I am not the target user for AI content tools. I do not run a blog that needs three posts a week. But I was curious to see what one would actually produce, so I gave it a piece of content I liked and expected it to write something new in the same style.
It gave me the same piece back, with the words moved around.
The argument was identical and the distinctive turns of phrase survived. Even the joke at the end survived. Nothing new had been added — no new claim, no new example, no point of view.
I thought maybe the prompt was bad. But more likely the prompt was not the problem.
The prompt was probably fine
Look at what the model was actually given: someone else’s finished text, and an instruction to produce something based on it. And for that setup, paraphrase is not a failure. It is the correct answer.
In this particular case the model is not being lazy and it is not being unimaginative. It has no other material to work with. The only substantive content in its context window belongs to someone else. Asked to write “based on” that, it does the one thing it can do.
And the obvious fix does not work here either — “Do not copy. Write it in your own words.” That instruction sits in the same context window as the source text, competing with it. The source wins, reliably, because it is concrete and the instruction is abstract.
The fix cannot be a better prompt
I believe it should be two steps instead of one.
Step one: read the reference and extract only its shape.
- What kind of hook is it — a contrarian question, a surprising number, a belief being corrected? What order do the beats come in? How many are there?
- What kind of call to action closes it?
Then throw away everything else like the topic, facts, examples, metaphors and especially phrasing.
What is left is a structure that contains particular parts, for instance: a question that contradicts what you would assume – evidence with weight behind it – the flip – send this to someone and this is a structure. It does not contain anything that belongs to the original author or their content.
Step two: generate. It receives exactly three things: the structure from step one, your topic, what you actually know — your position, your experience, what you think people get wrong. And here is the part that does the work: step two should never receive the reference. Not summarised. Not “for context.” Not at all.
If the model cannot see the source, it cannot reproduce it. Originality stops being something you request and becomes a property of the system. That is a much more reliable place to put a constraint than a sentence in a prompt.
What that can look like in practice
Say the reference is a trainer explaining why people stall on a lift, built around a memorable line about the nervous system being the real bottleneck.
Step one keeps only this: common belief stated → why it is wrong → the actual cause → one thing to try this week. Four parts. Confident, peer-to-peer. Ends with an invitation to test it.
No lift. No nervous system. No memorable line. Structure without content.
Step two gets that structure, your topic, and your own claim: “clients believe meal timing matters more than total intake, and in my experience timing changed nothing until the calories were right.”
Out comes a four-beat correction about meal timing, in your words, from your experience. Nothing connects it to the original except the shape.
How to explain what happens in simple words
A musician takes a chord progression off a song they like and writes their own song over it. The progression carries; the melody and the words are theirs. Nobody calls that plagiarism — it is craft.
Plagiarism is taking the melody and changing a few notes. The tool I tried took the melody.
And this is the part I keep coming back to: it did borrow something of mine — it asked at setup which phrases I use often, and put them at the front. So it wrapped someone else’s substance in my style. That is exactly the wrong way round.
Borrow the shape and write your own words — that is craft. Borrow the words and change the voice — that is not craft.
A smaller thing worth knowing
The same tool offered me three variants. All three were near-identical — one idea, rephrased three ways. That is what happens when three options come from a single request. A model asked for three variants produces three surface rewrites of one thought.
Three separate requests, each with a different assigned angle — a personal experience, or a correction of a common belief, or a practical procedure — produce three genuinely different results.
The difference is visible, isn’t it?
The harder problem underneath
While I was going through the setup, I noticed what the tool asked me for: my topic, my audience, how energetic I want to sound, whether I use humour, which phrases I say a lot. It was about my style. But there were no questions about my substance.
The product never learned what I actually know. Nowhere was I asked what I think, what I have seen work, what I disagree with, or what the people I work with keep getting wrong.
Which means even a perfectly designed second step would have nothing of mine to build from. It would produce something generic, or drift back toward the only real content available — the reference.
I believe that a short intake at setup can help. At least some questions about “things you believe that others in your field do not“, or “the objection you hear most often“, or one question per piece — “what is your take on this?” — can improve the outcome.
“But that costs the user time, and the whole point is saving time“
This is the first objection anyone will raise, and it is fair. So, honestly: a user is not paying to avoid thinking. They are paying to avoid filming, editing, hunting for topics, and the blank page. The opinion is the one thing they own — it is what makes them worth listening to. Asking for it costs twenty seconds; asking them to write a script costs an evening.
And it should get cheaper with use. The first few pieces ask a question or two. By the tenth, the system has heard enough positions to stop asking except on new ground. A cost that falls as you use the product is not a cost. It is an asset — and it makes a user reluctant to switch.
I did not see a finished result in that product
Partway through, the flow required a step that lived inside another company’s product — its own signup, its own card, its own API key to copy back. This is the bring-your-own-key pattern, and plenty of AI products are built on it.
I was not willing to hand payment details to a service I had never used, purely to finish evaluating something I had not yet seen work. So I stopped.
A trial that cannot reach the thing that decides the purchase is not a trial.
Other tools in adjacent categories handle this differently: you get a pot of credits and you spend them however you like. Burn them all on one output if you want. It costs the company real money, and they do it anyway, because a user who has seen the finished thing once can make a decision — and a user who has not, cannot.
If the differentiated part of a product sits behind a third party, the trial has to cross that boundary on the company’s account, not the user’s. Otherwise the funnel is asking people to pay for the right to evaluate.
I would treat that as the first thing to fix, ahead of anything in this post about prompts. A pipeline that produces better output does not help if nobody reaches it.
What I know and what I am guessing
I want to keep these apart, so the confident part does not carry the guesswork.
What I saw. I gave a tool a reference. It gave the reference back, reworded. Three variants, all the same idea. I never reached a finished result. That part is not a guess.
What I am fairly sure about. If a model can see the source while it writes, it will lean on it. That is how these systems work. It is not one product’s quirk.
What I am only proposing. The two-step design. I have not built it. I have not tested whether the structure survives without the content, or whether the output feels better to a real person and not just safer on paper. I would want to test it.
What I do not know at all. The money. Two steps plus retries is about double per generation, maybe more. For a product that sells credits that is a big number, and it lands on a company already paying for model calls. I have no idea if it adds up.
One more thing I would watch. Output that no longer echoes something proven can feel worse at first. The echo was doing a job — it felt like proof. Take it away and you take away a comfort, even if what replaces it is more honest.
So this is not a clean win. It makes the product better and the funnel worse. Someone has to choose that on purpose, instead of finding it in the numbers a month later.
That choice is the real product work. The prompt was never the hard part.
Leave a Reply