The Best AI Prompts for Sales Copy (Framework-Locked, Not Fluff)

Paste "write sales copy for my product" into ChatGPT and you get generic mush: adjective-heavy, allergic to a real claim. The best AI prompts for sales copy don't ask for copy. They assign a proven structure (PAS, the offer stack) and hand the model a swipe file to mirror.

The best AI prompts for sales copy give the model three things: a copywriting framework to fill in (like Problem-Agitate-Solution or a bullet-and-benefit structure), a real swipe example to match in tone and rhythm, and specific inputs about the offer, the reader, and the objection being overcome. Skip any of the three and the output reads like AI wrote it.

Sales copy is the one place AI output gets tested immediately. It either earns a click and a sale, or it sits there ignored. What decides the outcome isn't which model you use: Claude, ChatGPT, and Gemini can all produce strong drafts. It's whether you fed the model a structure to build inside and a real example of the voice you're going for. Structure keeps the logic tight so the reader can follow it. The swipe keeps the voice from defaulting to AI-generic. Skip either one and you end up editing a first draft from scratch anyway, which erases the time you thought AI would save you.

Why does "write me sales copy" produce generic output?

Because a vague prompt gives the model nothing to push against, so it defaults to the statistical average of every sales page it's ever seen, which is bland by definition.

Large language models predict the most probable next word given what you've told them. Tell them almost nothing and the most probable output is the safest, most generic version of sales copy that exists: broad claims, soft benefits, a CTA that says "get started today." That's not the model being lazy. It's the model doing exactly what you asked, which was nothing specific.

Three things fix this, and all three show up in every prompt later in this piece:

• A framework. Give the model a shape to fill (Problem-Agitate-Solution, a value-stack offer, a bullet formula) instead of an open blank page.
• A swipe example. Show it the cadence and voice you want, not just describe it.
• Real inputs. The actual offer, the actual reader, the actual objection. Not "professionals" as the audience. The 41-year-old operator who's tried three tools already and doesn't believe a fourth will work.

Miss the framework and the copy has no spine. Miss the swipe and it sounds like every other AI draft on the internet, because in a real sense it is. Miss the specific inputs and you get benefits instead of a reason to believe them. If you're building out prompts for other marketing tasks too, our broader library of ChatGPT prompts for marketing covers the same principle across formats beyond sales copy alone.

How do you build a prompt around the PAS framework?

You name the three beats explicitly in the prompt (Problem, Agitate, Solution) and tell the model what belongs in each one, so it can't blend them into one paragraph of vague positivity.

Problem-Agitate-Solution works because it mirrors how a reader actually processes a pitch: first they need to see their problem stated back to them accurately, then they need to feel what it costs to leave it unsolved, then and only then are they open to a solution. Skip straight to the solution, which is what a lazy prompt produces, and you're selling to someone who hasn't agreed yet that they have a problem worth solving. For the mechanics of the framework itself, our breakdown of the Problem-Agitate-Solution formula is worth reading before you prompt with it, so you know what "good" looks like when you're editing the output.

In the prompt, spell out what each beat is for:

• Problem: state the reader's problem in their language, specific enough that they think "that's literally me," not a category like "time management" that could describe anyone.
• Agitate: show the cost of inaction using logic, not invented statistics. What happens in three months if nothing changes? What's the reader already quietly worried about?
• Solution: introduce the offer as the mechanism that closes the gap, framed around what changes for the reader, not a features list.

A worked example. Say you sell a course on running Meta ads for local service businesses. A generic prompt gets you: "Struggling with your ad spend? Our course teaches you everything you need to know to succeed with Facebook ads." A PAS-structured prompt with real inputs gets you something closer to: "You've burned $4,000 on ads this year and you still can't tell which dollar did anything. Every month you don't fix targeting is another month a competitor with a worse offer out-advertises you, because their funnel is dialed in and yours isn't. This course rebuilds your campaign structure from the audience up, so you stop guessing which variable is burning the budget." Same offer, same model, completely different pull, because the prompt forced the structure and gave it something real to agitate with.

