How to prompt AI for cold email that doesn't sound like AI
A practical guide to prompting AI for cold email that reads as researched, not generic — specificity, constraints, real detail, and the stock phrases to ban.
By Norbelys Chirinos, Co-founder
Founder-reviewed ·How we research and correct articles
You can tell within the first line whether a cold email was prompted with “write a cold email about X” or with something more specific. The generic prompt produces generic copy — grammatically fine, structurally correct, and instantly forgettable. The fix isn’t avoiding AI. It’s prompting it like you’d brief a junior copywriter who has never met the prospect: give them the facts, name the traps, and tell them exactly what not to write.
This is the practical version of that brief — the prompt patterns that produce usable drafts and the phrases worth banning outright.
Why the lazy prompt fails
“Write a cold email to CEOs about our sales tool” gives the model nothing to differentiate the output from the millions of similar emails in its training data. It doesn’t know who the CEO is, why now, or what makes your tool different from the last five that pitched the same audience. So it fills the gaps with the safest, most generic language available — which is exactly the language every recipient has learned to skim past.
The model isn’t bad at cold email. It’s bad at guessing what you didn’t tell it.
Give it the same inputs a good rep would use
A prompt that works has the same shape as a proper campaign brief: a real audience, a specific trigger, a stated pain, and constraints on length, tone, and claims. The difference between a prompt and a brief is mostly formality — you can write the same information as a paragraph instead of a template and get the same lift.
What consistently improves output:
- Name the trigger. “They just posted three SDR roles” beats “growing company” every time — it gives the model a reason for the email to exist right now instead of an evergreen pitch that could have been sent any day.
- State the pain in the buyer’s words, not your product’s words. “Scaling outbound before deliverability controls exist” is a sentence a founder would recognize about themselves. “Struggling with email infrastructure” is a sentence only a vendor would write.
- Hand it real proof, not permission to invent some. A specific fact, a real customer result, or an honest constraint beats an adjective every time, and it keeps the model from manufacturing a stat that doesn’t exist.
- Set hard limits. Word count, one CTA, no fake familiarity, no urgency language unless something is actually time-bound. Constraints are what stop the model from drifting into its default voice.
Ban the tells
Certain phrases are AI-cold-email tells not because a model invented them, but because the model reaches for them by default when the prompt is vague — and years of generic outreach have already trained readers to recognize them as the first sign of a message not worth finishing.
| Instead of this | Prompt for this |
|---|---|
| “I hope this finds you well” | Skip the opener entirely — start on the trigger |
| “I wanted to reach out because…” | State the reason in the first six words |
| “As a [role], I’m sure you know…” | Name the specific thing they’d know, not the fact that they know things |
| “We’re passionate about helping companies like yours” | Name the one thing you help with, and the evidence for it |
| “In today’s fast-paced business environment” | Delete — it’s true of every era and says nothing |
| “I noticed you’re a leader in your space” | Reference the actual, specific trigger you found |
| “Just circling back” (as a follow-up opener) | Lead the follow-up with new information, not a nudge |
| “Let me know if you have any questions!” | Give one specific, low-friction next step instead |
None of these phrases are wrong in isolation. They’re wrong because they’re the default output of an underspecified prompt, and a reader who’s seen a thousand cold emails recognizes the pattern faster than they read the content.
A prompt structure that holds up
Write a [word count]-word first-touch email to [specific audience —
role, company size, context].
Trigger: [why this person, now]
Pain: [in their language, not ours]
Offer: [one sentence]
Proof: [a real, specific fact — do not invent one]
Constraints:
- No stock openers ("hope this finds you well," "wanted to reach out")
- One CTA, low-friction
- No claims we can't back up
- Tone: [direct / casual / formal — pick one]
Output 2 subject line options and one body draft.Run the same structure through a follow-up prompt, but require the second message to add a new piece of information rather than restate the ask — that one rule eliminates most of the “just following up” filler that makes sequences feel like nagging instead of a conversation.
Check the output before it goes anywhere
A draft that passes the eye test can still trip mechanical filters or read worse in a subject line than it does in your prompt window. Run the subject line through the subject line tester and the body through the spam word checker before you trust it — both catch things a careful read can miss, especially at the volume a good prompt lets you produce.
None of this replaces judgment. A well-prompted draft is still a draft. Cold emails that get replies are ones a human read, cut, and checked against what they actually know about the prospect — AI gets you to a better starting point faster. It doesn’t get you to done.
The full set of tested prompt structures — for openers, follow-ups, breakups, and reply drafts — lives in the Prompt Library.
Frequently asked questions
Will AI-written cold email hurt deliverability?
Not because it's AI-written specifically. Deliverability responds to authentication, complaint rate, and engagement, not to writing style. What does hurt deliverability is generic, low-effort copy at volume, because it gets ignored or reported more often — and that's a risk from bad prompting, not from AI itself.
Should I disclose that AI was used to write the email?
There's no legal requirement to disclose AI drafting for standard B2B cold email in most jurisdictions, and disclosure isn't the substitute for quality. The more useful standard is whether every claim in the email is true and every detail is real, regardless of who or what typed it first.
What details make a prompt actually work better?
A real trigger, the pain stated in the prospect's own language, one piece of verifiable proof, and hard constraints on length and tone. Vague prompts produce vague output because the model has nothing specific to work from.
Can I use the same prompt for every audience segment?
You can reuse the structure, but not the content. The trigger, pain, and proof fields need to change per segment, or every version of the email will read the same way to someone who receives more than one.
How do I know if a draft still sounds like AI?
Read it out loud. If it could have been sent to any company in the audience without editing a single sentence, it's still generic. A draft that sounds researched references something true and specific that only applies to that one prospect.