I Sent the Same Cold Email Brief to ChatGPT, Claude, and Gemini. Only One Wrote Something a Real Prospect Would Answer.

Most people assume all AI chatbots write the same generic, robotic sales copy. That assumption is costing you replies. I gave ChatGPT, Claude, and Gemini the identical cold email brief — same product, same target, same word limit — and the results were nowhere close to equal.

One model wrote something that sounded like a template. One overcorrected into "casual" and got weird. One nailed the tone that actually gets a reply. In this article, you'll see the exact prompt I used, the three outputs side by side, and the specific pattern that separates a cold email that gets ignored from one that gets answered.

The Brief: What I Actually Asked All Three Models

Here's the exact prompt I gave every AI, word for word:

"Write a cold email to a marketing director at a mid-size SaaS company, offering a done-for-you LinkedIn content service. Keep it under 100 words. No hype, no exclamation points, sound like a real person wrote it, and end with a low-pressure call to action."

That's it. No extra coaching, no examples, no back-and-forth. I wanted to see what each model defaults to when given a clean, realistic brief — because that's exactly how most people actually use these tools.

ChatGPT came back fast and polished, but it leaned into the classic "I hope this finds you well" energy. It used phrases like "I wanted to reach out" and "streamline your content strategy" — safe, professional, and completely forgettable. If you've gotten a cold email like this before, you already skimmed past it in your head.

Gemini went the opposite direction. It tried to sound casual and ended up sounding like it was trying too hard — one line literally said "Marketing is hard, we get it!" with an exclamation point I explicitly told it not to use. It ignored a direct instruction, which matters more than people realize.

Claude was the only one that sounded like an actual human wrote it in five minutes between meetings. Short sentences. Specific detail about the recipient's likely pain point. No fluff word. It closed with "Worth a quick look?" instead of "Let's schedule a call" — a tiny difference that completely changes how pushy the email feels.

Why Claude Won: It's Not About Vocabulary, It's About Restraint

Here's the insight most comparisons miss: the best AI-written cold email isn't the one with the smartest phrasing. It's the one that says the least while implying the most. Claude's win wasn't about better words — it was about what it left out.

Cold emails fail for one reason above all others: they try to sell too hard, too early. ChatGPT's version had three separate value propositions crammed into 90 words. Gemini tried to build rapport and pitch the offer in the same breath. Claude picked one single benefit and let the rest of the email support it.

This connects to a mental model worth stealing: think of a cold email like a movie trailer, not a movie. Your job isn't to explain the whole plot (the entire service, every feature, every benefit). Your job is to make someone curious enough to ask for the full story. Claude intuitively trailer-ized the pitch. The other two tried to show the whole film in one email.

There's a technical reason this happens too. Claude tends to weight instructions like "keep it under 100 words" and "no hype" more literally than ChatGPT, which often defaults to its trained politeness patterns even when told not to. Gemini, in my testing, is the most likely to add enthusiasm markers (exclamation points, "we get it," "exciting") unless you explicitly ban them in the prompt itself, not just imply it.

If you want AI to write emails that don't sound like AI, the fix isn't better vocabulary prompting — it's stricter constraint prompting. Tell it what to remove, not just what to include.

How to Use This Today: The 3-Step Cold Email Workflow

Here's exactly how to apply this in the next 10 minutes, no matter which tool you're using.

Step 1: Write the constraint-first prompt. Instead of "write me a cold email," structure it like this: "Write a cold email to [specific role] about [specific offer]. Under 80 words. No greetings like 'I hope this finds you well.' No exclamation points. No more than one benefit mentioned. End with a low-pressure question, not a call to schedule." Notice how much of this prompt is telling it what NOT to do.

Step 2: Run it through Claude first. Based on this test, Claude is currently the strongest at literal instruction-following for tone-sensitive copy. Use Claude 3.5 Sonnet if you have access — it's the version I used here and it consistently produces the least "AI-sounding" output for short-form persuasive writing.

Step 3: Edit for one detail. Take whatever Claude gives you and add ONE specific, real detail about the actual person or company you're emailing — something it couldn't have known. A recent LinkedIn post they made, a product launch, a specific metric on their site. This single addition does more for reply rates than any prompt engineering trick, because it proves the email wasn't sent to 500 people at once.

Do this for your next 10 outbound emails and track your reply rate against whatever you were doing before. The difference won't be subtle.

The Part Most People Get Wrong

Most people think the AI model matters more than the prompt. That's backwards. I could give ChatGPT a better-constructed prompt than the one above and it would likely beat Claude's default output — the model isn't the ceiling, the instructions are.

The real mistake is treating cold email generation as a one-shot task. People type a vague request, get generic output, and conclude "AI cold emails don't work." What actually happened is they never told the AI what bad output looks like — only what good output should include.

The second mistake: not testing tone-sensitive prompts across multiple models before committing. Every model has a different default personality baked in from training. ChatGPT defaults corporate-safe. Gemini defaults enthusiastic. Claude defaults measured and literal. Knowing these defaults means you can pick the right starting point instead of fighting your tool for ten prompts trying to un-teach its personality.

Key Takeaways

• Constraint-first prompting wins: Tell the AI what to remove (exclamation points, greetings, multiple benefits) before telling it what to include.
• Claude currently leads on tone-sensitive short copy: It follows literal instructions like word limits and banned phrases more consistently than ChatGPT or Gemini.
• Gemini defaults to enthusiasm: Explicitly ban exclamation points and hype words in your prompt, or it will add them anyway.
• One real detail beats ten prompt tweaks: A specific fact about the actual recipient does more for reply rates than any AI-generated phrasing.
• Trailer, not movie: Your cold email should create curiosity about one benefit, not summarize your entire offer.

What to Do Right Now

Open Claude right now and paste this prompt, swapping in your actual offer: "Write a cold email to [role] about [offer]. Under 80 words. No greetings, no exclamation points, one benefit only, end with a low-pressure question." Then add one real, specific detail about your actual recipient before you hit send.