Prompt engineering
Prompt engineering is the craft of writing instructions, examples and structure for a large language model so it produces reliable, useful outputs. The model is fixed; the prompt is the lever, which is why most production LLM work actually happens here rather than in the model itself.
Prompt engineering covers everything you put around the user's question before it reaches the model: the system message, the role and tone, the format of the answer, a few worked examples, the retrieved context from a retrieval-augmented generation pipeline, and the constraints that keep the output usable downstream. It is the cheapest, fastest lever in any large language model application — long before fine-tuning or swapping the foundation model, a better prompt usually closes most of the gap.
In practice, a good prompt is more like a small spec than a clever one-liner: it states the task, the inputs, the format, the edge cases and what the model should do when it does not know. Structure beats cleverness.
The honest take is that prompt engineering only really pays off in combination with evals. A prompt that works today can quietly degrade when the model is updated or when a new edge case appears in the data, and without a test set you will not notice until a user does. The rule of thumb: if a prompt matters enough to ship, it matters enough to measure.
