Knowing how to prompt AI to write like a human comes down to two moves: show the model human writing, and tell it who will read the result. Paste two or three paragraphs in the voice you want, describe the reader, and give the model a short list of habits to replace with something specific. The instruction "sound more human" on its own does almost nothing, because the model has no fixed picture of what that means.
Why generative AI writing sounds generic
A language model writes by predicting the most probable next word, over and over. Probable words are the ones everybody uses, so unguided output drifts toward the average of everything the model has read.
Some of those habits are now measurable. A team led by Dmitry Kobak tracked vocabulary across more than 15 million biomedical abstracts and found that "delves" appeared 28 times more often in 2024 than the earlier trend predicted. They estimate that at least 13.5 percent of 2024 abstracts were processed with a language model.
Wikipedia editors keep a field guide to the signs of AI writing, built from cleaning up machine-written articles. Their list goes well past vocabulary. The patterns that come up most:
- Inflated significance, where a minor event "marks a pivotal moment".
- Copula avoidance, meaning "serves as" or "functions as" where "is" would do.
- Negative parallelism, the "not X, but Y" sentence that argues with a point nobody raised.
- Heavy use of em-dashes, plus bold scattered through ordinary paragraphs.
Each habit is harmless alone. Together they produce prose that sounds like a hotel welcome letter drafted by a very enthusiastic committee.
The fix follows from the cause. If the model defaults to the average, the prompt's job is to make the average unavailable.
AI prompt examples beat adjectives for tone
Adjectives such as "conversational" or "authentic" describe a voice. A sample demonstrates one, and models copy demonstrations far more faithfully than they follow descriptions.
Anthropic's prompting best practices call examples one of the most reliable ways to steer tone and format. The guide recommends three to five of them, wrapped in <example> tags so the model can tell them apart from your instructions.
The model copies everything, so three samples that all open with a question will produce a fourth that does too. Paste writing you like, ideally your own, then add a line that separates voice from content:
<examples>
<example>[a paragraph you wrote]</example>
<example>[another one, on a different topic]</example>
</examples>
Write the announcement in the voice of the examples.
Match their sentence length and word choice.
Take no facts from them.The same guide notes that the formatting of your prompt influences the formatting of the reply, so write the request in the register you want returned.
Asking for relaxed, natural prose in a prompt laid out like a tax form sends a mixed signal. The model usually resolves it in favour of the tax form.
What a negative prompt in AI writing should look like
A negative prompt lists what the model should leave out. For text, a bare ban list is weak. Pair each ban with the thing to write instead.
Anthropic's guide says the same thing directly: tell the model what to do instead of what to avoid. Its own example swaps "do not use markdown" for a request for smoothly flowing prose paragraphs. A ban removes one option and leaves the next most probable habit in its place.
A ban also has a quieter cost. Naming a word puts it into the context, so a model told to avoid "tapestry" is, in a small statistical way, now thinking about tapestries.
Pairs worth writing down:
| Ban | Write instead |
|---|---|
| Padded vocabulary such as "delve" and "tapestry" | The plainest word that is accurate |
| "Serves as", "stands as" | "Is" |
| The "not X, but Y" sentence | State the claim on its own |
| Em-dashes | A comma, a colon or a full stop |
| A closing paragraph that summarises | End on the last useful point |
Each ban works better with its reason attached, which is the idea behind cause and effect AI prompts. Tell the model that your readers treat em-dashes as a sign of machine writing, and it will apply the same caution to punctuation you never listed.
The house rules this site is written under ban three of the habits in that table outright. You are welcome to check this article for them.
System prompts: where AI tools keep your voice
A system prompt is a standing instruction that applies to every message in a conversation. Put your voice there once and stop retyping it.
Consumer apps expose it under names like custom instructions or project instructions. Through an API it is the system field of the request. Anthropic's guide notes that setting a role in the system prompt focuses the model's tone, and that even a single sentence makes a difference.
The most useful sentence describes the reader. "Write for a finance lead who reads on a phone between meetings and distrusts hype" implies short sentences and no superlatives without naming either. A tone adjective gives the model a mood. A reader gives it decisions.
Keep it to one screen. Instructions buried in paragraph nine compete with everything above them.
A 4,000-word style guide in the system prompt is the prompting equivalent of the onboarding binder: thorough and admired, then consulted by nobody.
Generative AI prompt examples, before and after
The difference between a weak prompt and a strong one is almost always inputs. Three pairs:
| Before | After |
|---|---|
| Write a LinkedIn post about our launch. Make it sound human. | Write 120 words announcing the feature to customers who asked for it in March. Name the feature in the first sentence. No exclamation marks. |
| Rewrite this email to be less robotic. | Rewrite this email as a reply to a colleague you like. Keep every fact. Cut any sentence that could appear in another company's email. |
| Write a product description in a friendly tone. | Here are two product descriptions we wrote by hand. Write one for the new kettle in the same voice. Mention the 90-second boil time once. |
Every "after" gives the model a reader and a limit, and most hand it facts nobody else has. The average of the internet does not know your kettle boils in 90 seconds.
Specific inputs are also what moves detector scores, a separate question covered in the guide to what a Claude prompt to bypass AI detection really changes.
How to organize AI prompts into a style library
Once a voice prompt works, save it. A small prompt library means every draft starts from the same voice.
A layout that holds up:
- A voice file holding your samples and the reader description.
- A ban list, each entry carrying its replacement and its reason.
- Task templates for recurring jobs such as replies and summaries, each pulling in the voice file.
Date every change, since model upgrades shift defaults and you will want to know what moved. AI Prompt Engineering covers building prompt systems that stay reliable across those changes. If you are starting out with Claude, Claude AI for Beginners walks through writing clear prompts from the first chat.
Then read the draft aloud. The ear catches what the eye forgives, such as three sentences of identical length in a row or a paragraph that ends by restating itself. A second model pass against your ban list catches more.
If a sentence makes you wince when spoken, it will make your reader wince in silence, which is worse, because you will never hear about it.
A starter prompt you can copy
You are writing for [reader], who will read this [where and when].
Match the voice of the samples in <examples>: sentence length and word choice.
Take no facts from the samples.
Use plain words. Use "is" where it fits. State each claim directly.
Use a comma, a colon or a full stop where you would reach for an em-dash.
Give at most two examples inside a sentence; use a list for more.
End on the last useful point, with no summary.
If you need a fact you do not have, write [TK] and keep going.Fill in the reader first. It does the most work, and most people leave it out.