Prompt Engineering for Marketers: The Prompts I Actually Use
The short answer
A working marketing prompt carries four things: a role, real context the model could not otherwise know, constraints that rule out the generic answer, and a specified output format. Most prompts supply only the task, which is why they return copy that could belong to any business in any market. Below are ten prompts I use weekly, with the reasoning behind each so you can adapt rather than copy them.
Nearly every prompt guide written for marketers is a definition of prompt engineering followed by five prompts that say "act as an expert marketer and write me a Facebook ad."
That prompt produces exactly what you would expect. Confident, fluent, and indistinguishable from what your competitor gets when they type the same thing.
So this is the other version: the actual prompts, what makes each one work, and what changed in prompting practice this year that most marketing advice has not caught up with.
Why most marketing prompts fail
A prompt fails when the model has no way of knowing what makes your situation different. Ask for a Facebook ad for a Dubai interior fit-out company and you will get villa renovation copy assembled from the average of every fit-out website on the internet. The model is not underperforming. It has nothing else to work with.
Four elements fix most of it.
- Role. Sets the perspective and vocabulary. "You are a direct response copywriter who distrusts adjectives" produces different output from "you are a marketing expert", which produces nothing in particular.
- Context. The part that actually matters. Who buys, what they object to, what your client charges, what the last campaign taught you. This is what the model cannot generate, and it is the entire difference between generic and usable.
- Constraints. What to avoid, how long, what not to claim, which words are banned. Constraints eliminate the average answer, which is what you are trying to escape.
- Output format. Specify it or you will spend longer reformatting than you saved. A table, a numbered list, exactly five variations, each under 125 characters.
The working skeleton looks like this, and every prompt further down is a variation on it.
You are [role with a specific bias or standard]. Context: [Business, buyer, price point, market, what you already know works or fails.] Task: [One clear task.] Constraints: [What to avoid. What not to claim. Length. Banned phrases.] Output: [Exact format. Exact quantity.]
What changed in 2026
Three shifts worth knowing, because a lot of circulating advice predates them.
Chain-of-thought is now model-dependent. Telling the model to think step by step was the single most repeated tip for years. Reasoning models now perform that decomposition internally, and OpenAI's current guidance explicitly states that chain-of-thought instructions are unnecessary for its reasoning models, while Anthropic recommends general reasoning instructions rather than hand-written prescriptive sequences. On non-reasoning models it still pays, with Wharton Generative AI Labs research finding gains of roughly 11 to 13 percent. Match the technique to the model instead of applying it everywhere.
Examples changed jobs. Stacking worked examples to teach a reasoning model how to think no longer improves the reasoning and can compete with the model's own process. What examples are still genuinely good for is format alignment: showing the exact shape of output you want. That is a narrower and more reliable use.
The reasoning dial is the new lever. Several 2026 models expose a control over how much internal reasoning to apply before answering, such as Claude's effort levels and GPT-5's reasoning effort setting. Low for bounded mechanical work, high for analysis where being wrong is expensive. Most people leave it on one setting for everything, which is probably the most commonly missed improvement available right now.
The underlying shift
Prompting used to be about scripting how the model thinks. It is now about specifying the problem well: the outcome, the evidence available, the constraints, and what a good answer looks like. The forty-page prompt full of reasoning scaffolding is finished. Context and clear success criteria replaced it.
The prompts
Each of these is a template. Square brackets are yours to fill. The quality of what comes back depends almost entirely on how specifically you fill them.
1. The context block, which goes at the top of everything
This is the highest-leverage thing in this article. Write it once per client, paste it at the start of any thread, and every subsequent prompt in that conversation improves without further effort.
Context for everything in this conversation. Business: [what they sell, to whom, in which market] Price point: [actual numbers, and what that buys] Buyer: [role, age range, nationality mix, what triggers the purchase] Top three objections: [the real ones from sales calls, not the ones on the website] What has worked: [campaign, angle, or message with the result] What has failed: [same, with why] Tone: [two or three adjectives, plus one thing to never sound like] Do not use: [banned words, claims you cannot legally make, competitor names] Acknowledge this and wait for my first task.
The "top three objections" line does more work than everything else combined, because objections are specific to a business in a way that benefits and features never are.
2. Meta ad primary text
You are a direct response copywriter who writes for cold audiences and distrusts adjectives. Write 6 primary text variations for a Meta ad. Each variation must open on a different angle: 1. The most common objection, stated before it is raised 2. A specific number or cost 3. A situation the reader will recognise from their own week 4. What happens if they do nothing 5. A comparison against the obvious alternative 6. A direct statement of the offer with no setup Constraints: - First line under 90 characters, because that is what shows before the fold - No rhetorical questions as an opening line - No "unlock", "elevate", "seamless", "game-changing", "in today's fast-paced" - Do not claim results we have not agreed - Written for someone who has never heard of this business Output: a table with columns Angle, Primary text, First-line character count.
Six angles rather than six rewrites is the point. Ask for variations and you get the same message reworded, which tests nothing. Ask for different angles and you get something you can actually learn from when one wins.
