How to write tweets in your own voice with AI (without sounding like AI)

What we tested

In October 2026 we ran a blind comparison on 28 topics, 14 in English and 14 in Turkish. For each topic, three systems wrote a tweet: a one-shot generator with no information about the writer, a chat assistant given a one- or two-sentence description of the writer, and xVirals working from a style profile built from real posts. Three AI judges ranked the outputs without knowing which system wrote which, on human voice, concrete detail, fluency, scroll-stopping power and risk of invented facts.

Read the numbers below as direction, not proof. The judges were AI models, not people; 28 topics is a small sample; and the systems did not run under equal conditions. We are not naming the other tools for that reason. The patterns, though, were consistent enough to be useful.

Why AI tweets sound like AI

The one-shot generator, which knew nothing about the writer, produced the most recognisable AI pattern. Every one of its outputs — 100% in both languages — ended with a question to the audience ("What's your experience? Let's share!"). About a quarter used emoji, and almost none were written in the first person. On human voice it ranked last in almost every set.

The assistant that had a short description of the writer did better, but it fell into a template of its own: a contrarian opening, two to four one-line "insights", and a closing aphorism. In Turkish its lines often read like translations.

The common cause is the same. When a model has no examples of how you write, it falls back to the average tweet it has seen — and the average tweet is the one people have learned to scroll past.

1. Give the model your real posts, not a description

"Casual, witty founder who writes about SaaS" describes thousands of people. Your last 20 posts describe one. Rhythm, sentence length, whether you use capitals, how you open and how you stop — those survive only through examples.

In our test, drafts written from a style profile with real example posts were ranked first on human voice 99% of the time. That is the single biggest lever, and it works with any tool that accepts examples.

2. Start from something only you know

Generic drafts talk about a topic. Good tweets come from a person who was there: a number from your own dashboard, the exact error you hit, what a customer actually said. Give the model that detail in the prompt. Voice-matched drafts in our test used the first person and specific numbers far more often than the generic ones, and the judges ranked them first on concrete detail every time.

3. Check every specific detail before you post

This is the catch. A model that writes in a confident first-person voice will also invent the details that make it sound real — "I tried this on a small settings screen", "it took 18 minutes". In our test, judges flagged at least one invented detail in 23 of the 28 voice-matched drafts. Those drafts still won on voice, which is exactly the problem: the invented line reads the most human.

Treat every number, duration, name and "I did" in an AI draft as a placeholder. Replace it with what really happened, or delete it.

4. Cut references the reader can't see

When a tool researches a topic before writing, names from its sources can leak into the tweet: "following X's staged approach", "as the guide suggests". Your reader never saw that guide. Either name the source properly and link it, or rewrite the line in your own words.

5. Rewrite the first line, and keep it short

On scroll-stopping power the voice-matched drafts were only on par with the template-based assistant. Its short, contrarian first lines were what the judges rewarded on that measure. The lesson is not to copy the template — it is to spend most of your editing time on the opening line. Make it short, specific and something you would actually say.

6. Check the length

Voice-matched drafts run long: in our test, 36% of the winning drafts went over X's 280-character limit once links and emoji were weighted the way X counts them. Paste the draft into the Twitter character counter before you post, and see the X character limit guide for how the counting works. If it really needs more room, turn it into a short thread with the thread generator.

A checklist before you post an AI draft

• Did the model see real posts of yours, not just a description?
• Does it end with a question only because AI likes to? Delete it.
• Is every number, name and "I did" true? If not, fix it or cut it.
• Does it mention a source the reader can't see?
• Would you say the first line out loud?
• Is it under 280 weighted characters?

How xVirals does it

xVirals is built around step 1. It reads the public posts of an X account, builds a style profile from them, and drafts tweets, quote posts and replies with examples of that account's real writing in the prompt. It writes several candidates and ranks them, partly by how close they are to the account's own posts. It never posts for you and stores no X login: you review, edit and publish every draft yourself — which is where steps 3 to 6 happen.

The test above also showed us what to fix: invented details and drafts that run past 280. We're working on both, and we will update this guide when we have new measurements.

Frequently asked questions

Related

• X (Twitter) character limits
• Twitter character counter
• Twitter thread generator