Prompting for Marketers: The Part That Actually Matters

Output quality is set mostly by the material you supply, not the phrasing you use.

That is the finding people take longest to accept, because it is less interesting than the alternative. A carefully worded request with no context produces a competent generic answer. A plainly worded request with your actual customer language, your positioning and three real examples produces something usable.

The prompt is rarely the bottleneck. Which reframes the whole skill: the work is assembling what the model needs, not discovering the right way to ask.

---

What actually changes the output

Four inputs, in rough order of contribution.

1. Context — the material only you have. Customer language, product specifics, previous work, what you tried before, constraints. This is most of the difference and it is the part people skip, because pasting three paragraphs of background feels like more work than writing a clever instruction.

2. Task definition — what the output is for. Not "write a landing page" but "write the section that answers the objection about setup time, for people who already read the pricing page." A model given a purpose makes better choices than one given a format.

3. Form constraints — length, structure, what to leave out. Cheap to specify and disproportionately effective. "No introduction, no conclusion, no summary sentences" removes most of what you would delete anyway.

4. Phrasing. Real, and last. The gap between a well-phrased request and a plain one is smaller than the gap between context and no context, which is the opposite of how prompting is usually taught.

---

What does not work

Three widely-repeated practices worth dropping.

Magic phrases. "Take a deep breath", "this is very important to my career", "you are a world-class expert". The evidence for these is weak, it is model-specific, and it does not survive model updates. Anything whose value depends on a particular version is not a skill worth building.

Role prompting on its own. "Act as a senior copywriter" does very little by itself. What helps is the context a role instruction sometimes smuggles in — so supply that directly instead. "Write for somebody who has read three competitor pages already and is comparing" is a role instruction that carries actual information.

Longer as a proxy for better. A prompt padded with instructions the task does not need buries the ones it does. Length helps when it is context and hurts when it is exhortation.

---

The four-part prompt

A structure that covers most marketing tasks without becoming a template you follow mechanically.

1. The material. Everything relevant you already have. Previous work, customer quotes, the actual product page, the brief, the transcript. Paste generously — this is the part that does the work.

2. The task and its purpose. What you want, and what it is for. Who reads it and what they know already.

3. The constraints. Length, structure, register, and explicitly what to leave out.

4. The example. One piece of your own writing that gets it right, or one that gets it wrong with a note on why. A single concrete example outperforms a paragraph describing the style you want, and it is faster to supply.

Worked, for a marketing task:

Here is our current pricing page, our three most common sales objections as customers phrase them, and the last two emails we sent about pricing. [pasted]

Write the FAQ section for the pricing page. It is read by people who have already decided we might be right and are looking for a reason not to proceed.

Six questions, two to four sentences each. No introduction. Use the customers' own words for the questions, not ours. Do not claim anything the pricing page does not support.

Here is an FAQ we wrote that works, for tone: [pasted]

Nothing clever in it. It works because three of the four parts contain information the model could not otherwise have.

---

Iteration beats formulation

The first output is a draft of the prompt as much as a draft of the work.

What a disappointing first output usually tells you:

• Too general → you did not supply enough material
• Wrong angle → you did not say what it was for
• Right content, wrong shape → you did not constrain the form
• Confidently wrong facts → you asked it to supply knowledge instead of supplying the source

The most useful follow-up is rarely "try again, better". It is naming the specific problem: "The second section is generic. Rewrite it using the customer quotes above and nothing else."

Two techniques that reliably help:

Ask for options, not an answer. Five approaches at one paragraph each, then develop the one that is right. Cheaper than five full attempts, and you choose from a wider spread.

Ask it to critique before rewriting. "What is weak about this draft?" often produces a more useful list than a rewrite does — and you keep control of what changes.

---

Where prompting cannot help

Four limits worth knowing before you spend an afternoon rephrasing.

It cannot supply facts you did not give it. Asking for statistics produces statistics, and some will be fabricated — a real figure with the wrong source, a real source credited with something it does not contain, or a number with no origin. No phrasing prevents this. Supplying the source does. How to verify what comes back.

It cannot know your customers. What people said on calls, what they complain about, the objection that keeps coming up. That material has to be pasted in, and if you do not have it written down anywhere, that is the actual problem.

It cannot decide what is worth saying. The distinctive claim in a piece — what you tested, what you got wrong — comes from having done things. Why that ceiling exists.

It cannot verify its own output. A model checking a model produces agreement, not verification.

---

The marketing-specific discipline

Marketing output is published, so it needs a standard that most prompting advice does not apply.

Three habits:

Ask for the source, always. Any factual claim, with where it came from. Then check the ones that matter — a supplied citation is a lead, not a verification.

Constrain against the failure modes you know. "No superlatives. No urgency language. No claims we cannot support. No sentences that would be equally true of a competitor." These four instructions remove most of what makes generated marketing copy identifiable.

Separate generating from judging. Generate in one pass, evaluate in another, ideally not immediately. The judgement about whether something is good is the part that has to stay yours, and it degrades when it happens in the same motion as the generation.

---

What is worth building

Two things, and neither is a collection of clever prompts.

A context file. Your positioning, your customers in their own words, your product specifics, your constraints, three examples of writing that is right. Written once, pasted often. This is the single highest-return thing available in this whole subject, and it takes an afternoon.

A short set of patterns you actually use. Not fifty prompts. Six or seven shapes for the tasks you repeat, refined as you go. How to build one that survives a team.

What is not worth building: a library of prompts copied from elsewhere. They encode somebody else's context, which is the part that mattered.

---

Frequently asked questions

What makes a good prompt for marketing work?
Supplied material first — customer language, previous work, the actual product page — then what the output is for, then form constraints, then phrasing. The gap between good and bad phrasing is much smaller than the gap between context and no context.

Do prompt tricks like "act as an expert" work?
Role instructions do very little on their own. What sometimes helps is the context they smuggle in, so supply that directly instead: "write for somebody who has read three competitor pages and is comparing" carries real information, while "act as a world-class copywriter" does not.

Why does AI output sound generic?
Almost always because it was given nothing specific to work from. A model with no context produces the average of what has been written, which is competent and interchangeable. Pasting real customer language and real examples is what changes it.

How do I stop AI making up statistics?
No phrasing prevents it. Supply the source material and ask it to work only from that, then verify anything factual at a primary source. Asking a model for statistics produces statistics, some of which will be a real figure with the wrong attribution.

Should I use long or short prompts?
Long when the length is context, short when it would be exhortation. Padding a prompt with instructions the task does not need buries the ones it does, while pasting three paragraphs of relevant background is usually the single most effective thing you can do.

What should I build instead of a prompt collection?
A context file — your positioning, customers in their own words, product specifics, constraints, and three examples of writing that gets it right. Written once and pasted often. Prompt collections copied from elsewhere encode somebody else's context, which was the part that mattered.

How do I improve a disappointing output?
Diagnose rather than rephrase. Too general means insufficient material; wrong angle means you did not say what it was for; right content in the wrong shape means missing form constraints; confidently wrong facts mean you asked it to supply knowledge instead of supplying the source.