AI can write a marketing email in ten seconds. The problem is that most of what it produces sounds like AI wrote it. If you've felt that "close, but not quite right" feeling reading your own AI drafts, the issue usually isn't the model. It's the prompt.
Learning how to prompt AI for better marketing copy is quickly becoming one of the highest leverage skills in marketing. AI gets you most of the way to a finished piece of copy. The last stretch, the part that actually sounds human, still depends on how you ask for it.
The simplest fix is also the most overlooked. Tell the model outright: "make this sound more human, more genuine." If the first draft still feels stiff, say it again. Repetition works. AI models default toward a safe, corporate register unless you keep pushing them out of it.
This applies across tools. ChatGPT tends to be the most reliable for copywriting out of the box. Claude can get you just as far, but it usually takes a bit more back and forth to land the same tone. Whichever tool you use, don't accept the first draft as final. Treat it as a rough cut you're directing, not a finished asset.
Certain formatting habits are dead giveaways that a piece of copy came straight from a model with no editing. Bullet points in an email are one of the biggest. They make a message look like a corporate update instead of a note from a real person.
Em dashes are another. AI models default to them constantly, and readers have started to associate them with generated text. There's a system-level fix worth knowing: in Claude, you can set a permanent preference telling it to never use em dashes. Set it once in your system preferences and it holds across every future output, so you're not fixing the same thing draft after draft.
Beyond formatting, watch for generic phrasing that AI tools reach for by default, words like "leverage," "seamless," "robust," or "elevate." None of those are how a person actually talks. When you spot them, ask the model to replace them with something a colleague would say out loud.
AI tools are pattern matchers. If you give a model nothing to work from except a topic, it falls back to its default tone, which is polished but generic. Give it a real example of copy you already like, and ask it to match that voice, and the output improves dramatically.
This is worth doing even for small pieces of copy. Paste in a past email, social post, or paragraph that landed well, and prompt the model to write the new piece "in this exact tone." You're not asking it to be creative from nothing. You're asking it to copy a voice you already know works.
Pair that with a direct framing prompt: "write this like a colleague emailing a colleague, no company voice." That single instruction does more to fix tone than almost any other adjustment.
The most common mistake is treating the first AI draft as done. AI gets you roughly ninety percent of the way to something usable, but that remaining stretch is where your judgment, taste, and brand voice actually live. Publishing the first draft unedited is the fastest way to end up with copy that reads as generic across every channel.
The second mistake is prompting once and expecting perfection. Good AI-assisted copy usually takes two or three passes: one for structure, one for tone, and one for trimming anything that still sounds corporate.
Pick one piece of copy you're about to publish and run it back through your AI tool with three specific instructions: make it sound more human and genuine, remove every em dash, and rewrite it like a colleague emailing a colleague. Compare that draft to your original.
If you use Claude regularly, set "never use em dashes" in your system preferences today so you're not fixing it manually every time.