ChatGPT writes a competent cold email in seconds. The problem is that everyone's ChatGPT writes the same competent cold email, and both spam filters and tired buyers have learned the pattern. These are prompts built to avoid that: each one either forces a real detail into the email or strips out a tell. Copy them, swap the brackets for real inputs, and read the edit pass near the end before you send anything.
A model trained to be safe and average writes email that is safe and average. On its own that is not a disaster. The disaster is that most people prompt it the same way, from the same three listicles, so the inbox fills with near-identical openers, the same balanced three-item lists, and the same "I hope this email finds you well." Readers pattern-match it in half a second and archive it. A model also cannot know the one specific thing about a prospect that earns a reply unless you hand that thing to it.
So the prompts below are organized around the two moves that actually work: feed in a real detail, or remove a generated tell. If you want the long-form version of writing with AI, our guide to writing cold emails with AI covers the workflow. This page is the prompts.
The opener is the only line that proves you did not blast the email. These six turn a piece of real information into that line.
Why it works: it asks for a reaction, not a recap, and bans the compliment opener that reads as filler.
Why it works: the angles come from their own words, so the email sounds researched instead of guessed.
Why it works: timing is the strongest personalization signal there is, and this keeps it framed around the prospect.
Why it works: it hands you the buyer's real scoreboard, so you can lead with what they are graded on.
Why it works: proof lands harder as a number than as a brag, and the adjective ban keeps it from sounding like marketing.
Why it works: when one-to-one research is not realistic, this gets you personalization at the segment level instead.
Getting that one specific line right is the whole game. Our AI personalization guide goes deeper on sourcing the detail and editing it so it never reads as merge-field output.
The subject decides whether any of the work above gets seen. These four bias toward the plain, internal-note tone that gets opened.
Why it works: a subject that looks like a coworker wrote it beats a subject that looks like a campaign, every time.
Why it works: it bans the two most burned question formats and keeps the rest honest.
Why it works: when the subject and the first line agree, reply rate goes up and spam marks go down.
Why it works: it turns the model into an editor for lines you already wrote, instead of a generator of more.
Five prompts for the email itself, all built to kill the standard AI draft before it starts.
Why it works: the constraints kill the standard AI opener and the standard AI close in the same instruction.
Why it works: the "true for anyone" test is exactly what makes a template feel like a template.
Why it works: it gives you five real angles to test instead of one flat pitch you fell in love with.
Why it works: length is the enemy of cold email, and this is a ruthless editor that shows its work.
Why it works: most cold email is read on a phone in a hallway, and this matches the draft to how it will actually be read.
Most replies come from a follow-up, not the first send. These five keep the follow-ups from turning into empty "just bumping this" noise.
Why it works: it bans the empty bump that trained everyone to ignore follow-ups in the first place.
Why it works: repeating the same pitch louder does nothing, but a genuinely new angle can catch a different nerve.
Why it works: the breakup email often outperforms every email before it, as long as it stays clean and low-pressure.
Why it works: it produces a sequence plan you can load once, not four disconnected emails you write on the fly.
Why it works: intent signals are only useful if the follow-up does not read like surveillance.
The last three are for after the draft exists: catching the tells, handling the reply, and personalizing a batch without the seams showing.
Why it works: it turns the model into a detector of its own tells. Do a human read on top of it, since the model misses its own habits.
Why it works: the reply is where the deal starts, and having the branches ready means you answer in minutes, not hours.
Why it works: it is batch personalization that does not look batched, which is the whole point.
No prompt gets you all the way there. The draft is a starting point, and this four-step read is what turns it into something a person actually replies to. It takes about a minute per email.
One more thing that sits underneath all of this: sending near-identical AI drafts to a big list is one of the fastest ways to damage your sending domain, no matter how good the copy is. The specific ways that happens are in our guide to automation mistakes that hurt deliverability.
Only if you paste the output straight into a campaign. Every prompt here does one of two things: it forces a real, specific detail into the email, or it strips out a phrase that reads as generated. The prompts get you a strong draft in seconds. The specific detail you feed in, and the edit pass at the end of this page, are what make the result yours instead of the same email every other AI user is sending.
The writing tool has nothing to do with it. Inbox providers filter on sender reputation and engagement, not on whether a human typed the words. What does hurt you is sending the same AI output to a large list: identical bodies look like a blast, get low replies, and draw spam complaints, and that is what damages your domain. Vary the specifics per prospect and keep your volume inside a warmed-up sending limit.
Any current model writes a competent cold email. The bigger reasoning models follow multi-part instructions more reliably, which matters for the research and editing prompts where you give five or six rules at once. For quick subject-line variations, a faster model is fine. The prompt matters far more than the model version.
Assume the first line and the last line need work, always. The opener is where AI defaults to a cliche, and the close is where it hedges. Read the whole draft aloud once. If a sentence would be equally true sent to a different company, cut it or replace it with something specific. Most drafts need three or four small cuts, not a rewrite.
Using AI to draft the email is fine. The rules that matter are the sending ones: honor unsubscribe requests, do not use false sender names or misleading subject lines, and include a real physical address where required by CAN-SPAM and similar laws. The tool that writes the copy does not change any of that.
lemlist's AI builds the sequence and personalizes each email from real prospect data, then handles the multichannel follow-up you would otherwise copy-paste by hand. Same idea as these prompts, running on its own. The 14-day trial includes it with no card upfront.
Start your free trial