How to write cold emails with ChatGPT: give it a person, not a job title

ChatGPT writes to whoever you describe. Describe a job title and you get an email few people answer. The prompt, the research note it needs and the check.

73%

Of B2B buyers who say they actively avoid suppliers that send irrelevant outreach, per Gartner
Contents
  1. Why ChatGPT's cold emails all sound the same
  2. What goes in the research note
  3. The prompt
  4. Read the draft against the note
  5. Set it up once in a project
  6. What not to paste into ChatGPT
  7. Sending is a separate job
  8. Questions
  9. Sources

ChatGPT writes a cold email to whoever you describe. Describe a job title and every sentence has to fit everyone who holds it, which is how you get the email that opens with a compliment, asks for fifteen minutes and rarely gets a reply. Give it a research note on one person instead, three facts with a source each, and ask it to mark which fact each sentence uses. Then a person checks the draft against the sources before anything is sent. The prompt is below.

Why ChatGPT's cold emails all sound the same

A model can only be as specific as what you give it. "The data platform lead at a 40-store retail group" is true of whoever holds that seat at every retailer that size, so the draft can only say what is true of all of them: a compliment, a guessed pain, a line about the product and a request for a call. That is a mail merge with better grammar, and the reader can tell, because nothing in it could only have been written to them.

Buyers act on it. In a Gartner survey of 632 B2B buyers, 73% said they actively avoid suppliers who send irrelevant outreach. The fix is upstream of the prompt: find out three things about one person, and the email has something to say.

What goes in the research note

  • The person, and why they own the problem. A title is not a reason. "Runs the data platform, and the CFO owns the outcome" is.
  • Three facts, each with a link and the date you read it. The company's press page, its careers page and its last earnings call or annual report are the usual places. A job post often names the tools a team runs and the problem it is hiring to fix.
  • Why now: the one thing that changed recently.
  • What you could not confirm. Write it down so it stays out of the email.

The note in the example below took three sources and fits in five lines. How to have a model do the reading is in Claude for sales prospecting. ChatGPT searches the web and shows its sources on every plan, free included (OpenAI), so the same research prompt works in it.

The prompt

Paste the note where it says, and one sentence about what you sell. The last line matters most.

Prompt
Write a first email from [your name] at [your company] to [their name], [their role] at [their company].

Use only the facts in this research note. Do not add any other fact, number or claim.
[paste the research note: each fact with where it was read and when]

What we do, in one sentence: [what you sell, in the reader's terms]

Rules:
- 60 to 100 words, plain text, no links.
- Open with a fact from the note, stated plainly. No compliment and no "I came across".
- Name the problem that fact suggests, in their terms, in one or two sentences.
- Say what we would do about it in one sentence.
- End with one question they can answer in a line, and an easy way to say no.
- Never write "I hope this finds you well", "quick question", "touch base" or "circle back", and use no exclamation marks.

Write two versions. Under each, list every sentence with the note fact it uses, or "none".

That last instruction is the check. A sentence marked "none" is either the ask or a claim nobody can back up, and the second kind is the line to cut. Two versions give you a choice of first sentence, which is the one that decides whether the rest is read.

Read the draft against the note

The note

James Ashworth, data platform lead

Ardent Retail Group, 40 stores

  • Announced the Halden Stores acquisition on 11 March.

    Press release 11 March

  • Posted a data engineer role that names both warehouses, Redshift and Snowflake.

    Careers page 2 May

  • Said reporting between the two businesses does not match yet.

    Q1 call Earnings call, 30 April

The draft, each sentence marked with its source

To James Ashworth, Ardent Retail Group

Subject Two warehouses and the quarterly close

  • Hi James,

  • Ardent has run two data warehouses since the Halden acquisition, Redshift and Snowflake.

    Press releaseCareers page

  • On the Q1 call, your team said reporting between the two businesses does not match yet.

    Q1 call

  • With the migration due to finish this year, the gap will only get harder to close.

    Not in the note: cut it

  • Northward reconciles the two, so finance and analytics bring one number to the quarterly close.

    Your offer

  • Is the close the problem, or is it something else?

    The ask

  • Joel

Example. Ask for the draft with each sentence’s source marked, and the check is three links and five sentences. The one line with no source is the line to cut. Northward, Ardent and James are made up, and the draft is ours, written to show the shape.

