A ChatGPT Prompt for Cold Emails That Starts With Research

Most pages that rank for "ChatGPT prompt for cold emails" hand you fifteen or twenty templates with square brackets to fill in. You fill them in, and the email that comes back is polite, well formed, and the same in spirit as the four other pitches your prospect received that morning. The prompt wasn't the weak part. What you fed it was.

We write cold outreach with AI in three passes, and only the middle one is the "write the email" prompt everybody shares. The first pass decides what the email is about. The third reads the draft the way the recipient will. All three prompts are below, with a worked example and a short one for the follow-up.

Why template prompts produce the email everyone ignores

A model fills gaps with the most probable text. If your prompt says "write a personalized cold email to a Head of Support," all it knows is a job title and a genre, so it writes the genre: a compliment, a claim about saving time, a request for fifteen minutes. We covered the mechanics in why AI answers come out generic. Cold email is where that default costs the most, because the reader is looking for a reason to stop reading.

What keeps a stranger reading is one true, specific observation about their situation that connects to something you can fix. The model can't supply that observation. It can help you choose it and phrase it.

Step 1: Have the model pick the angle before it writes

Spend five minutes collecting raw material on one prospect: their role, a recent post, the company's open job ads, a product page, a line from a press release. Paste it in unedited. Then ask for angles and hold the draft back.

I'm going to write a cold email to one person. Before any drafting,
help me pick the angle.

What I sell (one sentence): [your offer]
Who it's for and the problem it removes: [buyer and problem]
Proof I can mention truthfully: [a result, a client type, a number]

What I know about this prospect (raw notes: role, company, recent
posts, job ads, news, product pages):
[paste notes]

Do three things:
1. List the 3 observations from my notes that most plausibly connect
   to the problem I solve. Use only what is in the notes. If something
   is your inference, label it "inference".
2. For each, write one sentence on why this person might care about
   it this month.
3. Tell me which one you would lead with, and what is missing from my
   notes that would make the email stronger.

Do not draft the email yet.

Two instructions in that prompt do most of the work. "Use only what is in the notes" stops the model from inventing a funding round or a product launch, which it will otherwise do with complete confidence. "Do not draft the email yet" keeps the conversation on the decision that matters. If the model tells you the notes are thin, believe it and go find one more fact.

Step 2: The drafting prompt

Stay in the same chat so the notes and the chosen angle carry over. The structure follows the four elements from our guide to writing prompts for work; here the criteria block carries the weight.

Now write the email using observation number [1, 2 or 3].

Rules:
- 60 to 90 words in the body. Plain text. No bullet points.
- Line 1: the observation, stated as a fact about them. No compliment,
  no "I hope you're well", no "I came across your profile".
- Next: the problem that usually follows from that observation, in the
  words a [their job title] would use.
- One sentence on what we do, with the proof from my notes. Do not add
  any claim, number, or client name I did not give you.
- End with one question they can answer in a single line. No calendar
  link, no "15 minutes".
- Subject line: 2 to 4 words, lowercase.
- Write it the way I would say it out loud to a peer.

Give me 2 versions that differ in the opening line. Then tell me which
one you would send and why.

A worked example

Take a typical task for a salesperson at a small firm that writes help-center content for software companies. The prospect is the Head of Support at a 150-person scheduling software company. The notes hold three things: the company has four open roles for support agents, the public help center still shows an old interface in its screenshots, and the prospect recently posted about response-time targets.

The angle prompt ranks the job ads first. Hiring four agents at once suggests ticket volume is outrunning the team, and that is this person's problem this quarter. It labels the outdated help center as an inference worth checking. It also points out what's missing: we gave it no proof.

We add one line of proof we could stand behind (for this exercise, a past project in which rewriting the 30 most-visited help articles cut "how do I" tickets) and run the drafting prompt. A typical first version looks like this:

Subject: four support hires

Hi Dana, you have four support agent roles open at once, which usually means tickets are growing faster than the team can absorb. A good share of that volume tends to be "how do I" questions a help center could answer, and yours still shows the old interface in its screenshots. We rewrite the 30 most-visited articles for software teams so those tickets stop arriving. Is ticket deflection something you're working on this quarter, or is hiring the whole plan?

