AI prompts for marketing: a practical template

AI prompts for marketing can save a small team from rewriting every first draft, or create a faster route to an embarrassing mistake. The difference is the brief you give the machine. A useful prompt tells it who the customer is, what it may claim, and what the finished piece must look like. Here is a repeatable way to build one.

Start with the business decision

A prompt is simply the instruction you give an AI writing tool. Most weak marketing prompts begin with a format: “write an email” or “make five ads.” That gives the tool a container but no job.

Begin with the decision you need the reader to make. A plumber may want a homeowner to understand why a same-day callout costs more. A retailer may want a previous buyer to notice a new product line. Those are different jobs, so they need different facts, objections, and proof.

Write one sentence before opening any AI tool: “After this message, the reader should understand or do ___.” If the blank is fuzzy, the draft will be fuzzy too. This one line prevents a common waste of time: asking for “engaging” copy, then rejecting several versions because none has a clear point.

Next, name the audience in plain terms. Include what they already know, what they care about, and what could make them hesitate. “Local homeowners whose boiler stopped working today” gives far more direction than “our target customer.” That context is not decoration. Google’s prompt design guidance says models use the details and constraints supplied in a prompt to tailor the response.

Give the model facts, not a blank page

An AI writing tool predicts plausible language. It does not know which current price, promise, or availability statement your business has approved. Give it a small source pack for each assignment.

For a campaign, that pack might include the offer, deadline, product details, customer question being answered, approved proof, and a list of claims the tool must avoid. Put the facts in a clearly labeled section called “source material.” Then add a direct instruction: “Use only these facts. If a needed fact is missing, write [FACT CHECK] instead of inventing it.”

That instruction will not make every draft true by itself. It does force uncertainty into view, which is where a human can handle it. The person approving copy still owns the claim. For larger teams, put that boundary alongside the review rules in an AI content policy.

Keep customer names, private conversations, contract terms, and other sensitive material out of a casual prompt unless your company has checked the account’s data settings and approved that use. A good draft is not worth a loose data habit.

Set the output shape before the writing begins

The output shape means the visible form of the finished work: an email, a landing-page section, a social caption, or a table of ad ideas. Specify it. State the length, the number of options, the reading level, and the exact pieces you need.

Instead of “write social posts about our sale,” try this:

Create four Instagram captions for the source material below.
Audience: homeowners considering a boiler replacement this month.
Goal: encourage a booking request without making savings claims.
Each caption: 70 to 100 words, one opening sentence that names a customer problem,
one concrete service detail, and one plain next step.
Do not use exclamation marks, invented testimonials, price comparisons, or urgency
we cannot prove.

Specific limits make the first draft easier to judge. Google’s documentation also recommends stating response format and constraints rather than letting the model choose them. That is useful because an owner can compare four captions against the same standard instead of arguing about a general feeling.

Avoid asking for a whole campaign in one giant prompt. Break the job into stages: first the message angle, then the email or ad, then the edit. OpenAI’s prompt engineering guide recommends separating high-level instructions from the details of an individual request. In practical terms, keep your standing rules in one block and add today’s assignment underneath.

Build a reusable marketing prompt

The stable part of your prompt is your instruction block. It should carry the things that rarely change: who you serve, the brand voice, claims that require approval, and the review standard. Your team can keep it in the document where approved marketing materials live.

Use this template as a starting point. Replace every bracketed instruction before use.

You are preparing a first draft for [business name].

Audience
[Who will read this? What do they know? What problem is on their mind?]

Business goal
[What should the reader understand, decide, or do after reading?]

Source material
[Paste only approved current facts, product details, customer language, and proof.]

Voice
- [Describe the level of formality in observable terms.]
- [State sentence and paragraph preferences.]
- [List words or claims the business avoids.]
- [Name any required legal or operational wording.]

Assignment
[Name the format, channel, length, and required sections.]

Accuracy rules
- Use only the source material for factual claims.
- Mark missing facts as [FACT CHECK].
- Do not create testimonials, guarantees, prices, deadlines, or results.

Output
[State exactly what to return, such as three subject lines followed by one email.]

This is deliberately less clever than the prompts people share online. Good. Your staff should be able to update it when a price changes or a new service launches. If only one person understands the instruction, it is a fragile business process.

For the voice section, borrow observable habits from your strongest existing work: how quickly it gets to the point, words customers actually use, and which kinds of humor are off-limits. Our AI brand voice guide shows how to turn those patterns into a one-page reference rather than a pile of adjectives.

Use examples when tone matters

Examples are often the fastest way to improve a recurring format. This approach is called few-shot prompting, meaning you show the tool a few input-and-output examples before asking it to make a new one. Google says consistent examples can guide a model toward the required pattern, especially for format and phrasing.

Choose two or three approved pieces that do the same job as the new request. For each example, identify the situation and the finished copy. Do not dump in a folder of old work. Mixed-quality samples teach mixed-quality habits.

Add a short note after the examples: “Follow the structure and level of specificity. Do not reuse phrases or facts from these examples.” That keeps an old promotion or retired price from wandering back into a new email.

Examples do not replace facts. They show style and structure. The source pack still carries the truth of the current assignment.

Make the human check part of the prompt

No prompt turns marketing approval into a button press. Add a final self-check that makes the model point out what needs a person’s attention, then have an actual person check it.

Ask the tool to return a short list after the draft:

Before the draft, list:
1. Every factual claim used and its matching source-material line.
2. Any phrase that needs legal, price, or availability approval.
3. Any missing detail that would make the message stronger.

That list gives the reviewer a practical starting point. It also makes it harder for a polished sentence to slip past without evidence. Anthropic’s prompting overview starts with the same principle: define what success looks like and test the result against it.

For a recurring campaign, keep a small scorecard. Did the draft use only approved facts? Does it sound like the business? Is the next step accurate? Could a customer misunderstand the offer? These are business checks, not technical ones. They belong with the person who would answer the customer’s reply.

Fix the usual prompt failures

When the output feels generic, the prompt usually lacks a real customer problem or source material. Add both before changing tools. When it sounds off-brand, replace words like “friendly” with visible writing rules and a couple of approved examples. When it invents details, narrow the source pack and require [FACT CHECK] markers.

When several staff members get different results from the same task, compare their prompts before comparing the AI. One person may be pasting last month’s offer while another is using the current one. A shared prompt and a named owner reduce that drift. It is the same operating discipline that makes a marketing stack audit useful: know what is in use, who owns it, and what it is allowed to do.

Worth watching

AI writing models and their instructions change often, so revisit a high-volume prompt whenever you change models or a campaign starts producing noticeably weaker drafts.

Keep a small library of approved prompts by job, not by tool. The task stays recognizable even when the software changes.