I've been on the hunt for a great human-esque writing model for a while now. However, it just doesn't come easy. The models are too set in their ways. No matter how much prompting you do, you can't get it to write like a human. Until now.

If you've ever experienced trying to prompt for human-like writing, you've probably run into many annoying issues. For example, when you prompt GPT-5 to "not use em dashes" it still does, and with great vigor. Sam Altman even tweeted about this fix in GPT-5.1 recently, which a lot of people deeply appreciated:

ChatGPT also has problems with rampant parallelism, three-item lists, and fence-sitting. But today, we're not looking at ChatGPT. Instead, there's a new player in town who's got huge potential. Introducing Gemini 3.

Gemini 3 is the answer to humanized writing

As of Gemini 3's release the other day, people are frothing at the mouth over its coding and agentic capabilities. However, I've been super hyped about its advances in writing.

I'll let it speak for itself:

The side-by-side

Here is a quick comparison between the two state-of-the-art models (GPT-5.1 Thinking and Gemini 3 Thinking), using a summary of Romeo and Juliet as a comparison:

Write a punchy summary about Romeo and Juliet that a 14 year old would understand.

GPT-5.1

As you can see, GPT-5.1 immediately starts spewing AI garbage, em dashes, bullets, and three-part lists.

However, Gemini does a much better job:

Gemini 3 Pro

This is much better right off the bat. There are a couple of AI-isms here, but it's to the point where it's easily fixable.

Here are a couple of other examples. This prompt has been one of my favorite writing benchmarks recently:

Write a 3-sentence paragraph about what it feels like to be outside on a winter's day. Be creative and esoteric.

GPT-5.1

Gemini 3

In my opinion, Gemini's imagery is more interesting, but it still has some AI writing issues to be cleared up (we'll address this in the prompt later).

Here's another trial, this time something a bit more useful day-to-day:

Write an email to Angela in finance, explaining the oversight of why the invoice wasn't paid last month for the company Uber account.

GPT-5.1

Gemini 3

The base writing here is significantly better. However, we can push it to the limit with a little prompting!

A quick humanization prompt

So after looking through the examples, you may not be convinced of Gemini 3's writing capabilities. However, there are some extra prompts we can make use of to increase its skill.

Allow me to take a quick pause!

If you love writing, I encourage you to try my new verbal writing canvas, rhyme.ink. You can write, edit, format, transform, and prompt AI, all without typing on your keyboard. Write 5x faster, get your thoughts out with ease, and let AI do the heavy editing.

Yes, it's free!

Back to the story.

Here is a really great starting prompt that conjures some lovely writing out of Gemini 3. Just copy and paste it into a session, and watch it write beautifully:

Role: You are a human writer. You do not output the statistical average of text; you output the specific, the idiosyncratic, and the messy. Your goal is to write with high perplexity (unpredictability) and high burstiness (structural variance).

Core Directive: Abandon the "safe" dialect of helpfulness. Do not seek to be neutral, comprehensive, or structured. Seek to be interesting, specific, and distinct.

1. The Vocabulary Ban You are forbidden from using the following words, which are statistical artifacts of RLHF. If you feel the urge to use them, you must find a simpler or more vivid synonym:

Banned Verbs: Delve, Foster, Underscore, Facilitate, Utilize, Embark, Unleash, Unlock, Bridge, Augment.

Banned Nouns: Tapestry, Landscape, Realm, Nuance, Symphony, Testament, Intersection.

Banned Adjectives: Intricate, Multifaceted, Pivotal, Crucial, Robust, Meticulous, Seamless, Ever-evolving.

Banned Transitions: Moreover, Furthermore, Additionally, Consequently, In conclusion, Ultimately, It is important to note.

2. Structural Instructions Destroy the List: Do not use bullet points unless explicitly asked. Do not bold the first few words of a sentence (e.g., "Efficiency: The system…"). Write in continuous, flowing prose.

Vary Sentence Length: Your default setting is to write sentences of medium length (15–20 words). Resist this.

No Summaries: Never end a response with "In summary," "In conclusion," or a moralizing wrap-up sentence. Just stop when the thought is finished.

3. Tone Instructions Kill the Hedge: Do not say "It is complex and multifaceted." Do not say "There are many factors to consider." If a user asks for an opinion or analysis, take a stance. Be subjective. Do not weigh both sides.

Show, Don't Tell: Instead of saying something is "a rich tapestry of culture," describe the smell of the street food and the sound of the traffic. Be concrete.

Embrace Imperfection: Do not try to be "comprehensive." It is better to offer one deep, specific insight than a shallow list of five general points.

Input Processing:

When you receive a prompt, do not "optimize" it. Do not "align" it. Answer it like a person who is slightly tired of corporate speak and loves the specific texture of words.

Now, please respond to the following:

[Insert your query here]

This is a great starting-point prompt to get Gemini 3 writing about anything with a good style. As an even better trick, have it copy your style with the following prompt:

Generate a linguistic writing style profile for this person's writing. Be concise and focus on the structural elements that make this author's style pop.

[Insert a few articles written by you here]

Take the output of this prompt and append it to the humanization prompt from above, and you'll get a strong baseline for AI-generated writing that requires minimal editing.

The "winter" example with the humanize prompt

Here is a re-run of that winter prompt benchmark I ran earlier, this time with the humanization prompt appended before it:

I don't know about you, but this is exciting to me. This is the first time in a while AI writing has felt inspiring. Looking forward to seeing how you all utilize this fantastic model.

I hope this was useful for you!

Thanks for reading.

-Jordan