It is close, and two things need a human. "A good share of that volume tends to be" is the model hedging a claim it can't back, so we cut it or replace it with what we have seen in our own projects. And the screenshot remark is a small criticism of the prospect's work in the second sentence of a first email. Whether that lands as useful or rude is a judgment call. The third prompt helps you make it.

Step 3: Read it as the recipient

Now switch sides. You are the [job title] who received this email
between two meetings, on a phone. You get several vendor pitches a day.

- At which exact word would you stop reading, and why?
- What in it sounds like a template?
- Is anything in it presumptuous or wrong about my situation?
- What would make you reply with one line?

Then rewrite the email once, fixing only those points. Keep it under
90 words.

This pass tends to catch an opening that takes too long to reach the point, a phrase that assumes too much about the prospect, and a closing question that is really two questions. Take the critique and treat the rewrite with suspicion. The rewrite often sands off the one specific line that made the email worth sending. If you need a narrower fix than a full rewrite, a targeted follow-up prompt does it faster.

The follow-up, without "just bumping this"

A follow-up that only says "bumping this to the top of your inbox" gives the reader no new reason to answer. Ask for one that adds something.

Write a follow-up to the email above, to be sent 4 working days later
if there is no reply.

- 40 to 60 words. Same thread, no new subject line.
- Do not mention that they haven't replied. No "just following up",
  "bumping this", or "circling back".
- Add one new, useful thing the first email did not use:
  [paste a fact, an example, or a short resource]
- End with an easier question than the first email, or with "if this
  isn't yours to decide, who should I ask?"

We stop after two follow-ups. A third rarely says anything the first two didn't.

What to paste in before you prompt

Input	Where to find it	What it becomes
A trigger: hiring, a launch, a new role, a pricing change	Careers page, company blog, the prospect's own posts	The first line
Their words for the problem	Their posts, interviews, job descriptions	The second sentence, in their vocabulary
One piece of proof	Your past work, only what you can document	The single sentence about you
Your offer in one sentence	You, written before you open the chat	A limit that keeps the model from describing your whole company
A sample of how you write	Two emails you sent and liked	The tone, pasted with "match this voice"
The reply you want	You	The closing question

If you can't fill the first row for a prospect, the answer is more research or a different prospect. No prompt changes that. When the reply is a yes, the next document is the proposal: How to Write a Business Proposal With ChatGPT, Step by Step.

Where this breaks

• Invented personalization. Models state guesses as facts. Check every claim about the prospect against your notes before you send. A wrong detail in line one is worse than no personalization at all.
• Volume. These prompts are built for one email to one person. Run through a mail-merge tool across 2,000 contacts, the "observation" turns back into a template. Mailbox providers also set rules for senders. As of October 2026, Google's email sender guidelines require SPF or DKIM authentication from everyone who sends to Gmail accounts, tell senders to keep reported spam rates below 0.3%, and add DMARC and one-click unsubscribe requirements for anyone sending more than 5,000 messages a day. Good copy does not rescue a domain that mailbox providers distrust.
• Law and consent. Rules on unsolicited commercial email differ by jurisdiction and by whether you write to a company address or a private person. Check what applies where your recipients are before you build a list.
• Personal data. Don't paste a contact export, private messages, or CRM notes marked confidential into a consumer chat account. Public, professional information about one prospect is enough for these prompts.
• The AI accent. Even a good draft carries stock phrases. We keep a list of words that sound like AI and strip them before sending.

Where we go deeper

Cold emails, follow-ups, rejections, and difficult conversations form one block of the second session of our two-week practicum. The assignment after that session is to build a template for your most frequent correspondence. If outreach is your daily work, that template can be this three-prompt sequence rebuilt around your offer, your proof, and a sample of your own writing. We describe what the practicum covers for sales work on the page for sales professionals.