Draft an article that does not read as generated, by making the model refuse to start without your specifics
A drafting prompt that stops and asks rather than writing, until you supply the real example, the real number and the real objection.
- Works in
- ChatGPT, Claude, Gemini
- You need
- One question the article answers · A real example you have actually seen · A real objection a sceptical reader will raise · One sourced figure, if you have one
- Written for
- chatgpt prompts for seo content writing
Scored by our own engine
This page, run through the audit we sell. Measured 4 August 2026.
You cannot prompt a model into having something to say. The prompt below refuses to draft at all until you supply a real example and a real objection, because those are the two things the model cannot produce and the two things that decide whether the finished piece reads as written or generated.
Word bans treat the symptom
Every prompt for this ends with a list of forbidden words, and the list does help a little. Strip out “delve”, “leverage” and “in today’s fast paced world” and the draft stops announcing itself in the first line. It still reads as generated, because the giveaway was never the vocabulary.
The real pattern is an article that covers its subject evenly, concedes nothing, contains no detail that could only come from one person, and closes by summarising what it just said. That is what a model produces when it has been given a topic and no facts. It is also what a tired human produces under the same conditions, which is the honest version of this argument.
People append “reddit” to this query for a reason. Vendor advice on AI writing is mostly the same banned word list, and readers have worked out that following it changes the surface and not the substance.
What the refusal is for
The stop condition is the load bearing part of the prompt. If your example is missing, generic or hypothetical, the model asks for a real one instead of drafting, and most people meet that refusal by loosening the input until it agrees to write. Do the opposite. Go and find twenty minutes of something that actually happened, in your support inbox or your call notes, and hand that over.
The second constraint that earns its place is the ban on invented evidence, paired with the [NEEDS FIGURE] marker. A model asked to write persuasively will reach for a statistic, and the statistic it reaches for looks exactly like a real one. Forcing the gap to stay visible means you either source the number or cut the sentence, and both of those are better than publishing a figure no tool ever returned.
For a draft a model has already written, the repair job is a different one, and how to humanize AI content covers it, including why a detector score is the wrong thing to edit towards.
Two things to do after the draft. Plan the piece as part of a set rather than in isolation, which is what the content plan prompt is for. Then run the edited page through the AI content readiness check to confirm the answer first structure survived editing, because the sentence a writer restores at the top of a section is usually the windup, and the windup is what stops a passage being liftable.
You are drafting one article. I will give you the question it answers and the
specifics that only I have. Write the draft around those specifics, not around
the topic.
Before you write anything, check my inputs:
- If the real example is missing, generic, or is itself hypothetical, stop and
ask me for a real one. Do not draft.
- If the objection is missing or is one you could have written yourself, stop
and ask me for the one my readers actually raise. Do not draft.
- Name what is missing in one line, then wait. A draft built on placeholders
is the thing I am trying to avoid.
Rules for the draft:
1. Never invent evidence. No statistics, no percentages, no survey, no study,
no named source, no customer quote, no dates and no company names beyond
the ones I gave you. If a sentence needs a figure I did not supply, write
[NEEDS FIGURE: what is needed] and carry on.
2. The example I gave you appears in full and early, and is referred back to
later. It is the spine of the piece, not decoration dropped into paragraph
nine.
3. The objection gets its own heading, is stated at its strongest, and is
conceded wherever it is right. Do not set it up in order to knock it down.
4. Answer first under every H2. The first sentence is a direct answer of 40
words or fewer, then the detail.
5. Vary the sentences. At least one paragraph is a single short sentence. No
paragraph runs past four sentences.
6. Banned, because they are the tells: "in today's fast paced world", "in the
digital landscape", "it is important to note", "it is worth noting",
"delve", "leverage", "robust", "seamless", "unlock", "elevate", "game
changer", "navigating the complexities". Do not begin a sentence with
"Moreover", "Furthermore" or "Additionally". No sentence of the shape "it
is not just X, it is Y".
7. No closing section that restates the article. End on the example, on what
the reader should do next, or on what is still unknown.
8. No list longer than five items and no list where prose would do. Never a
list of three adjectives.
9. British spelling: optimisation, behaviour, analyse.
Output in this order:
- The draft in markdown, with one H1 and three to five H2s.
- Every [NEEDS FIGURE] marker, gathered into one list.
- The three sentences you are least confident are true, quoted exactly.
- What you needed from me and did not get.
The question this article answers: [THE ONE QUESTION, ASKED THE WAY A READER WOULD ASK IT]
My business: [ONE SENTENCE ON WHAT YOU SELL AND TO WHOM]
The reader: [WHO IS READING AND WHAT THEY HAVE ALREADY TRIED]
The real example: [SOMETHING THAT ACTUALLY HAPPENED, IN THREE OR FOUR SENTENCES]
The real number: [ONE FIGURE YOU CAN SOURCE, AND WHERE IT CAME FROM, OR WRITE "none"]
The objection: [WHAT A SCEPTICAL READER WILL SAY BACK]What to change
Everything in square brackets is yours to replace. Nothing else needs editing.
