Does telling AI to be creative or original work?

In practice, not much. Those are adjectives and the model has no reliable way to check itself against them, so it tends to produce what ordinary writing looks like when it is trying to seem creative. Prohibitions give it something checkable instead: do not use the word realm either happened or it did not, while be original cannot be confirmed by either of you.

How long should a writing prompt be?

Long enough to remove the easy paths and no longer. A specific speaker, a few hard prohibitions and a locked format will usually fit well under 200 words, and that is where most of the improvement sits. Past that you tend to add instructions that contradict earlier ones, and contradictions push output back toward the safe middle.

Prompt craft· August 3, 2026

Why AI writing sounds generic

Q: Is it faster to write it yourself than to prompt for it?

Often yes, and that is a legitimate answer. These repairs are worth it for writing you produce repeatedly in the same shape, where the prompt is reusable. For a single paragraph you already know how to write, opening a blank document is quicker.

It is not that the model is stupid, and it is not that you picked the wrong one. Generic output is the model doing exactly what it was built to do. Once you know what that is, the fix stops being a mystery and becomes a checklist.

The short answer

AI writing sounds generic because a language model works from a probability distribution over what comes next, and it draws from the likely end of it. Likely phrasing is, near enough, expected phrasing. Left unconstrained, the output settles near the middle of everything the model has seen. You fix it by removing that middle as an option: ban the stock phrases by name, give the model a specific person to be, force it to commit instead of hedging, show it a sample of your own writing, and lock the shape of the output.

The mechanism, in one paragraph

A language model works by repeatedly answering one question: given everything so far, what could plausibly come next? It produces a spread of candidates with probabilities attached, then picks from the plausible end of that spread. It does not always take the single top option — there is deliberate randomness in the sampling, which is why the same prompt gives you slightly different text each time. But the spread it is drawing from is built out of what usually follows, and that is the problem: "plausible" and "predictable" point in the same direction. When the instruction is vague, nothing pulls the output away from the middle of that spread, so you get the middle — the phrasing that would be least surprising to the most people. Which is a fair definition of generic.

This is why "write me a post about productivity" cannot produce something sharp. There is no sharp version of that request. You have described a topic and left every actual decision — who is speaking, what is being claimed, what is being refused — to a system whose default is to not decide.

The three causes

1. The default is the average

Nothing in a bare prompt tells the model to be specific, so it is not. It reaches for the phrasings that appear most often around your topic. Those phrasings are common precisely because they are inoffensive and broadly applicable, which is another way of saying they carry almost no information.

2. You gave it no one to be

"Professional tone" is not a person. It is a category — an enormous one — and the model will write the centre of it. A named, specific speaker, with a job, an opinion and something at stake, narrows the target sharply. The difference between "write professionally" and "write as a freelance developer who has been burned by vague briefs and is explaining what a good brief contains" is the difference between a category and a human being.

3. You gave it nothing to refuse

This is the one people skip, and it is the one that does the most work. A prompt with no prohibitions has no edges. When you write "do not use the words 'leverage', 'seamless' or 'in today's landscape'; do not open with a definition; do not hedge", you have removed the easiest paths. The model has to find a less obvious route, and the less obvious route is where the interesting writing lives.

The tells

These are the patterns that show up again and again in unconstrained output. None of them are wrong on their own — the giveaway is the density. When several appear in one paragraph, the prompt underneath was almost certainly a single line.

The throat-clear opening. "In today's fast-paced digital landscape…" Nothing has been said yet, and a full sentence is gone.
Definition-first. Answering "how do I do X" by first explaining what X is, to a reader who plainly already knows.
Rule of three, everywhere. Every list is three items, every sentence has three clauses, whether or not the subject has three parts.
Symmetrical hedging. "While X has benefits, it also has drawbacks." True of everything, useful for nothing.
Vocabulary that nobody says out loud. Delve, tapestry, testament, multifaceted, realm, underscore, navigate the complexities.
The empty summary. A closing paragraph that repeats the opening with different words and adds no new claim.
Compliment-then-comply. "Great question!" before the actual answer.
Refusal to commit. Three options presented as equally good, with no recommendation — because a recommendation could be wrong, and the average is never wrong.

Five repairs

Each of these removes one path to the middle. They stack — in my experience the combination does considerably more than any one of them alone.

1. Ban the tells by name

Vague instructions get vague compliance. "Don't sound like AI" means nothing to the model — it has no notion of sounding like itself. A banned list is concrete and checkable.

banned-list.txt

Banned — do not use any of these:
delve, tapestry, testament, multifaceted, realm, underscore,
seamless, leverage (as a verb), navigate the complexities,
"in today's ... landscape", "it's important to note",
"whether you're a ... or a ...", "the key is"

Also banned:
- Opening with a definition of the topic
- Any sentence that would be true of any other topic
- A closing paragraph that only restates the opening
- Presenting options without picking one

Paste this block at the end of any writing prompt. If you only take one thing from this page, I would make it this one.

2. Give it a specific person, not an adjective

Replace the tone word with a human being who has a history and a bias. The more specific the person, the narrower the space the model writes from.

voice-layer.txt

Instead of: "Write in a professional, engaging tone."

Write:
You are a bookkeeper of 15 years who has watched small
businesses fail for the same three reasons. You are patient
but blunt. You have no interest in sounding clever. You use
short sentences and concrete numbers. When something is a
bad idea you say so in the first sentence, not the last.

The test: could this description apply to a different person? If yes, it is still a category, not a voice.

3. Force a commitment

Hedging is the safest possible output, so it is the default. Make it unavailable by requiring a single answer and a reason, and by explicitly banning the balanced non-answer.

commit.txt

Pick exactly one option and defend it.

