Cold emails written by AI are easy to spot. Not because the grammar is off — it usually isn’t — but because they all use the same three moves: a forced personal observation in the opener, a one-sentence pitch, and a soft ask that says “would love to connect” or something equally forgettable. If you’ve gotten five of these in a week, you recognize the shape immediately.
The problem isn’t the model. It’s that the default prompt is something like “write me a cold email to a marketing director at a SaaS company” — precisely vague enough to produce the blandest version that technically satisfies the request. The model isn’t guessing wrong; it’s pattern-matching on what cold emails in its training data looked like, and those were mostly generic to begin with.
Here’s what actually changes the output.
Give it the receiver, not a job title
“Marketing director at a SaaS company” is a category. Categories produce category-shaped output. Instead, paste in something real: a paragraph from their LinkedIn About section, a quote from a recent podcast they appeared on, a summary of a piece they published. Even two or three sentences of actual context about a specific person produces a noticeably different opener than job-title-only prompts do.
You don’t need to write the email yourself after that — you need to supply the raw material and let the model do the assembly. That’s a different workflow than asking it to generate from scratch.
Tell it specifically what you’re NOT trying to say
This is the instruction most people skip. Before you ask for the email, write out three or four things you don’t want in it: phrases that feel salesy, angles that would feel presumptuous, anything that sounds like it came from a sequence tool. Then include that list in the prompt explicitly.
Something like: “Don’t open with a compliment about their work. Don’t say ‘I came across your profile.’ Don’t mention ROI or ‘synergy.’ Don’t end with ‘would love to connect’ or ‘pick your brain.'”
The effect is real, and the mechanism is the same one that applies to any underspecified prompt: a model without a clear negative example defaults to the statistically safest version of the task, which is exactly the generic structure you’re trying to avoid. “Make it feel genuine” is underspecified. A banned-phrase list is not.
Specify the one thing you actually want to accomplish
Not the general goal (“get a reply”) — the one specific thing this particular email should do. Do you want them to watch a two-minute video? Reply yes or no to a single question? Agree to a 20-minute call next week, not “sometime”? The more specific the intended action, the tighter the email. Models are genuinely good at working backwards from a concrete outcome. They’re worse at inferring which of several plausible outcomes you actually care about.
Ask for three versions, pick from the seams
Ask the model to write three variations with meaningfully different openers — not different words, different strategies. One that leads with a question. One that leads with a specific observation about something they published. One stripped to five sentences with no opener at all.
You probably won’t use any of the three as-is. What you’re looking for is which opener doesn’t make you wince, and then you rewrite from there. The model is generating raw material for you to edit, not a finished deliverable. Treating the first output as something you just clean up a little is usually what produces the generic result you were trying to avoid.
The honest caveat
None of this makes AI a replacement for actually knowing who you’re writing to. If you don’t have anything real to say about the person or why you’re contacting them specifically, the model will invent plausible-sounding specificity — and that reads as fake because it is. The prompt technique helps with structure and avoiding clichés; it doesn’t manufacture genuine relevance from nothing.
Claude tends to do better at this particular task than ChatGPT in my experience, specifically because it pushes back when the brief is thin. Ask it to write a cold email with almost no context and a vague goal, and it’ll often ask a clarifying question rather than just generating. That friction is useful — it surfaces the gap in your brief before you send something hollow.
One more thing: the email that reads best to you is often not the one that gets a reply. Short and specific almost always beats long and polished. If AI is helping you write longer, more comprehensive cold emails, it’s probably making them worse.
Try writing your next one with a hard 100-word limit. Give the model that constraint up front, before anything else, and see what it cuts first.