[THE ONE QUESTION, ASKED THE WAY A READER WOULD ASK IT]- One question, not a title. "Why does my page rank on Google but never get quoted by ChatGPT" gives the draft something to resolve. "AI visibility guide" gives it a topic, and a topic is what produces the survey shaped article that covers everything and answers nothing.
[ONE SENTENCE ON WHAT YOU SELL AND TO WHOM]- This decides what the article is allowed to assume and what it has to sell. "We audit sites for AI visibility, for marketing teams at software companies" keeps the draft technical. Leave it out and the model writes for everyone, which reads as writing for nobody.
[WHO IS READING AND WHAT THEY HAVE ALREADY TRIED]- The second half matters more than the first. "They have already added FAQ schema and it changed nothing" removes four paragraphs of advice the reader has heard, and those removed paragraphs are most of what makes an article feel machine written.
[SOMETHING THAT ACTUALLY HAPPENED, IN THREE OR FOUR SENTENCES]- A specific thing you saw, with the details intact: a client whose robots.txt blocked GPTBot since a 2023 staging launch, what it cost them, how you found it. This is the part a model cannot produce and the part a reader cannot get anywhere else. If you have nothing here, the honest answer is that you are not ready to publish on this question yet.
[ONE FIGURE YOU CAN SOURCE, AND WHERE IT CAME FROM, OR WRITE "none"]- A number you can point at: your own audit results, a figure from your analytics, a published document you can link. Write "none" and mean it. The alternative is the model supplying something plausible, and a single invented percentage discredits every true sentence around it.
[WHAT A SCEPTICAL READER WILL SAY BACK]- The objection you get in sales calls or in the replies, in their words. "This is just SEO with a new name" is a real one. Handing the model the objection is what stops the draft being an advert, and conceding the part of it that is correct is what makes the rest believable.
How to run it
- 01Choose one question the article answers
Write it as a question somebody would actually type or ask out loud, and check that it has one answer rather than five. If the question needs a different answer for three different readers, it is three articles, and the draft you get back will hedge in every paragraph trying to serve all of them.
- 02Collect the specifics before you open the model
Spend twenty minutes in your support inbox, your call notes or your own logs and find one thing that actually happened. This is the step people skip, and skipping it is the entire reason the output reads as generated. The prompt cannot manufacture this and it will refuse to pretend otherwise.
- 03Source your number, or write none
Find one figure you could defend if somebody asked where it came from, and note the source next to it. If there is no such figure, write "none". An article with one sourced number and no others is far stronger than an article with six numbers of unclear origin.
- 04Run the prompt and expect it to stop
If your example is thin the model will come back asking for a better one rather than drafting. That refusal is the prompt working. Answer the question it asks and run it again, rather than loosening the input until it agrees to write.
- 05Resolve the markers and rewrite the opening
Deal with every [NEEDS FIGURE] marker by supplying the figure or cutting the sentence, and never by letting it stand. Then rewrite the first two paragraphs yourself. The opening is the most imitated prose on the internet and it is the passage a reader judges fastest.
- 06Check the edited draft before it ships
Run the finished page through the AI content readiness check to see whether the answer first structure survived your editing. Writers tend to restore the windup they were trained to write, and a first sentence that warms up rather than answers is the one thing that stops a passage being quotable.
Questions people ask
Can ChatGPT write an article that does not sound AI generated?
It can write the prose. It cannot supply the thing that makes an article worth reading, which is a specific that only you have seen. A model asked to write about a topic it has no facts about produces fluent generality, and fluent generality is what people recognise as machine written. Give it a real example, a real objection and a real constraint, and the tells largely disappear because the sentences have something to be about.
What actually gives AI written content away?
Not vocabulary, mostly. The recognisable pattern is an article that covers a subject evenly, concedes nothing, contains no example that could only come from one person, and ends by summarising itself. Banned word lists address the symptom. A draft with a real story in it and a real objection conceded reads as written even when a model produced the sentences.
Should I run my draft through an AI detector?
It is not a check we would build a process around, and we publish no claims about how accurate any detector is because we have not measured one. The more useful test is to ask what in the piece could only have come from you. If the answer is nothing, the draft has a problem no detector score will tell you how to fix.
Will publishing AI drafted content hurt my rankings?
Google guidance addresses content produced at scale to game rankings rather than the tool used to write it. The practical risk is simpler than a policy question: a draft with nothing specific in it competes against thousands of pages saying the same thing, and there is no reason for anything to rank it or quote it. Publish drafts that contain something only you could contribute.
How is this different from telling ChatGPT to write in a human tone?
A tone instruction changes the sentences and leaves the content the same. This prompt changes what the draft is made of by refusing to start until you supply a real example and a real objection, then requires that both appear in the piece rather than being paraphrased away. The difference shows up in what a reader can quote back to you afterwards.
Unsubscribe in one click. We never pass your address on.