EN
Ihor Nikolenko- Why the Prompt Writes the Fluff, Not the Model
- The Six Parts of a Strong Prompt
- Headlines by Formula, Not by Inspiration
- A Weak Prompt and a Strong Prompt, Side by Side
- Techniques That Kill Generic Copy
- Three Templates You Can Copy
- A Prompt Is a Consequence of the Method, Not a Replacement for It
- The Short Version
- Frequently Asked Questions
"Write a sales post about our course" — and the model produces text you could slap onto the site of any course on earth. Smooth, correct, empty. The owner concludes: AI writes fluff. In reality, the fluff was written by the prompt: it had no person the text was for, no fear it needed to remove, no outcome it needed to promise. The model filled the void with the most probable words — that's just how it works.
A marketing prompt isn't "the right magic words." It's the same thing as a copywriter's brief, just shorter and more precise: who it's for, what rung that person is on, what hurts, what you're promising, in what form, and what's off-limits. I've been putting together briefs like this for years, for human writers — now I hand them to a model instead, and the structure hasn't changed, only the speed has.
This article covers a six-part prompt structure, the headline formulas I use in the D.N.A. launch model, techniques that kill generic-sounding copy, and three ready-made templates you can copy today.
Why the Prompt Writes the Fluff, Not the Model
A language model continues text in the most probable way. If you give it little to work with, it defaults to the average of everything it has ever seen: an average audience, an average pain point, an average promise. That's where "effective solutions for your business" and "a unique opportunity" come from. It isn't laziness on the model's part — it's the absence, in the prompt, of whatever would make the text yours.
What's "yours" is a specific person with their own fear and desire, your real experience, your facts, your voice. The model knows none of that. All of it has to be in the prompt — or the text belongs to nobody.
The Six Parts of a Strong Prompt
1. Role. Who the model should be: "an experienced direct-response copywriter writing in English for business owners." The role sets the level and style, but it doesn't save you on its own — it's just the first part.
2. Context: who it's for and what rung they're on. The most important part, and the one most often skipped. Describe the person: who they are, what they do, what they've already tried. Name which rung of the awareness ladder they're standing on: just sensing the problem, aware solutions exist, comparing you to others, or ready to pay. Copy for "just sensing the problem" and copy for "comparing" are two completely different texts, and the model will only write the right one if you tell it which.
3. Fear and desire. One specific fear the text needs to remove, and one desire it leads toward. Not "wants more customers," but "is afraid of paying for another site that stays silent; wants to see the first lead from an ad before the budget runs out." Fear and desire are the fuel of the text; without them, it's running on empty.
4. Task and format. Exactly what to write and in what shape: a 120-word post, three headline options, an email with a subject line and one call to action, a page built from blocks. The format is a constraint — and it's exactly the constraint that produces quality.
5. Facts and limits. Everything that's true: your experience, numbers, timelines, terms, names. Plus a clear ban: don't invent figures, case studies, testimonials, or client names. Plus a stop-list of words that shouldn't appear: "unique," "effective," "innovative," "comprehensive approach" — anything you yourself would never say to a client out loud.
6. Voice samples. Two or three paragraphs of your own real writing — from emails, posts, conversations. A model picks up rhythm and word choice from examples far better than from descriptions like "write warm and human."
Six parts isn't a lot. A prompt like this takes up half a page, gets written once, gets saved as a template, and after that you only ever change parts 2 through 4.
Headlines by Formula, Not by Inspiration
The headline is half the result, and this is where the prompt needs to be especially precise. I give the model the same formulas I use myself in the D.N.A. launch model, and ask for several options on each:
- Outcome + condition. "Your first ad lead, one week after your site launches." A promise, plus the condition that makes it real.
- "Without X." "A website that sells, without an agency and without a spec document." Removes the fear of cost and complexity in one word.
- Contrast. "Two business owners, one ad campaign: why one gets leads and the other gets clicks." A scene with two outcomes that makes you want to keep reading.
- The customer's own question. "Why is my site silent when the ads are working?" — the exact wording of the pain point, straight from a conversation.
- Number + substance. "Three page blocks that make people submit a lead" — only when the number is real.
Ask for ten options, not one: the first three are usually generic, the strong ones show up further down the list. And run every one through a single test: could this headline go on a competitor's site unchanged? If yes, it isn't your headline.
A Weak Prompt and a Strong Prompt, Side by Side
| Part | Weak prompt | Strong prompt |
|---|---|---|
| Role | — | Direct-response copywriter, English, B2B |
| Who it's for | "for business owners" | A service-business owner who already paid for a site and ads with no results |
| Rung | — | Knows solutions exist, is comparing options |
| Fear and desire | — | Afraid of paying "for looks" again; wants to see leads within the first week of ads |
| Format | "write a post" | Under 120 words, one call to action, no emoji, three opening-sentence options |
| Facts and limits | — | The site is built from blocks in a week; don't invent numbers or case studies; word stop-list |
| Voice | — | Two paragraphs from my own emails to clients |
The left column gives you copy for everyone and no one. The right column gives you copy you can publish after two rounds of edits.
