We Gave ChatGPT, Claude, and Gemini the Same Cold Email Brief — One of Them Made a Critical Mistake
Cold email is one of the highest-leverage skills you can have right now, and AI is supposed to make it faster. But "faster" doesn't mean "better" — and after running the same brief through all three major AI tools, the gap in quality was bigger than we expected. One model wrote something genuinely impressive. One wrote something usable but forgettable. And one produced an email with a fatal flaw that would tank your reply rate before the prospect even finishes reading. By the end of this article, you'll know exactly which AI to use, which prompt to give it, and how to spot the mistake that most people miss.
We Ran the Same Brief Through All Three — Here's What We Found
The brief was simple and realistic: "Write a cold email from a freelance web designer to a small e-commerce business owner. The goal is to book a discovery call. The business sells handmade candles and their website looks outdated."
No extra coaching. No examples. No tone guidance. Just the brief — the way most people actually use these tools.
Here's what came back.
ChatGPT (GPT-4o) delivered a clean, structured email. Subject line: "Quick thought on your site." It opened with a specific observation about the candle business, mentioned a believable pain point (slow load time on mobile), and ended with a soft CTA to jump on a 15-minute call. It felt human. It didn't over-explain. It was good.
Claude (Claude 3.5 Sonnet) went deeper. It crafted a subject line that felt almost personal: "Noticed something about Wick & Glow's checkout page." The body acknowledged the quality of the product before pivoting to the website gap — a subtle but powerful move called mirroring the prospect's pride. The CTA was specific: "Would Thursday or Friday work for a quick call?" That specificity alone lifts reply rates. This was the strongest output of the three.
Gemini (Gemini 1.5 Pro) produced something that looked like a cold email but read like a brochure. It opened with "I hope this email finds you well" — a phrase that signals mass outreach from a mile away. It listed the designer's services in bullet points. It closed with "Please don't hesitate to reach out." Every single one of those choices is a reply-rate killer. That's the critical mistake — and we'll break down exactly why it's so damaging in a moment.
Why Claude's Email Actually Worked (The Psychology Behind It)
Claude didn't just write a better email by accident. It applied a principle that elite salespeople know but most AI users never think to ask for: lead with the prospect, not with yourself.
The worst cold emails open with "I am a web designer with 5 years of experience…" That's not an opener — that's a resume. Nobody asked. The prospect's first question is always "Why should I care?" and you have about three seconds to answer it before they hit delete.
Claude's version opened by naming the specific business, acknowledging what was working (the product), and then introducing the gap. That structure — compliment → gap → solution → ask — is what conversion copywriters charge thousands of dollars to write. Claude produced it without being asked.
Here's the prompt that gets you there consistently: "Write a cold email to [prospect type] from [your role]. Open with a genuine observation about their business. Identify one specific problem they likely have. Offer a clear, low-commitment CTA. Keep it under 120 words. Do not open with 'I hope this email finds you well' or list your credentials."
Those negative instructions matter. Telling Claude what not to do is just as important as telling it what to do. Most people skip that step — their outputs end up generic as a result.
One more thing Claude did that was subtle but smart: it used a two-option CTA ("Thursday or Friday?") instead of an open-ended "let me know if you're interested." Open-ended asks kill momentum. Two options force a micro-decision, and micro-decisions get replies.
How to Use These Tools Together for a Cold Email System That Scales
Here's something most AI comparison articles won't tell you: the real move isn't picking one AI and using it for everything. It's using each tool for what it actually does best.
Use Claude for the first draft. Its writing is the most contextually aware of the three. It understands nuance, mirrors tone naturally, and tends to produce copy that doesn't sound AI-generated on the first read. This is where you want your most important creative thinking to happen.
Use ChatGPT for volume and variations. Once Claude gives you a strong base email, paste it into ChatGPT and run this prompt: "Give me 5 subject line variations for this cold email. Three should be curiosity-based, two should feel hyper-personal. No clickbait. No questions." GPT-4o is fast, iterative, and excellent at generating options you can A/B test.
Use Gemini for research, not writing. Where Gemini genuinely shines is pulling together information — company news, recent product launches, LinkedIn summaries. Run this prompt before you write anything: "Summarize [company name]'s recent activity, any pain points for a business in their category, and three things a [your service] could improve for them." Use that research to feed into your Claude prompt. That's when personalization gets real.
Here's the actual workflow in three steps:
- Research with Gemini (5 minutes) — get the context and pain points
- Draft with Claude (5 minutes) — use the prompt from the previous section, inject the Gemini research
- Optimize with ChatGPT (5 minutes) — generate subject line variants, test different CTAs
Fifteen minutes. One highly personalized cold email system. You can repeat this for every prospect on your list.
The Part Most People Get Wrong
Most people treat AI like a vending machine for cold emails. They type in one line, take whatever comes out, and send it. That's wrong — and it's why their reply rates stay at 2%.
The output you get from any AI is only as good as the constraints you give it. Gemini's terrible email wasn't Gemini being a bad tool — it was Gemini filling in the blanks with the most average, statistically common cold email patterns it had ever seen. When you don't give it guardrails, it defaults to generic. Every time.
The specific mistake in Gemini's email — "I hope this email finds you well" — is a trust signal in reverse. It tells the reader this email was written for nobody in particular, which means it was written for everyone, which means it was written for spam. One phrase. Reply rate gone.
The fix isn't complicated. Before you finalize any AI-generated cold email, run it through this quick filter: Does this email name a specific person, business, or detail that proves I actually looked at their work? If the answer is no, you haven't written a cold email. You've written a template. And templates get deleted.
Key Takeaways
- Claude 3.5 Sonnet: Produces the most psychologically effective cold email copy out of the box — use it for first drafts.
- ChatGPT (GPT-4o): Best for generating variations, subject line testing, and rapid iteration once you have a strong base.
- Gemini 1.5 Pro: Use it for prospect research and context-gathering, not for writing the email itself.
- Negative constraints: Telling AI what not to write (no "I hope this finds you well," no credential lists) is as important as your actual instructions.
- The two-option CTA: Replace "let me know if you're interested" with a specific choice ("Tuesday or Thursday?") — it creates momentum and drives replies.
What to Do Right Now
Open Claude right now and paste this prompt: "Write a cold email from a [your role] to a [specific prospect type]. Open with a genuine observation about their business. Identify one specific problem they likely have. Offer a clear, low-commitment CTA asking for a 15-minute call. Keep it under 120 words. Do not open with 'I hope this email finds you well.' Do not list credentials."
Replace the brackets, hit send, and compare what you get to the last cold email you wrote yourself. You'll immediately see the gap — and you'll know exactly where to tighten it.