Open the three sources and read the draft one line at a time. A model fills a gap with what is likely, and a likely fact about a company you have not read about is a guess. In a cold email the reader is the one person who knows when it is wrong. Hunter's report on 31 million emails sent through its platform in 2025 points the same way: emails a person had edited drew replies at 5.2%, against 4.4% for fully automated ones.

Then send the draft to yourself with the name fields empty. The free mail-merge preview shows any {first_name} that would reach an inbox as written.

Set it up once in a project

A ChatGPT project keeps chats, files and instructions together, and its instructions apply to the chats inside it (OpenAI). Put the parts that never change there: what you sell, who you sell to, the rules from the prompt above, and two emails of yours that got replies, as files. Then each chat is one account: paste its note and ask for the draft. Projects are on every plan; Free allows five files per project, and Plus allows 25.

What not to paste into ChatGPT

A research note is about a person, so treat it as personal data. On ChatGPT's individual plans, Free, Go, Plus and Pro, OpenAI may use what you type to train its models unless you switch it off under Settings, Data controls, Improve the model for everyone (OpenAI). Even then, rating a reply with a thumb lets OpenAI use that whole conversation for training. Business and Enterprise workspaces are not used for training by default.

Whatever the plan, keep the note to what the person and their company published about the job, and leave out anything personal that is not.

Sending is a separate job

Since June 2026, ChatGPT can draft an email and, when you choose to, send it from a connected Gmail or Outlook account, on the web, on Plus, Pro, Business and Enterprise (OpenAI). Draft there if you like, and choose to send only after you have read the draft against the note. Complaints count against your domain: Google asks bulk senders to keep their reported spam rate below 0.10% and never reach 0.30%. In the US, commercial email needs a working opt-out and a postal address under the FTC's CAN-SPAM rules, whoever wrote it. OpenAI's usage policies list spam among the uses they do not allow.

The free email domain checker shows whether your domain is set up to send.

Questions

Can ChatGPT write a good cold email?

From a research note, it writes a first draft worth editing. From a job title, it writes the email every holder of that title already gets. The difference is the note, and a person still checks every fact before sending.

Should I use ChatGPT or Claude?

What you give the model matters more than which model it is. Both search the web with sources, both keep your rules in a project, and both can send from a connected Gmail account (ChatGPT on Plus, Pro, Business and Enterprise, on the web). The difference that matters most is what each does with what you paste in, on the plan you use; the Claude side is in Claude for sales prospecting.

Can I give ChatGPT a list of prospects and get an email for each?

You can, and every one will be written from whatever the list says, which is a name, a title and a company. That is the job title problem two hundred times. Write one note per person, then one draft per note.

Will Gmail mark an email as spam because ChatGPT wrote it?

No published Gmail rule says so. What Gmail and Outlook check, and what readers notice instead, is in do AI-written emails go to spam?

If you want this run for your accounts, with every email checked before it goes, see how we run it.

Sources

The outside pages this article links to, once each, and what it cites them for.

  • gartner.com73% said they actively avoid suppliers who send irrelevant outreach
  • help.openai.comChatGPT searches the web and shows its sources on every plan, free included , so the same research prompt works in it
  • hunter.ioreplies at 5.2%, against 4.4% for fully automated ones
  • help.openai.comA ChatGPT project keeps chats, files and instructions together, and its instructions apply to the chats inside it
  • help.openai.comOn ChatGPT's individual plans, Free, Go, Plus and Pro, OpenAI may use what you type to train its models unless you switch it off under Settings, Data controls, Improve the model for everyone
  • help.openai.comuse that whole conversation for training
  • help.openai.comSince June 2026, ChatGPT can draft an email and, when you choose to, send it from a connected Gmail or Outlook account, on the web, on Plus, Pro, Business and Enterprise
  • support.google.combelow 0.10% and never reach 0.30%
  • ftc.govIn the US, commercial email needs a working opt-out and a postal address under the FTC's CAN-SPAM rules, whoever wrote it
  • openai.comOpenAI's usage policies list spam among the uses they do not allow

If you want this applied to your own accounts, tell us what you are trying to fix.

Contact us