Prompt engineering

What it is

A prompt is the text and context sent to a model, including system instructions, user input, retrieved documents and any examples. Prompt engineering is the discipline of shaping those inputs so the response is accurate, on brief and in the right format. It overlaps with retrieval design, tool use and evaluation, because prompt quality alone cannot fix missing or wrong source data.

Why it matters

For marketing and search teams, prompts drive the quality of content drafts, classification tasks, data cleaning and internal research at scale. Good prompts reduce rework, hallucinated detail and inconsistent tone across a team. Prompts also shape how AI search features are tested, since the same question phrased differently can produce very different answers about your brand.

How it works

Practitioners state the role and task plainly, supply the source material rather than relying on model memory, give one or two worked examples, specify the output format, and set explicit limits such as word count or what not to include. They then run the prompt across a test set, score the outputs, and iterate, keeping successful prompts in a shared library with version notes. For production use, prompts are stored in code or a prompt management tool alongside evaluation cases.

When it applies

Applies whenever a team uses language models repeatedly for the same task, and whenever consistency, accuracy or brand tone matter more than a one off answer.

Examples

• A content team uses a prompt that supplies the approved product spec sheet and asks for a 60 word description in UK English with no superlatives.
• An SEO analyst prompts a model to classify 5,000 queries into intent buckets, with three labelled examples and a rule to return only the label.
• A brand tests ten phrasings of the question "who are the best providers of X" in an AI assistant to see which competitors appear and which sources are cited.

How it is measured

• Accuracy or acceptance rate of outputs against a human reviewed sample
• Edit distance or time to publish between first draft and final copy
• Format compliance rate, such as percentage of outputs returning valid JSON or the specified length
• Cost and token use per completed task across prompt versions