3. Google Ads responsive search headlines
Write 15 responsive search ad headlines and 4 descriptions. Target keyword: [keyword] Search intent: [what the person is actually trying to do] Landing page promise: [the exact promise the page makes] Requirements: - Headlines 30 characters maximum, descriptions 90 maximum. State the count after each. - At least 3 headlines must contain the keyword naturally - At least 3 must contain a number - At least 2 must address a hesitation rather than a benefit - No exclamation marks - Every headline must be true given the landing page above Output: two numbered lists with character counts.
The last constraint prevents the most common and most costly failure, which is a headline that promises something the landing page does not deliver. That is a quality score problem and a conversion rate problem at the same time.
4. Competitor teardown
Below is the homepage and services copy of a competitor. [Paste it.] Analyse it as a strategist, not a reviewer. Tell me: 1. Who they are positioned for, and who they are quietly excluding 2. Which three claims are doing the most persuasive work 3. Which claims are unsubstantiated and could be challenged 4. What objection they are handling that we are not 5. What they are avoiding mentioning, and why that might be 6. The single angle we could take that they could not credibly copy Be specific. Do not summarise the copy back to me. Output: numbered, under 500 words total.
Point six is the one worth the exercise. Everything else is analysis. That one produces positioning.
5. Content brief for SEO and AEO
Build a content brief for an article targeting: [keyword or question] Market: [country or city] Reader: [who they are and what they already know] What we can say that nobody else can: [your data, experience, or access] The brief must include: - A working title phrased as the reader's question - 6 to 9 H2 headings, each phrased as a question a real person would ask - Under each heading, the 40 to 80 word answer that should appear first - 5 FAQ questions distinct from the H2s - Specific data points to source, with what each one needs to prove - 3 internal linking opportunities and the anchor text for each - What this article will contain that the currently ranking pages do not Do not write the article. Output: structured outline only.
The last requirement before the output line is the one I care about. If the model cannot answer it, the article does not need to exist, which is the same test that decides whether AI-assisted content holds its ranking.
6. Landing page diagnosis
Below is the copy from a landing page. [Paste it.] Traffic source: [channel and what the ad promised] Current conversion rate: [number] Offer: [what the visitor is being asked to do] Read it as a sceptical first-time visitor who arrived from that ad and has 8 seconds. Tell me: 1. What they understand in the first 8 seconds, quoted from the page 2. Where the ad promise and the page diverge 3. The three biggest reasons someone leaves without converting 4. Which section is doing the least work and should be cut 5. What is missing that the buyer needs before committing 6. One change likely to produce the largest single improvement Be blunt. Do not compliment the page. Output: numbered, with the specific text you are referring to quoted.
"Do not compliment the page" is not a joke instruction. Without it you get two paragraphs of praise before anything useful, and the praise is worthless because it is unconditional.
7. Objection mining from real enquiries
Below are [number] real enquiry messages and sales call notes. [Paste them.] Extract: 1. Every distinct objection or hesitation, grouped by theme, with a count 2. The exact words customers use for each, verbatim, not paraphrased 3. Which objections appear before pricing is discussed, and which after 4. Any question asked repeatedly that our website does not answer 5. Vocabulary customers use that differs from our marketing vocabulary Output: a table, ordered by frequency. Do not invent objections that are not in the source material.
This one produces better ad copy than any copywriting prompt, because it returns your customers' actual language instead of the model's approximation of it. Point five regularly finds a mismatch worth fixing across an entire site.
8. Lead qualification rubric
Build a lead scoring rubric for [business type] in [market]. Average deal value: [number] Sales cycle: [length] What made our best 5 clients good: [description] What made our worst 5 clients bad: [description] Produce: - 6 scoring criteria, weighted to total 100 - For each, what a 1, a 3 and a 5 looks like in observable terms - The score threshold below which a lead should not get a call - Three questions to add to our enquiry form that would let us score accurately - One criterion commonly used in this industry that we should ignore, and why Output: a table plus the three form questions.
The final criterion is deliberate. Every industry has a scoring habit that correlates with nothing, and being told which one yours is tends to be the most useful line in the output. Scoring is also where most of the value sits in reducing cost per qualified lead rather than cost per lead.
9. Repurposing without sounding repurposed
Below is an article I published. [Paste it.] Write 3 LinkedIn posts from it. Each must stand alone and make sense to someone who never reads the article. Post 1: built around the most counterintuitive point Post 2: built around a specific number and what it implies Post 3: built around a mistake I see people make, and the correction Constraints: - First line must work as a standalone hook, under 12 words - No hashtags, no emoji, no "thoughts?" ending - Written as one person talking, not a brand - 120 to 180 words each - Do not link to the article in the body Output: three posts separated by headings.
10. The critic pass
This is the highest return prompt in the list and it is one line. Run it on anything before it goes out.
