You can tell within 30 seconds.
The words give it away. Delving into things. Leveraging assets. Navigating complexities. Bridging phrases like "moreover" or "furthermore" wedged between paragraphs. A summary at the end that restates the opening.
That voice is the statistical average of every "professional" blog post the model ever saw. The median of all good business writing. When you ask for a blog post and nothing else, that average is what shows up.
It is not a personality. It is a default.
Where the voice comes from
The model trained on a lot of text. It learned what good business writing usually looks like. "Usually" is the issue. The median has no edge, no voice, no risk.
When you type "write a blog post about onboarding," the model picks the safest path through its training data. Five paragraphs. Intro, three sections, conclusion. A bullet list every two hundred words. Vocabulary tuned to sound polished. No specific human in mind.
You did not tell the model whose voice to use. You got the average.
Five things that change it
1. Name a writer
Vague tone instructions do almost nothing. "Casual" produces median-casual. "Professional" produces median-professional.
Name a writer the model has read enough of to imitate:
Write in the voice of Patrick McKenzie (patio11). Direct, parenthetical asides, structural arguments, willing to be opinionated.
Write like Paul Ford. Long sentences that are not boring. Specific. Funny without trying.
Write like an old-school blogger. Mid-2000s personal blog energy. First person. Loose.
The model has read those writers. It can imitate the cadence. It cannot guess what you mean by "casual."
2. Ban the vocabulary
Put this at the bottom of your prompt:
Do not use any of these words or phrases: delve, leverage, utilize,
navigate (as a verb on an abstraction), moreover, furthermore, that
said, realm, landscape, plethora, paradigm, crucial, tapestry,
ultimately. If you find yourself reaching for one, rewrite the
sentence instead.
The bottom of the prompt is where the model pays the most attention. That is the right spot for the ban list.
3. Force off-median structure
A few constraints push the model away from the safest shape:
- No three-item parallel lists.
- One paragraph in your reply must be a single sentence.
- Start at least one sentence with a contraction.
- Use one parenthetical aside.
- Do not write a summary paragraph at the end. Stop on the last useful idea.
These all violate what "median professional prose" looks like. Obeying them forces a different shape of output.
4. Give the model your own writing
This is the single biggest lever. Paste a few hundred words of your real writing into the prompt and tell the model to match the rhythm:
Here is an unedited paragraph I wrote last week. Match this voice:
the sentence length, vocabulary, willingness to use fragments, the
specific kind of humor. Don't smooth it out.
[200+ words of your writing]
Now write the [thing you need].
The model imitates the specific text in front of it more than it imitates training-set averages. If you have a Substack or a few blog posts, paste one in.
5. Be specific about the thing
Vague topics produce vague writing. Compare these:
Write about productivity for knowledge workers.
Write about the specific feeling of having 23 browser tabs open at 4:00 PM on a Friday because every task is half-finished. Name the feeling. Do not generalize beyond that scene.
Specific topics force the model to commit. Vague topics give it permission to fall back on median phrasing.
What humanizer tools cannot do
Tools that promise to "humanize AI writing" mostly swap synonyms, sprinkle in typos, or replace banned words with similar ones. They are surface-level cleanup. The output still has the shape of AI prose. It just has different words.
The real fix is upstream. If your prompt has voice anchors and specificity, you do not need a humanizer because the first draft already reads as human.
The two minute check
After the model replies, re-read with this checklist. If you spot any of the patterns below, the prompt was missing a voice anchor:
- The word "delve" or "delving"
- Three items in parallel structure
- "Crucial role"
- A paragraph that ends with "ultimately"
- A summary paragraph that just restates the intro
Run it back through the prompt with the bans above. The voice usually cleans up.
Quick way to try it
Paste a prompt you wrote recently into FixMyPrompt. The rubric specifically scores voice and tone constraints. Most prompts that score low here are missing the voice anchor. Adding one is a small edit with a big payoff.