Here's the trap. You ask an assistant to write a cold email, it produces something fluent and plausible in four seconds, and you send it. It reads exactly like the other nine emails that person received that morning, because they were all written the same way by the same models from the same kind of instruction.
The drafting step was never the bottleneck. Writing a competent paragraph is the thing these tools are best at and it is the least valuable part of the job. What determines whether the email works is what you knew before you started writing, and what you forbade.
So here are four prompts, in the order you use them, with the constraint block that matters more than any of them.
Prompt one: research, before any drafting
This is where the value is. You are asking for raw material and specifically not for prose, because the moment it starts writing it stops looking.
The third item is the one worth keeping. A list of what your prospect does not care about is more useful than a list of what they might, because it removes the two sentences you were going to waste. The final instruction matters too: without it you get confident inference presented as research.
Prompt two: the first touch
Ninety words is doing real work as a constraint. Almost every bad cold email is bad because it is long enough to include the parts that make it generic, and a hard ceiling forces the model to spend its budget on the specific observation.
Prompt three: the follow-up that is not a bump
The instruction not to reference the previous email is the whole prompt. A follow-up that says "just bumping this" transfers the cost of your persistence onto the recipient. One that arrives with something new does not.
Prompt four: the breakup, written to be answerable
This one produces replies more often than the other three combined, in most people's experience, and the reason is that it is the only email in the sequence that costs the recipient nothing to answer.
The constraint block that does the actual work
Everything above works because of what it forbids rather than what it asks for. Keep a standing block and append it to every drafting prompt:
- No em dashes. They are the single clearest tell of unedited model output in 2026, and this site bans them in its own editorial rules for that reason.
- No "not just X, it's Y" constructions, and no sentence that begins by naming a problem the recipient has not told you they have.
- No adjectives in front of your own product. Powerful, seamless, game-changing and unlock all mean the sender could not think of a specific benefit.
- No claim the research does not support. This is the one that prevents the most damage, because a confidently wrong sentence about someone's business ends the conversation faster than a boring one.
- Name the register. British or American English, and one real person's writing to imitate if you have one. Otherwise you get the average of everything, which is what generic means.
Where this stops working
At volume. These prompts assume you are pasting real research about a real person, and that is roughly a five-minute job per prospect even with an assistant doing the reading. Fifty a week is achievable; five hundred is not, and pretending otherwise is how the output goes back to being generic.
That ceiling is the same one we hit costing the manual approach: a complete outbound motion runs on $17 a month and tops out around fifty messages a week. Past that you are buying sending capacity rather than better writing, and the honest comparison of what that costs is in the itemised sending model.
The short version: spend your prompt budget on the research step and on the list of forbidden moves. The part where it writes the paragraph will take care of itself, and it was never the part that decided whether anyone replied.