How do you use a swipe file so the AI matches your voice?

You paste 200 to 400 words of copy that already sounds right, yours or copy you genuinely admire, into the prompt and instruct the model to match its sentence rhythm and tone, explicitly telling it not to copy the words or claims.

This is the step most people skip, and it's the one that does the most work. A framework tells the model what to say and in what order. A swipe tells it how to say it: short sentences or long, blunt or warm, contractions or formal, second person or third. Without a swipe, the model reaches for its default voice, which is a polished, hedge-everything corporate tone that no real copywriter would ship.

What makes a good swipe example:

• It's actually good. Pull from your own best-performing email, page, or ad, or from a piece of copy you'd genuinely study. Feed the model mediocre copy and you get mediocre output that at least sounds like you.
• It's close in format. Swiping an email's voice for a landing page headline works fine. Swiping a 3,000-word VSL script for a 40-word ad won't transfer cleanly, the rhythm doesn't scale down.
• It's long enough to have a pattern. A single sentence doesn't give the model enough signal. Two or three paragraphs does.

Be explicit in the prompt that you want the pattern, not the content: "Match the tone, sentence length, and rhythm of this example. Do not reuse its specific words or claims." Without that instruction, some models will lift phrases wholesale, which is a problem if the swipe isn't yours to begin with.

How do you prompt for an offer stack instead of a single pitch?

You list every component of the offer (the core product, the bonuses, the guarantee, the urgency) as separate line items in the prompt and tell the model to stack value before it ever mentions price.

An offer stack works because a single price feels arbitrary, but a price attached to five named things, each with its own value, feels like a deal even before a discount shows up. Most AI-generated offers fail here because the prompt just says "make this offer sound good," so the model writes one paragraph of adjectives instead of building the stack piece by piece. For the underlying logic of why stacking works, the Grand Slam offer framework is the clearest breakdown of the mechanism, and it's worth prompting from once you understand why each piece earns its place.

In the prompt, feed the model the stack as structured inputs, not prose:

• Core offer: what they get, one sentence.
• Bonuses: list each one with what it's worth to the reader specifically, not a generic dollar tag.
• Guarantee: the actual risk reversal, stated plainly.
• Urgency or scarcity: only if it's real. A fake countdown is easy to spot and it's the fastest way to burn trust with an operator audience.

Then ask the model to write the stack in ascending order (core offer first, then each bonus building on it), close with the guarantee, and only introduce price after the full value is on the page. That sequencing is the entire trick. An AI given the raw stack elements and told to hold the price until the end will do it. An AI just told "write an offer" will lead with the price, because that's the fastest way to answer a vague prompt.

How do you get sharper headlines, bullets, and CTAs out of AI?

You ask for volume, not a single answer: request options built around different angles, then pick and combine, because the first few outputs from any model are almost always the safest and least interesting ones.

Headlines, bullets, and CTAs are short enough that AI can produce dozens in the time it takes you to write three by hand, which changes how you should prompt for them. Instead of "write me a headline," ask for headlines built around named angles. For a course on freelance copywriting, that prompt might produce:

• Curiosity: "The One Sentence That Gets Freelance Clients to Reply First"
• Benefit: "Land Your First $2,000 Copywriting Client in 30 Days"
• Number-led: "7 Cold Emails That Actually Got a Yes"
• Mistake-named: "Why Your Portfolio Isn't Landing You Clients (It's Not the Work)"
• Direct claim: "Learn to Write Copy Clients Pay For, Not Copy That Sits in a Folder"

That's five distinct angles from one prompt instead of five versions of the same headline. Pick whichever fits the platform and audience temperature, or blend two.

