14 min read · 3,138 words
The AI prompts for marketers worth reusing are the ones that make the model disagree with you, not the ones that make it write faster. Most published prompt libraries optimise for the second thing, which is why almost none of them survive contact with a real week of marketing work. Below are the twelve that have survived mine.
I run a content operation with AI stitched through most of it, briefs, drafts, research summaries, the lot. Early on I collected prompts the way most marketers do, dozens of them, saved in a note, most beginning with some version of “act as an expert marketer.” A year in, I use twelve. Not because the rest failed outright. They failed quietly, by producing something plausible enough that I stopped checking it. A prompt that generates a paragraph is doing the cheapest thing a model can do. A prompt that finds the hole in your segmentation, or argues the case your CFO will make against your own plan, is doing the expensive thing, and the expensive thing is what a fifteen-year career in brand and communication has actually taught me to value.
Why are most published prompt libraries useless?
Because they are written to be shared, and volume shares better than judgement. A list titled “50 ChatGPT Prompts for Marketers” performs well precisely because fifty is a bigger number than twelve, and every one of those fifty is built to produce output: a caption, a headline, a subject line. Output is the least scarce thing a language model makes. You can generate a hundred taglines before lunch and be no closer to knowing which one is true.
McKinsey’s 2026 State of AI survey of 1,719 respondents across 97 countries found 80% report AI has improved their individual productivity, yet only 37% attribute any measurable EBIT impact to it, and just 6% qualify as what McKinsey calls AI high performers. That gap is the prompt library problem in one statistic. Everyone feels faster. Almost nobody can point to what changed because of it.
What actually makes a prompt worth keeping?
It survives a second use, and a third, because it changes a decision rather than filling a blank page. I test every prompt I write against one question: if the model’s answer agreed with me by default, would I have learned anything? If not, I rewrite it to force disagreement, a weak point, a missing proof, an objection I had not considered, before I let it produce anything I would call finished work.
This is where the jagged frontier research matters more than most marketers realise. The Harvard, Wharton and BCG field experiment, run across 758 BCG consultants, found those using AI completed 12.2% more tasks, worked 25.1% faster and produced work judged 40% higher quality, with below-average performers gaining the most, a 43% jump. But on a task designed specifically to sit outside AI’s actual capability, human consultants working alone got the right answer 84% of the time, while those using AI got it right only 60 to 70% of the time. The tool did not fail loudly. People stopped checking it, the researchers call it falling asleep at the wheel, and that is exactly the failure mode a prompt library optimised for output volume walks you straight into.
The table below is the index. Each prompt gets its full text and its failure mode further down.
| Prompt | What it does | Watch for |
|---|---|---|
| Positioning stress-test | Attacks a positioning statement as your most skeptical buyer | Softens back to encouraging if not told not to |
| Interrogating a brief | Surfaces unstated assumptions before work starts | Jumps to solutions if allowed to in the same pass |
| Pressure-testing a segmentation | Checks whether segments change any real decision | Validates neat-looking segments that change nothing |
| Red-team the plan | Argues the strongest case against your own plan | Gives token pushback unless asked to fully inhabit the skeptic |
| Competitor message teardown | Extracts a competitor’s real claim from public pages | Returns adjectives instead of claim structure |
| Research report to board slide | Converts a report into one action-ready slide | Restates findings instead of stating implications |
| Category entry points | Lists real buyer trigger moments in their own words | Drifts into abstract need-states with no brief in them |
| Customer interview questions | Drafts non-leading questions for a research call | Writes leading questions that flatter you |
| Case study into three formats | One case study becomes a one-pager, a post, a call script | All three collapse into the same opening sentence |
| Holding statement | A factless acknowledgment for an unfolding situation | Invents plausible detail to fill gaps |
| AI citation audit | Checks whether a page is structured to be extracted | Turns into a generic SEO review |
| Remove the AI tells | Strips the tics that mark a draft as generated | Adds folksy filler instead of cutting structure |
The prompts for pressure-testing strategy and positioning
1. The positioning stress-test
What it is for: before a positioning statement is locked, this forces the model to attack it from the most skeptical, well-informed buyer you can imagine, rather than validate it.
Here is our positioning statement: [paste it]. Take the position of our most skeptical, well-informed buyer, someone who has been sold to before and has evaluated this category for two full budget cycles. List the five reasons they would not believe this claim, ranked by how damaging each one is. For each reason, tell me whether the fix is a proof point we are missing, a claim we should drop, or a claim that simply is not true yet. Do not soften the list to be encouraging.
Failure mode to watch: leave out the last line and the model quietly drifts back into cheerleading within two exchanges.
2. Interrogating a brief
What it is for: before anyone starts building against a creative or campaign brief, this surfaces what the brief assumes without saying so.
