How to write tweets with AI that don't sound like AI

Ask a general-purpose AI for a tweet with no context and you get the average of everything it has read: a hook like “Here's the thing:”, a list of three, an emoji or two, and a closing question. Any one of those is fine. Together they form a pattern readers scroll past.

AI writing isn't bad by nature. A model with nothing specific to go on fills the gaps with the most common phrasing, and the most common phrasing is, by definition, what everyone else is posting.

Common tells

• Openers like “Let's dive in”, “Here's the thing” or “In today's fast-paced world”.
• Words people rarely use in conversation: “delve”, “realm”, “tapestry”, “a testament to”.
• Every post ending in a question to “drive engagement”.
• Lists of exactly three adjectives.
• Emoji used as bullet points.
• Confident claims with no example, number or name behind them.

The right-hand column is your job. The best AI-assisted tweets start with something only you know.

Before you ask for a single tweet, write down how you sound and give it to the model every time:

• Three voice traits, each with what it means and what it doesn't.
• Do's and don'ts, including words you never use.
• Two or three of your own best posts as examples.
• Who you're writing for and what they care about.

How to find your brand voice walks through writing this. Pasting it into every chat gets tedious; tools with a saved brand kit read it automatically.

Specificity is the biggest single fix. “Write a tweet about pricing” gets a generic tweet. A prompt with a real situation gets something worth posting.

The second prompt included the real numbers, what the founder expected, and what actually happened. The model only had to shape it.

Treat every AI draft as a first draft from a fast, eager intern. A quick edit pass:

1. Cut the first sentence if it's warm-up. Start at the point.
2. Replace every vague claim with a specific one, or delete it.
3. Remove anything on your banned-words list.
4. Read it aloud. If you wouldn't say it, rewrite it.
5. Check the length with the character counter and preview it.

If you use a general chatbot, a fixed prompt structure saves time and keeps results consistent. Fill in the brackets each time:

• Who it's for stops the model writing for everyone.
• Voice and banned words cut most of the generic phrasing before it appears.
• What happened is the raw material. Without it, the model invents something plausible and bland.
• The point keeps the draft from wandering.
• Three versions give you options to combine, which is usually faster than regenerating.

Once a version is close, ask for targeted changes, like “make the first line shorter” or “replace the second sentence with the actual number”, instead of starting over.

1. Keep a running list of raw notes: things that happened, opinions, numbers.
2. Pick a note and give the AI the note plus your voice guide.
3. Ask for two or three versions, not one.
4. Pick the best, edit it by hand, and fix any line that sounds generic.
5. Preview it, then post or schedule.

Extudio's tweet generator is built around this. It reads your brand kit (voice, do's and don'ts, glossary) before every draft, takes a rough idea as the starting point, and lets you rewrite a single tweet and accept, dismiss or regenerate it. AI generation is on the Creator and Pro plans; the editor and preview are free.

X's rules focus on spam, manipulation and inauthentic behavior rather than on how a post was drafted. The real risk is to your reputation: posts that are generic, wrong, or claim experiences you didn't have.

Keep the facts and stories yours, check anything the model states as fact, and review every post before it goes out. Then you're writing with a tool, not handing over your voice. For the writing itself, how to write a good tweet still applies.