For bullets, the mechanism is different: a good bullet creates curiosity about a benefit without fully explaining it, which is its own skill separate from headline writing. Tell the model explicitly to hold back the explanation. Curiosity is the point, not clarity. "Write 10 bullet points about the benefits" produces flat, fully-explained bullets that kill the curiosity a good bullet needs. "Write 10 bullets that make the reader curious about [BENEFIT] without fully explaining how it works" produces something closer to what actually gets read.

For CTAs, give the model the actual next step, not "buy now" as a category, but what literally happens when someone clicks, and ask for versions at three temperature levels: direct ("Start today"), low-friction ("See how it works"), and specific-outcome ("Get my first draft in 10 minutes"). Test across those temperatures instead of guessing which one your audience wants.

How do you edit AI sales copy so it doesn't read like AI wrote it?

You strip a specific set of tells the model defaults to (certain words, rule-of-three lists, hedge phrases, and unnaturally even sentence length) and then read the draft out loud, because your ear catches rhythm problems your eyes skip past.

Every model has default patterns that show up regardless of how good your prompt is. Watch for:

• Vocabulary tells. Words like "unlock," "elevate," "seamless," "robust," "dive in," and "game-changer" show up because they're statistically common in training data, not because they're the right word. Ban them in the prompt and cut any that slip through anyway.
• The rule of three. AI defaults to three-item lists and three-part sentences ("faster, easier, and more effective") because that pattern is everywhere in its training data. Real writing doesn't hit three every time. Break the rhythm on purpose.
• Hedge phrases. "Can help," "may improve," "designed to." A model trained to avoid liability drifts toward soft claims. Direct response copy makes a direct claim.
• Even sentence length. AI output often runs sentences of similar length back to back, which reads smooth but flat. Real copy snaps: a long sentence that sets up context, then a short one that lands the point.

Our full breakdown on writing copy with AI without sounding like AI goes deeper on the edit pass itself if this is the step where your drafts keep stalling. The short version: treat the first AI output as a structural draft, not a finished one. The framework and swipe get you most of the way there. The last stretch is you, reading it out loud, cutting whatever sentence you'd never actually say.

What's the exact prompt to use?

Below is the prompt that combines everything above: a named framework, a slot for your swipe example, structured offer and reader inputs, and explicit rules against the tells that make copy read as AI-written. Fill every bracket. Don't skip the swipe slot. An empty one is the difference between this working and this producing the same generic draft you started with.

You are a direct-response copywriter. Write [SALES PAGE SECTION / EMAIL / AD] copy using the Problem-Agitate-Solution framework.

OFFER:
- What you're selling: [PRODUCT OR SERVICE]
- Price: [PRICE OR RANGE]
- Core mechanism: [HOW IT ACTUALLY WORKS, 1-2 SENTENCES]
- Real proof you have: [RESULTS, TESTIMONIALS, GUARANTEES, CREDENTIALS, ONLY WHAT'S TRUE]

READER:
- Who they are: [SPECIFIC AUDIENCE, NOT "EVERYONE"]
- Problem they're living with right now: [THE DAILY, CONCRETE VERSION]
- What they've already tried that didn't work: [1-3 THINGS]
- The objection most likely to stop them from buying: [NAME IT]

STRUCTURE (follow this order, don't label the beats in the output):
1. Problem: open on the reader's specific problem, in their language.
2. Agitate: show what it costs to leave this unsolved for another [TIMEFRAME]. Use logic and consequence, not invented statistics.
3. Solution: introduce [PRODUCT] as the mechanism that closes the gap, not a features list.
4. Proof: back the claim with the real proof listed above, nothing invented.
5. Objection: name and answer the objection above in one paragraph.
6. CTA: one clear next step: [SPECIFIC ACTION].

