Copywriting with AI produces sales copy worth using when the prompt supplies what the model cannot know: who the buyer is, what problem they are trying to solve, what the offer actually includes, what proof exists, and what single action the reader should take. Seven prompt rules cover those inputs, and an eighth constraint keeps the result honest: the model may only use urgency, scarcity and results claims that you can document.
This article walks through each rule with the prompt text that implements it and the reason it is there, then combines them into one master prompt, shows an illustrative before-and-after, and ends with a revision routine for improving a draft instead of regenerating it.
Why generic sales prompts produce generic copy
A prompt like “write persuasive sales copy for my course” fails because it leaves every important decision to the model. With no buyer, no offer details and no proof, the model fills the gaps with the most common patterns in sales copy it has seen: exclamation marks, “transform your life”, invented testimonials and a countdown. The result reads like every other page because it is assembled from every other page.
The model is good at structure, rhythm and variation. It cannot know your customer, your offer or your evidence. Anthropic’s prompting guidance puts it plainly: be explicit about what you want and explain the reason behind instructions, because the model generalizes better from the why than from a bare rule. Each rule below is a piece of context that only you can provide. For a broader look at this failure mode, see why ChatGPT gives generic responses.
The seven rules, one at a time
Rule 1: Give the model a role and a reader
A role tells the model which conventions to follow; a reader tells it whom to write for. The original version of this article suggested a role like “a direct response copywriter with 20 years of experience”. The role is useful, but the years add little; what improves the output is a precise description of the reader.
You are a direct-response copywriter. You write plain, specific copy
and you never claim results the client cannot document.
The reader: [role or life situation], who is trying to [goal] and has
already tried [alternative] without getting [result]. They are
skeptical of [common objection].Writing the honesty constraint into the role itself means it applies to every sentence that follows. More on how roles shape output is in the guide to role prompting.
Rule 2: Describe the problem and the desired outcome in the buyer’s words
Sales copy works when the reader recognizes their own situation in it. The best source for that language is not your imagination but what customers actually say: support tickets, sales call notes, reviews, survey answers.
<customer_language>
[Paste 5-10 real quotes from reviews, tickets or calls describing the
problem and what they wanted instead. Remove names.]
</customer_language>
Use the phrases and situations in <customer_language> to describe the
problem and the outcome. Do not exaggerate the pain beyond what these
quotes describe.The last sentence replaces the old “anatomy of pain and pleasure” rule, which tends to push models into melodrama. Buyers notice when their mild inconvenience is described as a crisis.
Rule 3: Specify the offer exactly
The offer is what the buyer gets, on what terms. A model that is told “an online course” will guess at the rest; a model given the facts can make them persuasive without inventing anything.
<offer>
What it is: [...]
What is included: [modules, hours, files, support, access period]
Price and payment terms: [...]
Guarantee or refund terms, exactly as written in our policy: [...]
What it does NOT include: [...]
What makes it different from [main alternative]: [...]
</offer>The “does NOT include” line is there to prevent the model from implying extras such as live support or lifetime access that the offer does not have. Pasting the guarantee exactly as written matters for the same reason: a model paraphrasing a refund policy can change its meaning.
Rule 4: Supply the proof, and forbid anything else
Proof is anything that makes a claim believable: documented results, real testimonials, credentials, a demonstration. It is also where AI copy most often goes wrong, because models readily generate realistic-sounding numbers and quotes.
<proof>
[Documented results with their source, real testimonials with
permission to use them, credentials, case studies.]
</proof>
Only use proof from <proof>. Do not create statistics, percentages,
testimonials, customer counts or ratings. If a sentence would be
stronger with proof that is not here, write [PROOF NEEDED: what kind]
instead.This is a legal constraint, not just a stylistic one. The FTC’s advertising guide for small businesses explains that advertisers need a reasonable basis for objective claims before they publish them, and that testimonials should reflect what typical customers experience; a generic “results may vary” line is not enough to fix a misleading one. AI-generated testimonials raise further issues, covered in the FTC checklist for AI-generated testimonials and reviews.
Rule 5: Use real urgency only
Urgency gives the reader a true reason to act now rather than later, and it is only legitimate when the reason exists. The original article recommended prompting for lines like “Only 50 slots left” and “Price returns to normal in 4 hours” regardless of whether they were true. That advice is withdrawn here.
The FTC staff report Bringing Dark Patterns to Light lists false low-stock messages, false high-demand messages, baseless countdown timers that reset, and false limited-time offers among deceptive design patterns. A model will happily produce all of them if asked for “intense urgency”.
<real_deadlines>
[Only true constraints: cohort start date, enrollment close date,
actual seat limit, price change date already announced, seasonal
relevance. Leave empty if none.]
</real_deadlines>
Use urgency only from <real_deadlines>. If it is empty, give the reader
a reason to act now based on the cost of the problem continuing, not on
a deadline. Never invent stock levels, seat counts, countdowns or
expiring prices.The fallback in the second paragraph is the useful part. When there is no deadline, the honest reason to act is that the problem keeps costing the buyer something every week it continues, which you can describe using the customer language from rule 2.
