Writing AI Social Media Captions That Don't Sound Like AI Wrote Them

There’s a specific, recognizable cadence to a bad AI-generated caption: an em-dash-heavy opening line, a rhetorical question, three bullet-style benefits, and a closing line that starts with “So.” It’s not that the AI is incapable of writing well — it’s that a generic prompt gets a generic answer, and most people asking for “a caption about X” are giving the model nothing to differentiate on.

The problem isn’t the model, it’s what you gave it

Ask five different people to write a caption about the same product launch and you’ll get five different captions, because each person brings their own voice, their own sense of what matters, and their own relationship with the audience. Ask an AI model the same bare prompt five times and you’ll get five variations on the same generic voice, because you didn’t give it anything to differentiate on either. The fix isn’t a better model — it’s giving the model the same inputs a good human writer would have.

What actually needs to be in the prompt

• A real voice reference, not an adjective. “Write it in a friendly tone” produces the same generic friendliness every time. “Write it the way you’d explain this to a friend who just asked what’s new, keeping the dry humor and skipping the exclamation points” gives the model something specific to hold onto.
• What NOT to say, not just what to say. Telling a model to avoid the clichés you’ve actually seen it produce — the em-dash opener, the rhetorical question, the “in today’s world” framing — works better than a generic “sound natural” instruction, because it’s concrete instead of vague.
• The actual constraint you’re writing under. A launch caption written with no context defaults to generic excitement. A launch caption written with “this is the third time we’ve shipped this feature and we’re being upfront that it’s still in beta” produces something with actual texture, because the constraint gives the model something real to write around.
• One specific detail, not a category. “Mention that it’s fast” is generic. “Mention it cut our own posting time from twenty minutes to ninety seconds” is a sentence a human would actually write, because it’s a real number attached to a real claim.

Why “personality” profiles work better than one-off prompting

The pattern above — voice reference, exclusions, constraints, specifics — is a lot to retype into every single generation. That’s the actual argument for setting up a persistent brand voice once instead of re-prompting from scratch each time: you write the voice reference, the tone exclusions, and the general constraints one time, and every generation after that inherits them automatically. It’s the difference between briefing a new freelance writer from zero every single time versus working with someone who already knows how you sound.

That’s what a personality profile in Babs actually is under the hood — a reusable, structured version of exactly the inputs above, so a generation prompt can be as short as “launch post for the new integration” and still come out sounding like you, because the voice, the exclusions, and the constraints are already loaded in rather than retyped per post.

The real test

Read the caption out loud. If it sounds like something you’d actually say to a person, in your actual voice, it passed — regardless of what wrote the first draft. If it sounds like a caption, rewrite it. The goal was never “AI wrote a caption.” It was “a caption went out that sounds like you,” and who typed the first draft is genuinely the least important part of that sentence.