Here is a brief: [paste it]. Read it as a suspicious strategist, not a builder. List every assumption it makes without stating it as an assumption, about the audience, the competitive set, the budget, or what success looks like. Then list the three questions you would ask the brief owner before starting work, ordered by how much each answer would change the output. Do not propose solutions yet.
Failure mode to watch: let it propose solutions in the same pass and it stops asking hard questions and starts producing ideas nobody asked for.
3. Pressure-testing a segmentation
What it is for: before budget follows a segmentation model, this checks whether the segments are genuinely distinct or just labelled differently.
Here are our proposed segments, with the defining attributes of each: [paste]. For every pair of segments, tell me the single sales or marketing decision that would actually differ between them. If you cannot name one, say so plainly. Then identify which segment is not really a segment, but a subset of another one wearing a different label. Assume I will be asked to defend this in front of a CFO who thinks segmentation is jargon for having an opinion.
Failure mode to watch: without the CFO framing, it validates segments that look analytically neat and change nothing.
4. The red-team prompt
What it is for: before a plan goes to leadership, this argues the counter-case with full commitment instead of a token objection.
Here is the plan I am about to take to my leadership team: [paste it]. Argue against it as if you were the most respected skeptic in the room, someone whose objections have killed plans like this before. Build the strongest possible case that this is the wrong plan, using only points a smart, well-informed person could actually raise, not strawmen. Then tell me which of your own objections you find most convincing, and why.
Failure mode to watch: ask it to poke holes and you get polite pushback; asking it to fully inhabit the skeptic is what produces a real objection.
The prompts for research and customer intelligence
5. Competitor message teardown
What it is for: reading a competitor’s public materials, this extracts the claim they are actually making rather than the adjectives wrapped around it, useful the week a Zoho or a Freshworks changes its homepage and your sales team asks what it means.
Here is [competitor]’s homepage, latest product page and last two press releases: [paste or links]. Extract the claim they are actually making about why a buyer should choose them, not the language every company in this category uses. Tell me what they are implicitly saying about us or the incumbent by what they emphasise or avoid. Then tell me which one of our claims becomes weaker the day they say this convincingly.
Failure mode to watch: without the instruction to strip generic language, the teardown returns a list of adjectives instead of a claim.
6. Research report into a board slide
What it is for: converting a dense report into the one slide a board with ninety seconds on the agenda can actually act on.
Here is a research report: [paste or summary]. I need one slide for a board with ninety seconds for this topic. Give me a headline that states the implication, not the finding, not “AI adoption is rising” but what that means for us. Three supporting data points, each tied to a source. One recommended action. One number that would change your recommendation if it moved, and in which direction. No more than forty words outside the data points.
Failure mode to watch: skip “implication, not finding” and the slide restates the report instead of telling the board what to do about it.
7. Category entry points
What it is for: generating the actual moments that push a buyer into the market, the input a campaign brief can be built from, not a persona document. This is the argument I have made before about why category entry points beat awareness as a metric, and the prompt is how I generate the raw list.
We sell [category] to [buyer type]. List the specific situations, moments or triggers that would cause a real buyer to start actively looking for a solution like ours, phrased the way the buyer would think it, for example “our current vendor missed a delivery window twice this quarter,” not “supply chain optimisation needs.” Rank them by how often you estimate they occur against how much budget authority typically exists at that moment. Flag which of these we currently have zero content or campaign against.
Failure mode to watch: drop the buyer-phrasing constraint and you get abstract need-states no brief can be written from.
8. Customer research interview questions
What it is for: drafting interview questions for a customer research call that surface what happened, not what the customer thinks you want to hear, which is the difference between a real journey map and one that flatters the team that built it.
I am interviewing a customer who [context: churned last quarter, just renewed, chose a competitor after evaluating us]. Write ten interview questions that would surface what actually happened, not what they think I want to hear. Avoid any question I could answer for them through how I phrase it. Include at least two questions built to surface a specific moment or event, not an opinion. Tell me which three questions are most likely to get a defensive, guarded answer, and suggest a softer way to ask each.
Failure mode to watch: left unprompted, the model writes the leading questions a marketer would ask, and gets back flattering nonsense.
The prompts for content, citation and rewriting
9. Case study into three formats
What it is for: one approved case study, turned into three genuinely distinct assets instead of the same paragraph reformatted three times.
Here is our full case study: [paste it]. Produce three versions. First, a hundred-word version for a sales one-pager, opening with the number, not the customer’s name. Second, a LinkedIn post written in first person from the account team’s point of view, using one concrete detail the case study itself does not lead with. Third, three lines a salesperson could say out loud on a call, written to be spoken, not read. Do not reuse the same opening sentence across the three.
Failure mode to watch: without the no-repeat rule, all three collapse into the same sentence wearing different formatting.
10. The holding statement
What it is for: drafting a factless acknowledgment for a situation that has not yet been confirmed, so the first thing published is not a guess.