Do not list the alternatives with pros and cons.
Do not say "it depends" without immediately saying what it
depends on and then choosing for the most common case.
End with the strongest argument against your own pick,
stated in one sentence, and why you still chose it.

The last line matters: it gives the model somewhere to put its uncertainty, so it stops smearing it across the whole answer.

4. Show it your own writing

Describing a voice is far weaker than demonstrating one. Two or three paragraphs you actually wrote will do more than any adjective, because the model can copy rhythm and sentence length directly rather than inferring them from a label.

sample-match.txt

Here are three paragraphs I wrote:

"""
[paste 150-300 words of your own writing]
"""

Match this: average sentence length, how often I start with
a conjunction, how blunt I am, how much I explain versus
assert. Do not match the topic.

Before writing, list the three most distinctive things about
how I write. Then write, obeying them.

Asking it to name the patterns first is the part I would not skip — it has to commit to an analysis it can then be held to.

5. Lock the shape

An unspecified format defaults to the safest one: an intro, some evenly-sized paragraphs, a summary. Specifying the shape removes that template and often removes the padding with it.

format-lock.txt

Format, exactly:
- One sentence stating the conclusion. No preamble.
- Three short paragraphs, 40-60 words each, each making a
  different argument. No paragraph may begin with "Additionally",
  "Moreover" or "Furthermore".
- One line naming what would change your mind.

Total under 220 words. Do not exceed it.

A hard word count is doing quiet work here: padding is the first thing cut when the budget is real.

Before and after

What you changed	Unconstrained	Constrained
Opening	"In today's competitive market, email remains a vital channel…"	"Your follow-up is too early. Three days is not patience."
Recommendation	"There are several approaches, each with trade-offs…"	"Use the second one. The first breaks as soon as you have two clients."
Ending	"In conclusion, by following these strategies you can…"	"If your reply rate is already above 8%, ignore all of this."

Notice what the right column has in common: every line takes a position that could turn out to be wrong. That is the whole difference. Writing that cannot be wrong is writing that says nothing.

The repair prompt

When you already have generic output and do not want to start over, run it through this rather than asking for a rewrite. Asking a model to "make it better" usually returns the same text with more adjectives.

repair.txt

Here is a draft:

"""
[paste the draft]
"""

Step 1. Quote every sentence that would still be true if the
topic were something else entirely. List them.

Step 2. Quote every sentence that avoids taking a position.

Step 3. Rewrite the draft with all of those deleted — not
softened, deleted. Replace each one with a specific claim,
a number, or nothing at all.

Rules: no sentence may open with "In today's", "It's important
to note", "Whether you're". No closing summary paragraph.
The result should be shorter than the original. Tell me the
before and after word count.

Making it quote the offending sentences before rewriting is the important part. It forces the model to locate the problem instead of paraphrasing around it.

What this will not fix

Worth being straight about the limits, because this kind of advice is easy to oversell.

It will not give you an opinion you do not have. Constraints stop the model writing the average; they cannot supply a point of view. If you do not know what you think, the output will be well-shaped and still empty.
It will not supply facts you did not give it. Forcing specificity makes a model more likely to produce confident numbers. Sometimes those numbers are invented. Every figure still needs checking.
It will not make you an expert. A convincing voice on a subject you do not understand is a liability, not an asset. It reads well right up until a real practitioner replies.
It is not a way to beat AI detectors. That is not the goal here and it is not a game worth playing. The goal is writing worth reading, which is a different target that happens to be more useful.

Where to go next

Everything above is repair work — it assumes a draft or a prompt already exists. If you would rather not produce generic output in the first place, the underlying structure is four layers: goal, constraints, voice, format. I broke that down with five complete worked examples in how to write an AI prompt that actually works. The constraints layer in that piece is the same idea as the banned list here, applied earlier.

Common questions

Does switching to a different AI model fix this?

Not on its own. Newer models have better defaults and fewer obvious tells, but every one of them is still drawing from a distribution of likely continuations, so an unconstrained prompt still lands near the middle. In my experience, changing the prompt moves the output further than changing the model does.

Does telling it to "be creative" or "be original" work?

In practice, not much. Those are adjectives, and the model has no reliable way to check itself against them — it tends to produce what ordinary writing looks like when it is trying to seem creative, which usually means more metaphors. Prohibitions give it something checkable instead: "do not use the word realm" either happened or it did not, while "be original" cannot be confirmed by either of you.

Won't a long banned list make the writing stiff?

Sometimes, and it is a real trade-off. If the output starts reading like it is dodging something, cut the list to the five bans that matter for that piece. The tells you personally notice most are the right ones to keep — a banned list is meant to be pruned, not accumulated.

How long should the prompt be?

Long enough to remove the easy paths, and no longer. A specific speaker, a few hard prohibitions and a locked format will usually fit well under 200 words, and that is where I find most of the improvement sits. Past that you tend to be adding instructions that contradict earlier ones, and contradictions push the output back toward the safe middle.

Is it faster to just write it myself?

Often, yes — and that is a legitimate answer. These repairs are worth it for writing you will produce repeatedly in the same shape, where the prompt is reusable. For a single paragraph you already know how to write, opening a blank document is quicker than engineering your way to it.

Free · No cost

Fifteen prompts, already built this way

The repairs above are the manual version. The free sampler is 15 prompts across 10 categories — strategy, sales, content, email, proposals, productivity, social, research, finance and AI workflows — each one already carrying its own constraints, voice and format, so you can paste and go. It is a PDF, it costs nothing, and it downloads the moment you ask for it.

Get the 15 free prompts →

One email address. The download starts immediately — you do not have to wait for it to arrive in your inbox.

Prefer to watch? The prompt structure behind all of this is broken down on screen in Episode 2 of The AI Nomad.