — which fixes the model's role, rules, tone, and output format — plus the user request itself and any examples you attach. Done well, it is less about a single magic phrase and more about engineering a template that produces the right answer across the messy variety of real inputs you'll actually see in production.

It works because today's models are extraordinarily sensitive to framing. Stating the task explicitly, supplying a couple of worked examples (few-shot prompting), specifying the exact output shape, and asking the model to reason before it answers all measurably change behavior. Pushing reasoning further is its own technique — see chain-of-thought — and the amount of guidance you can pack in is bounded by the model's context window, so concise, well-ordered prompts matter.

A concrete example: suppose you want a model to extract invoice totals. A vague "find the total" prompt yields inconsistent formats and occasional commentary. An engineered prompt assigns a role ("You are a precise data-extraction service"), states the rule ("Return only a JSON object with the field total_usd"), shows one example input and output, and instructs the model to return null when no total is present. The wording barely changed in length, but the output became parseable and stable — which is the entire point of prompt engineering.