How to Write AI Email Prompts That Get Replies Instead of Getting Deleted

"Write me a follow up email" is one of the most common prompts anyone types into an AI tool, and it is also one of the most reliably disappointing. What comes back is fluent, properly capitalized, generically polite, and almost never something you would actually send without editing half of it. The greeting is "I hope this email finds you well." The body restates what you already said in a message the model never actually saw. The closing offers to "circle back at your convenience." None of it is wrong, exactly. It also could have been sent to anyone, about anything, which is precisely the problem: an email that reads like it was written for everyone reads, to the one person actually receiving it, like it was written for no one in particular.

Why a generic email prompt produces a generic email

A language model handed no recipient, no relationship, and no specific ask fills every one of those gaps with the safest, most broadly applicable version of professional writing it has seen, and safe, broadly applicable writing is functionally indistinguishable from every other email using the same safe, broadly applicable phrasing. Ask for "a follow up email about the project" with nothing else and the model cannot know whether this is the third gentle nudge to a slow client or the first check in with a new hire, so it writes something polite enough to fit either situation and specific enough to fit neither. The fix is not asking for a "more personal" or "less generic" email, since those instructions are exactly as vague as the ones that produced the generic draft in the first place. The fix is supplying the specific facts a real email would actually contain.

The five layers an email prompt needs

Who is receiving it and what the relationship already is: a first cold email to a stranger, a message to a manager you talk to daily, and a note to a vendor who missed a deadline all need a different opening and a different level of formality, and a model told none of this defaults to a generic middle ground that fits none of them well. The actual goal of the email, stated as an action, not a topic: "get a yes to a 15 minute call next week" is an instruction the model can write toward; "touch base about the partnership" gives it nothing to build the email around and produces a message that says a lot while asking for nothing. The specific facts, dates, or context that belong in the body: what already happened, what already got discussed, what number or deadline is relevant, since a model asked to "remind them about the invoice" with no invoice number or amount writes a vague reminder that reads as vague to the recipient too. Tone and length: a two line internal nudge and a four paragraph client update are different writing tasks, and stating the tone directly (direct and brief, warm but firm, formal) keeps the model from defaulting to the safest, most padded version of politeness. A constraint against filler: naming the phrases to skip, "I hope this email finds you well," "just circling back," "at your earliest convenience," removes the model’s default opening and closing lines more reliably than hoping it skips them on its own.

Four email tasks worth knowing how to structure

Cold outreach: give the model the recipient’s actual role or situation, one specific and true reason you are reaching out to them and not a hundred other people, and the one concrete ask, then request a short email under 120 words, since a longer cold email just reads as a longer reason to delete it. A follow up after no response: state how many times you have already reached out and what, if anything, changed since the last message, and ask for something shorter than the original, not a repeat of it, since a follow up that just restates the first email reads as a nag rather than as new information. A difficult or sensitive message, declining a request, pushing back on a deadline, delivering bad news: state the actual constraint honestly and ask the model to lead with the point rather than burying it in three paragraphs of preamble, since over softening a difficult message tends to make the actual point harder to find, not easier to receive. A status update to a client or stakeholder: give the model the real facts of what happened, what is blocked, and what happens next, and ask for a format with those three sections named explicitly, since an unstructured prompt tends to produce a wall of paragraph text that busy recipients skim past rather than read.

Subject lines are a separate prompt, not an afterthought

A subject line and the email body are two different writing tasks competing for two different kinds of attention, the first for whether the email gets opened at all and the second for whether it gets acted on once it is, and asking for both in the same breath tends to produce a subject line that is really just the first sentence of the email trimmed down. A better approach treats the subject line as its own short prompt: give the model the actual goal of the email and ask for five distinct options, specific rather than vague, with an explicit length limit, since most inboxes truncate anything past roughly 50 characters on a phone screen. "Quick question about Tuesday’s deadline" tends to outperform "Following up" for the same reason a specific email body outperforms a generic one: it tells the recipient exactly what they are opening before they open it.

A worked example, before and after

Weak prompt: "write a follow up email to a client who hasn’t responded." No name, no history, no ask, so the model produces something like "I wanted to follow up on my previous email and see if you had any thoughts," a message that restates nothing and asks for nothing specific. Structured prompt: "Write a short follow up email, under 80 words, to a client named Priya who I sent a proposal to 8 days ago and haven’t heard back from. This is my second follow up. Ask directly whether she needs anything else from me to make a decision, and offer a 15 minute call this week as an alternative to email. Tone: direct and friendly, not apologetic. No phrases like ‘just circling back’ or ‘I hope this finds you well.’" The result names the actual person, states how many times this has already happened, asks a direct question instead of a vague one, and offers a specific alternative, all of which give Priya something concrete to respond to instead of another message she can leave unread.

Common mistakes

Asking for a "professional" or "friendly" email with no other detail and getting the model’s generic idea of professional or friendly instead of something matched to the actual relationship. Burying the actual ask in the third paragraph instead of leading with it, which works fine for a printed letter and poorly for an inbox someone is skimming on a phone. Requesting one subject line instead of several distinct options to compare against each other. Sending the first draft exactly as written instead of checking whether it still sounds like something you would actually say out loud to that specific person.

None of this turns an AI drafted email into something to send unread, and a quick pass to confirm it still sounds like you, not like a template, is worth the extra thirty seconds every time. What a structured prompt does is remove the specific reason generic AI emails read the way they do: no recipient, no real ask, and no constraint against the same handful of polite phrases every other generic email reaches for. Give the model the actual person, the actual goal, and the actual facts instead, and the email stops reading like it could have been sent to anyone.