Something has happened: [one line, whatever is confirmed, nothing more]. Write a holding statement of no more than sixty words that acknowledges the situation, states our first priority, commits to an update by a specific time, and contains zero facts beyond what I have given you, no speculation about cause, blame or scale, even if my prompt implies more. Flag anything in what I wrote that reads as an assumption rather than a confirmed fact.
Failure mode to watch: models fill gaps helpfully by default, so without an explicit ban on speculation it invents plausible detail that was never confirmed.
11. The AI citation audit
What it is for: checking whether a page is actually structured to be lifted by an AI answer engine, the practical extension of the case I made about how brands get cited by AI engines.
Here is a page from our site: [paste the text or URL content]. Acting as an AI answer engine deciding what to extract, identify the single passage you would quote if asked a question this page could answer, and tell me if it is under sixty words and self-contained. List every sentence that depends on an earlier sentence to make sense, these cannot be extracted alone. Tell me what specific number, named source or comparison is missing that would make this page more citable.
Failure mode to watch: skip the dependency check and it produces a generic SEO review, not an extraction-readiness audit.
12. Remove the AI tells
What it is for: stripping the tics that mark a draft as generated before it goes out under a real name, including mine.
Here is a draft: [paste it]. Rewrite it to remove anything that reads as generated rather than written: rule-of-three lists where a real writer would pick one, symmetrical sentence pairs, throat-clearing openers, and any claim so hedged it says nothing. Vary sentence length noticeably. Where the draft makes a safe, general claim, either sharpen it into a specific one or cut it entirely. Do not add new adjectives to compensate for what you cut.
Failure mode to watch: ask it to sound more human instead and it adds folksy filler rather than removing the actual structural tics.
So what should actually go in your prompt library?
Fewer prompts than you think, and almost none of them should ask for finished copy on the first pass. Every prompt above earns its place by making the model take a position I have to respond to, agree with, argue against, or go and find evidence for, before it is allowed to write a single sentence I will publish. That is a slower first draft and a much shorter revision cycle, which is the trade a fifteen-year career in brand work has taught me is almost always worth making.
The gap between adoption and discipline is visible in the numbers. BCG’s survey of 300 CMOs globally, published in June 2026, found that 96% say AI is driving an end-to-end transformation of their function while only 8% are running campaigns where multiple AI agents operate autonomously. Nearly everybody has been transformed and almost nobody has changed how the work is actually done. India’s own data backs the direction. BCG’s 2026 CMO survey found 53% of Indian CMOs expect AI use cases to generate 5 to 9% incremental topline growth, and 57% said the marketing function itself now funds AI investment, both ahead of the global figures BCG reported. Gartner’s 2026 CMO Spend Survey of 401 marketing leaders found CMOs now allocate 15.3% of marketing budgets to AI, while only 30% consider their organisation ready to scale it. The money is arriving faster than the judgement to spend it well, and a prompt library is one of the few places an individual marketer can close that gap without waiting for the organisation to catch up.
The cost of not closing it is not hypothetical. Research from BetterUp Labs and Stanford, covered in Harvard Business Review, found 41% of employees had received AI-generated “workslop,” content that looks finished but carries no real thought, and each instance cost the receiving colleague nearly two hours to fix. A related MIT Media Lab estimate found 95% of organisations investing in AI report no measurable return. That is what happens when a prompt library is optimised to fill the page. The prompts that pay are the ones that make the model argue with you, not write for you, because copy is cheap and judgement is not.
Frequently asked questions
What is the single most useful AI prompt for a marketer to start with? The positioning stress-test. Most marketers already have a positioning statement sitting somewhere, and testing it against a skeptical buyer persona surfaces gaps faster than any other single prompt, usually within one exchange, before you spend budget building campaigns on a claim that will not hold.
Do these prompts work the same across ChatGPT, Claude and Gemini? The structure works everywhere; the discipline that makes it work does not come free. All three models default to being agreeable, so the instructions forbidding softened answers, invented facts or premature solutions matter more than which model you run the prompt in.
How do I know if a prompt is worth adding to my own library? Ask whether the model’s answer taught you something you did not already believe. If a prompt’s output would have agreed with you regardless of what you fed it, it is not testing anything, it is decorating a decision you had already made.
Should AI write the first draft of marketing copy at all? It can, but the prompts that matter more sit before that draft: interrogating the brief, stress-testing the positioning, checking the segmentation. A well-briefed first draft from a human writer using a sharpened brief usually beats an AI first draft from a vague one.
What is the biggest risk in using AI prompts for strategic marketing work? Falling asleep at the wheel, the term researchers use for trusting confident-sounding AI output on tasks it is not actually good at. The Harvard and BCG field experiment found this cut task accuracy from 84% to as low as 60% on exactly this kind of work.
My honest view is that a prompt library is a poor substitute for judgement, and a genuinely good one is judgement written down so you do not have to rebuild it from scratch every Monday. Which of your own prompts, if you are honest, have you kept only because you never checked whether the output was any good?
