The instinct is to blame the model. That's the wrong diagnosis. The model wrote exactly what you asked for. The problem is that a cold email is not a writing task — it's a research task with a writing step at the end, and you skipped the first 90% of it.
The thing nobody wants to hear about cold email
Reply rate is mostly determined before a single word is written. It comes down to three things, roughly in this order:
- Is this person actually a fit? (list quality)
- Is this a reasonable moment to contact them? (timing / trigger)
- Does the email respect their time? (copy)
Copy is third. It matters — bad copy can sink a perfect list — but great copy cannot rescue a bad one. If you email a CFO about a developer tool, no sentence exists that fixes that. This is why "give me a better cold email prompt" is usually the wrong request. The prompt can only improve the third variable.
What a well-built prompt can do is refuse to let you skip the first two. That's the real design goal, and almost no cold email prompt on the internet does it.
The failure mode: confident fabrication
Here is the specific way AI cold emails die.
You give the model a company name and nothing else. It needs an opening line about the prospect. It has no information about the prospect. So it generates something plausible:
"I noticed Acme has been scaling its engineering team rapidly this year — impressive growth."
The recipient works at Acme. They know whether they've been scaling. If they haven't, you've just told them you're a bot. If they have, they've received forty of these this month and the sentence carries no information. Either way, you lose.
The model didn't malfunction. It filled a gap, because filling gaps is what it does. The fix is architectural: build the prompt so that gaps cause a stop, not a guess.
Merge variables versus real personalization
If you use {{ variables }} in your cold email templates, you already understand half of this. But there's a distinction that most outbound stacks blur, and blurring it is expensive.
Tier 1 — mechanical fields. {{ firstName }}, {{ company }}, {{ role }}. These populate from a CSV. They are table stakes. They add exactly zero credibility, because every automated email in the recipient's inbox has them. Getting them right is invisible; getting them wrong is fatal. That's the whole value.
Tier 2 — research fields. {{ triggerEvent }}, {{ specificObservation }}, {{ relevanceReason }}. These are the entire game. And critically: they cannot be mail-merged. There is no column in your CRM that contains "the thing that happened to this specific company last week that makes my product relevant right now."
The most common outbound mistake is treating Tier 2 like Tier 1 — dropping in a semi-generic variable like {{ industry }} and calling it personalization. It isn't. Personalization is not inserting a variable; it's proving you spent time.
Tier 3 — offer fields. {{ whatWeDo }}, {{ proofPoint }}, {{ theAsk }}. Constant across the campaign. Write once, test relentlessly.
The economics follow directly. Tier 2 costs 3–5 minutes of human or enrichment time per prospect. If you're sending 2,000 emails a week, you cannot afford it — which means you're sending Tier 1 emails, which means you're playing a pure volume game and should stop pretending otherwise. Pick a lane. Both lanes work. The middle doesn't.
Anatomy of an email that gets answered
Subject line. Under five words. Lowercase. It should look like something a colleague typed, not something a marketing department approved. Avoid "quick question" — it's been burned to the ground.
First sentence. About them. Always. This is the only sentence most recipients read, and its job is to answer "why is this in my inbox." If your first sentence would make sense sent to their competitor, delete it.
The bridge. One sentence connecting what you observed to a cost, risk, or friction they're already experiencing. Not to your feature. They don't care about your feature yet.
Proof. One concrete thing. A named customer that resembles them, or a number. "Trusted by industry leaders" is worse than saying nothing, because it consumes attention and delivers no information.
The ask. One. Answerable in under ten seconds. This is where most emails collapse — they open well and then ask for a 30-minute call from someone who has known you for eleven seconds. The correct first ask is a yes/no question or a one-line reply. Earn the meeting on the second exchange.
Sign-off. First name. Nothing else.
Total: under 90 words for most B2B. Under 60 for SMB. Under 40 for a LinkedIn DM.
Three angles, not three rewordings
When people ask AI for variants, they get the same email with synonyms swapped. That tests nothing. If Variant A and Variant B say the same thing in different words, whichever wins tells you nothing you can apply to the next campaign.
Real variants differ in strategy:
- Observation-led — opens on something specific you saw. Highest credibility, highest research cost.
- Proof-led — opens on a result you got for a company like theirs. Works when your proof is genuinely strong and genuinely comparable.
- Question-led — opens with a question only a researched sender could ask, and withholds the pitch entirely until they reply. Usually the highest reply rate on cold lists, and the one most people are too impatient to try.
Test across angles. Then, within the winning angle, test wording.
The follow-up rule
"Just bumping this to the top of your inbox" is a confession that you have nothing to add. Every follow-up must carry new information — a different angle, a relevant example, a piece of research, or an explicit permission-to-close ("if this isn't a priority, tell me and I'll stop"). Two to three follow-ups, spaced four to seven days. Then stop.
What to actually do with this
Build your prompt as a gate, not a generator. Before it writes anything, make it grade your research inputs and tell you where they're weak and where to find better ones. Let it refuse. A prompt that refuses to write is doing more for your reply rate than one that cheerfully produces a hundred variations of nothing.
Then hold it to hard constraints: no invented facts, one ask, first sentence about them, word ceiling enforced. Make it output a self-check at the end — the single most useful question being "could this email be sent to their competitor unchanged?" If the answer is yes, it isn't a cold email. It's a broadcast with a name field.