Prompt Engineering

Examples inside a prompt can show a language model what a successful answer should look like without changing its trained parameters. Lilian Weng's prompt-engineering survey explains this as few-shot prompting: supply several input-and-answer pairs, then present the new input.

The examples communicate both the task and the expected response pattern. Their selection and ordering can also bias the result, and they consume space that could otherwise hold the user's material or the answer. Weng describes prompting as an empirical practice whose effects vary across models. Examples are therefore part of the system's behavior to evaluate, rather than interchangeable illustrations attached to an instruction.

“It is an empirical science and the effect of prompt engineering methods can vary a lot among models, thus requiring heavy experimentation and heuristics.”

“At its core, the goal of prompt engineering is about alignment and model steerability.”