Techniques That Kill Generic Copy
A critic after the author. A second request in the same conversation: "Find everything in this text that anyone could have written about anything, and rewrite those parts using the specific facts given above." A model is quite good at spotting its own clichés when you ask it to.
Ban rhyming paragraphs. A model loves opening every paragraph the same way and repeating the "not X, but Y" construction. Say it outright: every paragraph starts with a different sentence type; the contrast construction appears no more than once per piece.
A scene instead of a statement. Instead of "you're losing leads," write "the site has been in the works for three months and it's still silent." Ask the model to rewrite every statement about the problem as a scene from the life of the person described in part 2.
A limit on sentence length. "No sentence longer than 20 words in the first paragraph" — and the text starts to breathe.
The read-aloud check. The last filter is you: read the text out loud. Anything you stumble over, or would never actually say to a client in conversation, gets cut or rewritten in your own words. That's the exact moment the text becomes yours.
Three Templates You Can Copy
Template for a post or an email:
You are a direct-response copywriter, writing in English. Who it's for: [who this person is, what they do, what they've already tried]. Awareness rung: [just sensing the problem / knows solutions exist / comparing / ready to pay]. Fear: [one specific fear]. Desire: [one specific desire]. Task: [a post under N words / an email with a subject line]. One call to action: [action]. Facts you can use: [list]. Don't invent numbers, case studies, or testimonials. Don't use these words: [stop-list]. Here's my voice — two examples: [text samples]. Give me three opening-sentence options and one full draft.
Template for headlines:
Based on the context above, give me three headlines for each formula: outcome + condition; "without X"; a contrast between two scenarios; the customer's exact question. Up to 60 characters each. No adjectives like "unique," "effective," "innovative." Note which fear each headline addresses.
Template for the critic pass:
Read the text above as a skeptical customer with the fear described earlier. Flag every sentence that could go on a competitor's site unchanged. Rewrite those sentences using only the facts given in the prompt. If facts are missing, ask me a question instead of making something up.
These three templates cover most marketing copy: posts, emails, hero sections, ads. From there, the only thing that changes is the "who it's for" and "format" parts.
A Prompt Is a Consequence of the Method, Not a Replacement for It
Notice something: everything in this article is about what needs to be in your head before the prompt. Who your person is, what rung they're on, what they're afraid of, what you're promising them, and how you're proving it. The model speeds up the writing, but it doesn't replace those answers. Without them, any prompt gives you fluff — just faster.
Those exact answers are the first steps of the D.N.A. launch model: the avatar, fears and desires, the promise, the proof. If you want them in a working form you can turn into a prompt in five minutes, start with the "D.N.A. Launch Model" guide, and if you want to walk the whole path from avatar to first sale with your own copy reviewed along the way, that's the D.N.A. launch model course.
The Short Version
- Fluff isn't written by the model — it's written by a prompt with no person, no fear, no promise, no facts.
- A strong prompt has six parts: role, who it's for and their rung, fear and desire, task and format, facts and limits, voice samples.
- Write headlines by formula: outcome + condition, "without X," contrast, the customer's question, number + substance. Ask for ten options.
- Headline test: if it could go on a competitor's site unchanged, it isn't yours.
- Techniques against generic copy: a critic pass after the author pass, banning rhyming paragraphs, a scene instead of a statement, the read-aloud check.
- Numbers, case studies, and testimonials come only from your own facts; the ban on making things up should be spelled out directly in the prompt.
- A prompt speeds up the method, it doesn't replace it: the answers about the avatar and the promise need to exist before the prompt does.
Frequently Asked Questions
Can I trust a model to write copy with zero edits?
For drafts of internal materials, yes. For anything a customer will see, no: the final read-aloud edit stays with you, or the text loses its voice and may end up making a promise you never intended.
Do these prompts work the same way in ChatGPT, Claude, and Gemini?
The structure is the same for all of them. Different models hold onto style and limits differently, so run the template through two of them and compare the results on your own copy.
How many voice samples should I provide?
Two or three paragraphs of real writing is usually enough. More isn't always better: the model can start copying specific phrases instead of picking up the rhythm.
How do I store prompts so the whole team uses them?
One shared document, or a project inside the assistant, holding a template for each type of copy with parts 1, 5, and 6 already filled in. The variable parts 2 through 4 get filled in for each task. That way, everyone writes in the same voice.
