The cold email prompt pack

Most people ask a model to write a cold email and get back something that reads exactly like every other cold email written by a model. The reason is almost always the prompt: it asks for a tone instead of supplying facts, so the model fills the gap with the average of everything it has seen.

These nine are structured to prevent that. Each supplies specifics, forbids the standard failure modes explicitly, and asks for something that can be checked. The critique prompts matter as much as the drafting ones, because the fastest improvement in cold email is usually deletion.

One boundary worth stating. These help you write. They do not tell you whether you may send. Consent and legal basis under CAN-SPAM, the GDPR or CASL are yours to establish, and this page is not legal advice.

1. Research a prospect

"Here is a company: [name, website, one-line description]. Here is the person: [name, title]. List five specific, checkable facts about this company that would be relevant to someone selling [what you sell]. For each fact, say where you would verify it. Do not include anything you are inferring rather than reading. If you can only find two, give me two."

2. Draft an opener

"Write a cold email of at most 90 words. Recipient: [name, title, company]. What I sell: [one sentence]. The specific thing I noticed about them: [fact from prompt one]. Rules: no adjectives about my company, no phrase that would be true of any other recipient, no more than one question, end with a question rather than a call to action. Include a plain unsubscribe line as the final line."

3. Critique a draft

"Here is a cold email: [paste]. For each sentence, tell me whether it would still be true if I swapped the recipient for a different company in the same industry. List every sentence that survives that swap, because those are the sentences doing no work. Then tell me the single sentence a busy reader would stop at, and why."

4. Cut it in half

"Here is a cold email of [n] words: [paste]. Rewrite it at half the length without losing the specific fact about the recipient or the question at the end. Tell me what you removed and why each removal was safe."

5. Write the follow-up

"Here is the first email I sent: [paste]. Write the follow-up for four days later, in the same thread. It must not restate the first email, must not apologise for following up, and must introduce one new specific thing: a different angle, a concrete example, or a reframing of the problem. At most 60 words."

6. Handle a specific objection

"A prospect replied to my cold email with: [paste their reply]. Write a response that answers the actual objection rather than talking around it. If the honest answer is that we are not a fit, say so and say why in one sentence. Do not offer a call unless the reply suggests they want one. At most 80 words."

7. Rewrite for a different seniority

"Here is an email written for [title]: [paste]. Rewrite it for [different title] at the same company. Change what the message leads with based on what that role is actually measured on, not on a tone shift. Tell me in one line what you changed the lead to and why."

8. Generate subject lines

"Here is the email body: [paste]. Give me eight subject lines. Rules: none longer than six words, none containing my company name, none phrased as a question, none using a word that would trip a spam filter such as free, guarantee or urgent. Two of the eight should be plain and boring on purpose. Rank them by how likely a busy person is to open without feeling sold to."

9. Audit a whole sequence

"Here is a four-step sequence: [paste all four]. Check three things. First, does each step introduce something new, or is any step a restatement? Second, does the sequence escalate in pressure, and if so where does it cross from persistent into annoying? Third, does the final step make it easy to say no? Give me a specific edit for each problem you find."

The knowledge base holds your own facts, offers and voice, so a draft argues from what is true about your business rather than from what sounds plausible. kb_search is what an agent uses to find that material.

Every draft an agent writes waits in the approval queue with the thread, the evidence, the confidence and the cost visible. Approve, edit-then-approve, or reject with a reason that feeds the next run. External sends are review by default in the policy table, so this is the path whether you configured it or not.

Model spend is metered and capped. There is a soft alert at eighty percent of your allowance and a hard stop at a hundred, and the hard stop applies to AI only. Manual sending never stops because a model allowance ran out. Bringing your own key keeps model spend off the allowance entirely.