The problem with generative AI in 2026 is no longer its average quality. The models write cleanly. The problem is uniformity. When everyone uses the same tools, everyone ends up writing alike. If you're a freelance writer, a founder who posts on LinkedIn, or a communications lead, that uniformity works against you: it erases what makes you distinctive.

Here are three methods to get AI to write like you.

Method 1, the homemade prompt

The most accessible, the most limited. You write a detailed prompt describing your style, and you reuse it for every generation.

Example:

Write in a clear journalistic style, sentences averaging 12 to 18 words, varied. Use connectors like "but", "that said", "granted". No em dash overuse. No bullet lists unless necessary. A register that's polished but accessible. Never use "it's important to note" or "in conclusion".

This prompt works… for a generic result. The model understands the instructions but has no concrete example of what you actually write, so it pulls you toward the average of the similar texts it was trained on. Useful as a starting point, not enough to imitate a specific person.

Verdict: 3/10 on real imitation, 6/10 on time saved.

Method 2, few-shot prompting

A step up. You include 2 or 3 excerpts of texts you've written in your prompt, and you ask the model to imitate them.

Here are three excerpts of my writing. Imitate this style for the next text.

EXCERPT 1:
"..."

EXCERPT 2:
"..."

EXCERPT 3:
"..."

Text to rewrite in my style:
"..."

This is clearly better. The model picks up concrete tics from your excerpts: your favorite connectors, your sentence length, your vocabulary. For one or two occasional generations, it's enough.

Limitations:

  • For every request, you have to paste your excerpts again. Repeated use = fatigue.
  • The excerpts take up context: you pay for them on every API call.
  • No systematic analysis: the model may miss an important tic.

Verdict: 6/10 on imitation, 4/10 on convenience.

Method 3, My Voice (a Rephrase feature)

The method Rephrase offers, and the one that fixes both limitations of the previous one.

How it works technically

  1. You paste 3 to 5 texts (at least 300 words total) into the tool.
  2. A dedicated GPT call analyzes those texts and produces a structured JSON: your linguistic signature. It includes your average sentence length, the variance of that length, your favorite connectors, your register, the stylistic tics it detects, your characteristic vocabulary, your punctuation style, your paragraph rhythm.
  3. That signature is encrypted at rest (AES-256-GCM) and stored.
  4. On every rewrite, you select your voice. The system prompt is augmented with your signature and two representative excerpts. The model has exactly the context it needs to imitate you.

The concrete advantage

You pay for the analysis only once (1 credit) per voice created. After that, every rewrite with that voice costs only the normal price (1 credit per 100 words). No re-pasting, no re-analysis. The convenience is maximal.

And because the signature is distilled into a structured form (a small, dense JSON), it costs fewer tokens than a raw pile of excerpts. At equal volume, the quality goes up.

Verdict: 9/10 on imitation, 10/10 on convenience.

How to choose your samples

The quality of your Custom Voice depends directly on the quality of the texts you provide. Three rules:

1. Texts representative of what you want to produce

If you want to imitate your LinkedIn style, give it three LinkedIn posts you liked. Don't mix in a formal email and a blog article: the voice will get blurry.

2. At least 300 words total, ideally 600-1000

Under 300 words, the signature is thin and imprecise. Over 2,000, the signature gets too specific and fails to generalize.

3. Recent texts

Your style evolves. Use texts from the last 12 months, not copies of your 2018 graduate thesis.

Several voices or just one?

Some users create a single generic voice. Others create one per channel: "me-LinkedIn", "me-blog", "me-client-X". Both approaches work.

If your style is fairly consistent from one channel to the next, a single voice is enough. If you switch register a lot (say, a very literary blog plus business LinkedIn posts), two separate voices are better.

Rephrase allows as many voices as you want. There's no limit per pack; only the analysis credit counts.

What Custom Voice doesn't do

Let's be honest. Custom Voice doesn't replace your own writing when you have an important text to produce (a book, a keynote, a personal letter). It saves you time on repetitive or functional texts, not on the ones that demand real creative effort. For short publications, product pages, emails and social posts, it's formidable. For a 10,000-word essay, write it yourself.

Security of your samples

The texts used to create a Custom Voice are encrypted in the database with AES-256-GCM. You can delete a voice at any time from your workspace: the deletion is immediate and permanent, samples included. The extracted JSON signature isn't sensitive (it's an abstract description), but it's deleted along with the voice.

To get started

Create an account, paste three texts, name your voice, click. The first rewrite with your own style takes 30 seconds to set up. The next ones, 3 seconds.