Prompt engineering for marketers: the parts that actually matter

Most prompt advice is either magic incantations or generic instructions to "be specific". Here's what consistently changes output quality in actual work.

1. Give it the constraints, not just the task

The difference between mediocre and useful output is nearly always in the restrictions.

Weak: "Write Google Ads headlines for a plumbing business."

Better: "Write 10 Google Ads headlines for an emergency plumbing service in Cape Town. Maximum 30 characters each. Must not use exclamation marks or the words 'best' or 'affordable'. At least four must reference speed of response. Do not claim 24/7 availability."

The second produces usable output because it encodes what you'd otherwise reject manually.

2. Tell it what to do when it doesn't know

Models fill gaps by default. That's the root of most hallucination in practical use.

Add explicit handling: "If the data provided doesn't support a conclusion, say so rather than estimating. Never invent statistics."

In production prompts I write, this instruction earns its place more than any other. The audit tool on this site has an entire paragraph about what to do when the input data is an error message rather than real data — because that happened, and without the instruction it would have confidently invented performance metrics.

3. Give examples of what good looks like

Two or three examples of the output you want beats any amount of description.

If you want ad copy in a specific voice, paste three ads you consider good. The model will infer register, rhythm and structure more accurately than you can describe them.

4. Specify the output format precisely

If you need JSON, say so and describe the shape. If you need markdown with specific headings, say which. If you're feeding the output into something else, state that.

"Return only the inner HTML for an email body. No markdown, no code fences, no html or body tags. Use only h2, p, ul and li."

Vague format instructions produce output you then have to clean up, which erases the time saved.

5. One task per prompt

Asking for ad copy, a landing page outline, and a keyword list in one prompt produces three mediocre things. Three focused prompts produce three good ones.

This is especially true in automations, where a single overloaded prompt is harder to debug when one part of the output degrades.

6. Iterate on the prompt, not the output

If you're editing the result heavily every time, fix the prompt instead. A prompt you use weekly is worth twenty minutes of refinement — that compounds; editing the output doesn't.

Keep the ones that work. A short personal library of five well-tuned prompts beats a downloaded pack of 500.

What doesn't matter as much as claimed

Role-playing preambles. "You are a world-class marketer with 20 years of experience" adds little on modern models. Specific constraints add a lot.

Politeness. Doesn't change quality.

Extreme length. Long prompts aren't automatically better. Relevant prompts are. Padding dilutes.

Where marketers get the most value

In rough order of return:

1. Structured extraction — turning messy input into consistent data
2. First drafts at volume — ad variations, subject lines, outlines
3. Summarising — calls, transcripts, long documents
4. Analysis with data supplied — give it the numbers, ask for interpretation
5. Editing your own writing — critique is more reliable than generation

Note what's last. Generating finished public-facing copy from scratch is the weakest use, and it's what most people try first.

The honest limitation

Models are confident regardless of correctness. They don't know what they don't know, and they won't tell you when they're guessing unless you build that in.

Everything above is about constraining that tendency. Review anything that goes out under your name.