How to not sound like AI on X (practical rules, not vibes)

You can usually tell an AI-written tweet in the first five words. It's not just the vocabulary - "game-changer," "delve into," "unlock the power of" are easy to catch and easy to avoid once you know to look for them. The bigger tell is structure: AI drafts tend to open with throat-clearing instead of the actual point, pad the middle with context nobody asked for, and close with a hollow call to action or a "thread:" teaser that promises more than it delivers. Fixing the vocabulary is step one. Fixing the shape is what actually makes a post read like a person wrote it.

The tells, specifically

Before the fixes, it helps to name exactly what reads as artificial, because "sounds like AI" is vague enough to be useless as a checklist.

• Buried hook. The interesting part - the actual news - shows up in sentence two or three, after a sentence of setup nobody needed.
• Banned-phrase vocabulary. "Game-changer," "delve into," "unlock the power of," "in today's fast-paced landscape," and anything that sounds like it belongs in a press release.
• Hollow enthusiasm. Exclamation points and hype words standing in for a specific reason to care.
• The generic close. A "thread:" teaser, a vague call to action, or a question that isn't really asking anything ("What do you think? 👇").
• No specific numbers, names, or details. Generic claims instead of the one real detail that would prove a human actually did the thing being described.

Rule 1: lead with the hook, not the setup

Cut the first sentence of most AI drafts and read what's left - it's usually better. AI models tend to open with context ("After months of work, we're excited to share...") before getting to the actual point. A person telling a friend what happened doesn't do that; they lead with the thing itself. "Shipped dark mode today" beats "I'm excited to announce that after weeks of development, we've finally shipped dark mode" every time, because the second version makes the reader wait for information they could have had in the first four words.

Rule 2: edit every draft, don't post it raw

A good AI draft should need 30 to 60 seconds of editing, not zero. Change at least one word or phrase to something you'd actually say, ideally swapping in a real number or detail from your own work - the specific error message, the actual time it took, the real reason you made the call you made. That one concrete detail does more to make a post sound human than rewriting the whole thing.

Rule 3: ask for several versions, not one

Asking an AI for "a tweet about this" tends to produce the most generic, median version of what a tweet about that topic could look like - because that's what you asked for. Asking for several different angles - a hook, a contrarian take, a question - gives you options to pick from and edit, instead of committing to whatever came out first.

Rule 4: give the model your actual voice

A model that's never seen how you actually write will default to a generic "helpful assistant" tone, because that's the safest average it can produce. Feeding it real examples of how you actually talk - even rough ones - narrows that average toward something that sounds like you specifically, not like a tool trying to sound like everyone.

Rule 5: ask "would I actually say this out loud?"

This is the fastest filter there is. Read the draft back and ask whether you'd say it to a friend in that exact wording. If the answer is "no, but it sounds professional," that's usually the tell that it needs editing - professional and generic are close neighbors, and X rewards neither.

Why this matters more than it seems

Posts that read as AI-generated get scrolled past faster than posts that read as a specific person saying a specific thing - not because people are hunting for AI tells, but because generic writing reads as having nothing new to say, regardless of who or what wrote it. The fix isn't "write without AI help" - it's making sure whatever AI help you use starts from something real and gets edited by someone who'd actually say it that way.

This is the whole premise behind how Puentiac writes drafts. Instead of generating from a generic prompt, it starts from something that actually happened - a feature you shipped, a bug you fixed, a decision made in a real Claude Code or Cursor session - and writes toward your voice, not a generic one. No hashtags, no links, no corporate phrasing, by design. You still review and edit before anything goes out, which is exactly the step that turns a decent AI draft into a post that sounds like you.

Frequently asked questions

The takeaway

AI slop isn't really about using AI - it's about posting the first generic thing it generates without anchoring it to something real or editing it into something you'd actually say. Lead with the hook, start from something specific, and always read it back before you post. That's the whole fix.