Review what you just produced as a sceptical senior marketer who thinks it is mediocre. Identify: - Every claim that is vague, unsupported, or could apply to any competitor - Every sentence that could be deleted without losing meaning - Anything that sounds like AI wrote it - The single weakest part Then rewrite only the weakest part.
Models are markedly better at evaluating output than producing it first time. A generate-then-critique cycle in the same thread produces better results than any amount of effort spent perfecting the initial prompt.
Three habits that matter more than wording
Feed real data. Paste the actual enquiries, the actual ad performance, the actual competitor page. Every prompt above improves by an order of magnitude when the input is real rather than described. Marketers who get poor results are almost always describing their situation instead of showing it.
Iterate in the same thread. The second and third responses are reliably better because the accumulated context is better. Starting a fresh chat for each task throws away everything the model learned about your client.
Use delimiters. Mark pasted material clearly with section headings, triple quotes or XML-style tags. It sounds fussy and it measurably reduces the model confusing your instructions with your source material.
What not to use prompts for
- Statistics you will publish. Models produce plausible numbers with plausible sources, both invented. Every figure needs verifying against a primary source. This is the fastest way to damage your credibility.
- Regulatory or compliance questions. Advertising rules, licensing, anything with a fine attached. Check the regulator, not the model.
- Final strategic judgement. Models are useful for generating options and terrible at knowing which one fits a market they cannot see.
- Anything where you cannot evaluate the answer. If you could not tell a good output from a bad one in that domain, a prompt is not the tool you need.
Common mistakes
- Asking for "engaging" or "compelling" copy. These words mean nothing operationally. Specify the angle and the constraint instead.
- Accepting the first output. The first response is a draft. Treating it as a deliverable is the actual reason people think AI writing is bad.
- Prompting for volume. Thirty posts of nothing is worse than none.
- Not banning the vocabulary. A short list of forbidden words removes most of what makes output recognisable as machine-written.
- Building a prompt library and never updating the context. The template is the smaller half. The context is where the quality lives.
Where this sits in a real workflow
None of this replaces knowing what to make. It compresses the distance between deciding and having, which is a genuine advantage when you are running campaigns across several accounts and the constraint is your own hours.
The prompts above came out of exactly that pressure, applied across paid media and content work in the UAE market, and refined through a prompt engineering certification from IIT Madras that mostly taught me how much of the result depends on the input rather than the phrasing. If you want the wider picture of where AI genuinely helps in a marketing operation and where it is decoration, that is covered in what AI marketing automation actually costs.
Frequently asked questions
What makes a good marketing prompt?
A role, real context the model could not otherwise know, constraints that rule out the generic answer, and a specified output format. Most prompts supply only the task. The quality of your context matters far more than clever wording.
Does "think step by step" still work?
Depends on the model. Reasoning models do that internally, and OpenAI's guidance says the instruction is unnecessary for them while Anthropic recommends general reasoning instructions over prescriptive sequences. On non-reasoning models it still helps, with Wharton research finding gains around 11 to 13 percent.
Are examples in prompts still useful?
Yes, but for format alignment rather than teaching reasoning. Showing the model the exact shape of output you want works. Stacking examples to demonstrate how to think no longer improves reasoning models and can compete with their internal process.
How long should a prompt be?
Long enough to carry context and constraints, short enough that every line earns its place. A focused fifty-word prompt with a clear role, real context and a defined output format beats a five-hundred-word prompt scripting the model's thinking.
Is a prompt library worth building?
Yes for structure, no for output. A saved prompt stops you forgetting the constraints and format that make results usable. It cannot supply the client-specific context, which is what actually changes quality.
What should marketers not use prompts for?
Statistics you will publish without verifying, regulatory or compliance questions, final strategic judgement, and anything where you could not tell a good answer from a bad one.
What is the reasoning effort setting?
A dial on several 2026 models controlling how much internal reasoning is applied before answering. Low for mechanical work, high for analysis where errors are expensive. Most people leave it on one setting for everything.
The short version
Role, context, constraints, format. Feed real data rather than descriptions of it. Iterate in one thread. Ban the vocabulary that gives it away. Run a critic pass before anything ships.
The prompt template is the small part. What you put in the context block is the whole game, and it is the one thing nobody can copy from an article like this.
Want a context block built for your business?
Send me what you sell, who buys it and the three objections you hear most, and I will write the context block for your account so every prompt you run afterwards starts from something specific rather than generic. Message me on WhatsApp or LinkedIn.
Jabir K
AI digital marketing consultant in Dubai with 6 years of UAE experience and roughly AED 1.8M in managed ad spend. Prompt engineering certified through IIT Madras. Works with UAE SMEs on paid media, lead pipelines and AI search visibility.
- OpenAI guidance on prompting reasoning models
- Anthropic guidance on extended thinking and reasoning instructions
- Wharton Generative AI Labs, chain-of-thought effectiveness research
- PromptHub research on few-shot examples and reasoning models
- Industry analysis of prompting practice, mid-2026
Model behaviour and provider guidance change frequently. Techniques described here reflect practice as of September 2026 and should be tested against the specific model you use rather than assumed.