Prompt engineering

Models do what the words say, not what the writer meant. A vague request gets a generic answer, and the person then spends longer editing than it took to ask properly. A clear prompt is also repeatable. Once one works, it can be saved and reused by a whole team. That is the first step towards putting a task in a workflow that runs without being retyped.

1. State the goal and who the output is for.
2. Give the facts the model cannot know, such as pricing, tone rules or the customer's situation.
3. Specify the format: a five-row table, three bullet points, 150 words.
4. Add one or two examples of good output.
5. For hard tasks, ask for the reasoning steps before the final answer.
6. Test on several cases, not one, then adjust wording based on where it fails.

A prompt is the text given to a model. Prompt engineering is the habit of writing it deliberately. A good prompt usually covers five things: the task, the background the model needs, the role or audience, the format of the answer and an example of what good looks like.

Compare "write some cold emails" with "write three cold emails to operations managers at logistics firms, under 90 words each, one clear question per email, in the tone of the example below". The model is the same. The second gets a usable draft.

• Piling on instructions that contradict each other.
• Skipping the background and expecting the model to guess it.
• Trusting the first answer on factual questions. Models can produce hallucinations, so check claims that matter.