Rule 6: Set tone and reading level, and show an example
Tone instructions written as adjectives (“enthusiastic and urgent”) are vague; each model interprets them differently. Two things work better: a concrete reading target and an example of copy whose voice you want.
Tone: plain and direct. Short sentences, common words, active voice.
Aim for copy a busy reader understands on one pass.
No exclamation marks. No words like revolutionary, game-changing,
ultimate, unlock or skyrocket.
Match the voice of this example (not its content):
<voice_example>
[Paste 100-200 words of your own copy that performed well or that
sounds like your brand.]
</voice_example>Examples are among the most reliable ways to steer format and tone, according to both Anthropic’s and OpenAI’s prompting guides. The banned-words list is a short, targeted use of negative instructions; long lists of prohibitions work less well, as explained in negative prompting explained.
Rule 7: Choose a structure and one call to action
A structure gives the copy a logical order; the call to action (CTA) tells the reader exactly what to do next. AIDA (Attention, Interest, Desire, Action) is a long-standing copywriting framework and a reasonable default, but it is not the only option: PAS (Problem, Agitate, Solution) suits short pain-led copy, and a simple “situation, offer, proof, action” order suits readers who already know the problem.
Structure: [AIDA / PAS / situation-offer-proof-action].
Label each section in your draft so I can see the structure, then
give me a clean version without labels.
Call to action: one action only: [e.g. "Enroll in the March cohort"].
Use the same wording on every button or link. No secondary CTA.Asking for labeled sections first makes the structure checkable: if the “Desire” section contains no proof, you can see that immediately. One CTA prevents the common AI habit of ending with three different invitations.
The master prompt
The master prompt combines all seven rules into one reusable template. Fill in the tagged blocks once per offer and keep the instructions fixed, so differences between drafts come from your inputs rather than from rewording the prompt.
You are a direct-response copywriter. You write plain, specific copy
and never claim results the client cannot document.
Task: write [sales page section / email / ad] copy of about [N] words.
<reader>[rule 1]</reader>
<customer_language>[rule 2]</customer_language>
<offer>[rule 3]</offer>
<proof>[rule 4]</proof>
<real_deadlines>[rule 5]</real_deadlines>
<voice_example>[rule 6]</voice_example>
Instructions:
1. Describe the problem using the reader's own language. Do not
exaggerate it.
2. Present the offer exactly as specified, including what it does not
include if that prevents a wrong assumption.
3. Use only the proof provided. Mark gaps as [PROOF NEEDED].
4. Use only real deadlines. If none, use the cost of waiting.
5. Tone: plain, short sentences, no exclamation marks, no hype words.
Match the voice example.
6. Structure: [framework]. One CTA: "[exact CTA]".
Output:
A) Draft with section labels.
B) Clean final copy.
C) A list of every factual claim in the copy and which input block
it came from.Output C is the step most people skip and the one that catches the most problems. If any claim cannot be traced to an input block, the model made it up.
Before and after: an illustrative example
The offer below is invented for illustration: a six-week online cohort that teaches small accounting firms to automate client onboarding paperwork. The real deadline is that enrollment closes when the cohort starts; there is no seat limit. The only documented proof is a quote from a past participant who gave permission to use it.
Before, from “write urgent, persuasive sales copy for my automation course”:
STOP wasting hours on paperwork! Our revolutionary course has helped
over 5,000 accountants save 20+ hours a week. Only 7 spots left, and
the price doubles at midnight! Don't miss out on the opportunity to
transform your firm forever. Enroll NOW!Every number in that draft is invented: the participant count, the hours saved, the remaining spots and the price change. Each one is a claim the seller would need to substantiate, and the scarcity lines match patterns the FTC report describes as deceptive.
After, from the master prompt with the real inputs:
If every new client still means chasing the same six documents by
email, this cohort is for you.
Over six weeks, you build an onboarding flow for your own firm: an
intake form, automatic document requests and a checklist that shows
what is still missing. You leave with the flow running, not a set of
notes.
"We stopped sending the same reminder email three times per client."
- Past participant, two-partner firm
The cohort is live, with weekly sessions; it is not a self-paced
library. Enrollment closes when the March cohort starts.
Enroll in the March cohortThe after version is calmer and more convincing: the problem is specific, the outcome is concrete, the proof is real, the deadline is true, and there is one action. It also tells the reader what the offer is not, which filters out buyers who wanted a self-paced course and would have asked for a refund.
Revising a draft instead of regenerating it
Regenerating from scratch throws away what worked. Targeted revision prompts improve one dimension at a time and keep the rest stable.
- Objections: “List the three objections this reader is most likely to have, based on <reader> and <customer_language>. For each, show where the copy answers it, or add one sentence that does, using only the inputs.”
- Specificity: “Find every sentence that could appear on a competitor’s page unchanged. Rewrite each with a detail from <offer> or <customer_language>.”
- Cut: “Remove 20% of the words without losing any fact or the CTA.”
- Headline options: “Write five headlines, each built on a different element: the problem, the outcome, the proof, the format, the deadline.”
The objections prompt is the updated form of the original article’s “fine-tuning” step. The difference is the grounding: objections come from your customer language, and answers come only from your inputs. When the copy is ready, test competing versions against each other rather than trusting any single draft; the landing page copy prompts show how to apply the same inputs to a full page.