VOICE: match the tone, sentence length, and rhythm of the swipe example below. Do not reuse its specific words or claims, only its cadence.
[PASTE YOUR SWIPE EXAMPLE HERE: 200-400 WORDS OF COPY YOU ALREADY LIKE]

RULES:
- No em dashes.
- No invented statistics, studies, or numbers I didn't give you.
- Vary sentence length. Short sentences should actually be short.
- Do not use: unlock, elevate, supercharge, game-changer, seamless, robust, dive in, unleash.
- Target length: [WORD COUNT] words.

Output the copy only. No explanation before or after.

Run it once, then run it again with the same inputs but ask for a second version that opens differently. Compare both openings against your swipe and keep whichever one actually sounds like it.

Where this breaks

AI can't generate proof, only you can. It will happily produce a testimonial-shaped sentence or a specific-sounding result if you let it, and that's the fastest way to end up with a claim you can't back up: copy that's a liability instead of an asset. Feed it only the proof you actually have. If you don't have enough proof yet, that's a signal to fix before you fix the copy.

Swipe files transfer cadence, not judgment. If you paste in copy that isn't in your voice because you admired it rather than because it sounds like you, you'll get sharper copy that still isn't yours. The fix isn't a better prompt. It's picking a better swipe.

Long-form output degrades. A single prompt asking for a full 2,000-word sales page will drift: the model repeats itself, loses the thread on the objection you named, and the back half comes out noticeably weaker than the front half. Prompt section by section (headline, then problem and agitate, then offer stack, then close) for anything longer than an email or a short ad, and stitch it together yourself.

And once you know these frameworks well enough to know when breaking them works better, the AI won't tell you that. It will keep filling in PAS or the offer stack exactly as instructed, because that's what you asked for. Frameworks are a floor for someone still building the skill, not a ceiling for someone who's outgrown needing one.

Frequently Asked Questions

What's the single best AI prompt for sales copy?

There isn't one universal prompt, because "best" depends on your offer, your reader, and the format you're writing in, whether that's an email, a page, or an ad. What stays constant is the shape: a named framework, a real swipe example, and specific offer and reader inputs. The template above is that shape built for PAS. Adapt the sections for other formats.

Which AI model writes the best sales copy, ChatGPT, Claude, or Gemini?

The model matters less than the prompt structure. All three can produce strong drafts when you give them a framework and a swipe example, and all three produce generic output when you don't. If you're choosing between them for copywriting specifically, Claude vs ChatGPT for copywriting covers the practical differences worth knowing.

Can AI write a full sales page from one prompt?

It can draft one, but expect to edit heavily, especially the proof and personalization sections, which the model can't generate on its own. Longer pages also tend to weaken toward the end of a single output. You'll get a better result prompting section by section and assembling it yourself than asking for the whole page at once.

Do I need to know copywriting frameworks to use these prompts?

Not in depth, but knowing the basic skeleton of whatever framework you're using helps you catch it when the AI cuts a corner, like blending the agitate and solution beats together instead of keeping them separate. Once you've run a few of these prompts, you'll start recognizing the shape well enough to spot it missing.

How long should my swipe example be?

Two to three paragraphs, roughly 200 to 400 words, is usually enough for the model to pick up a consistent pattern without so much text that it starts pulling in unrelated tangents. Pull it from your own best-performing copy when you have it, or from copy you'd genuinely study if you don't yet.

Will Google penalize sales copy written with AI prompts?

No. Google's guidance targets low-quality, unhelpful content regardless of how it was produced, not the fact that AI was involved in writing it. Whether Google penalizes AI content covers this in more depth if it's a concern before you publish AI-assisted copy anywhere.

Put these prompts to work

None of this replaces reps. The framework gets you a structured draft, the swipe gets you a voice-matched draft, and the editing pass gets you something you'd actually publish. But knowing which framework to reach for and which swipe fits your offer is a judgment call that gets sharper the more copy you write and rewrite. It's faster with other operators doing the same reps next to you, comparing what's converting instead of guessing alone.

If you want to build this skill with people running real offers and comparing notes on what's actually working, not theorizing about it alone, join Asset Academy on Skool.