AI cold email: 6 prompts for under 80 words, and the 0.3% spam line Gmail holds every sender to

In brief

An AI cold email works when you give the model the research and cap it under 80 words: one fact about the prospect, one problem, one ask. Instantly's 2026 benchmark puts the average reply rate at 3.43%. Gmail tells every sender to stay below a 0.3% spam rate, so volume without research risks rejection.

Three black plinths in a graphite studio: a brass magnifying glass on a stack of index cards, a cream envelope beside a chrome desk bell above the OpenAI, Claude and Gemini marks, and a glass jar of folded paper notes.

What should you give AI before it writes a cold email?

Give it the research, the problem and the ask, and nothing it has to invent. A model working from a job title alone writes the "I came across your profile" opener that every buyer has read a hundred times. Paste this into ChatGPT, Claude or Gemini:

```

Write a cold email of under 80 words to [name], [role] at [company].

What I found about them: [one fact you checked, with the link: a post, a job ad, a launch, a quote]

The problem it suggests: [one sentence]

What I offer: [one sentence, with one number if you have it]

The ask: [one small yes: a 15 minute call, a reply, a yes or no]

Plain text. No "I hope this finds you well", no "I came across", no flattery, no exclamation marks. Use only the facts above. If you need a detail I didn't give, write [NEEDED: detail] instead of inventing it.

```

The 80 comes from data, not taste. Instantly's 2026 benchmark, built on billions of cold email interactions from 1 January to 18 December 2025, found that "Elite performers average fewer than 80 words per first-touch email". It also found that 58% of all replies come from step one, so the first email carries most of the result.

Per 1,000 cold emails: 34 replies on average, and Gmail's ceiling is 3 spam reports
Average senderreplies, 3.43%34 repliesTop quarterreplies, 5.5%55 repliesTop tenthreplies, 10.7% or more107+ repliesGmail ceilingspam reports, never reach 0.3%3 reportsGmail targetspam reports, stay under 0.1%1 reporteach square is one email in 1,000; replies from Instantly's2026 benchmark, spam rates from Gmail's sender guidelines

Read the chart from the bottom. Gmail's sender guidelines say "Keep spam rates reported in Postmaster Tools below 0.10% and avoid ever reaching a spam rate of 0.30% or higher." At the average reply rate, that ceiling is about one spam report for every 11 replies (our division: 34.3 replies per 1,000 over 3 reports). AI makes 1,000 emails cheap to write. It doesn't move the 3.

Which ChatGPT prompts handle the rest of a cold email?

Five more prompts cover the rest: the research, the subject line, the follow up, a whole list and a critique. Research comes first, because the fact in line one is the email. With web search on, send:

```

Using only [company]'s own site, its press releases and [name]'s public posts from the last 90 days, list 5 facts that suggest they have [the problem you solve]. Give the link and the date for each. If you find nothing, say so. Don't stretch a fact to fit.

```

Open every link. A model with web search can misdate a post or merge two people with the same name, and a wrong fact in line one ends the email. The subject line comes next:

```

Write 5 short subject lines for the email below. No clickbait, no "Re:" or "Fwd:" on a first email, nothing in capitals. Each one must describe what the email is actually about. [paste the email]

```

The last rule is law, not style. The FTC's CAN-SPAM guide says "The subject line must accurately reflect the content of the message." A fake "Re:" breaks it. For the follow up:

```

Write a follow up to the email below, to send [number] days later, in under 50 words. Add one new thing: [a second fact, a short result, a useful link]. Don't write "just following up" or "bumping this". End with the same ask, or an easy way to say no.

```

For a list, keep the research in a column so the model can't make it up:

```

Here is a CSV of prospects with the columns name, company, role, fact, link. For each row, write one opening line of under 25 words that uses only the fact in that row. If the fact column is empty, write [NEEDED] and nothing else.

```

The empty row rule matters most. A row with no fact gets no email. Last, have the draft read by someone who'd delete it:

```

You're a busy [role] who gets [number] cold emails a week. Read this one: [paste]

Quote the sentence where you'd stop reading, the phrase that sounds like a template, and the reason you'd ignore it. Then rate from 1 to 5 how likely you'd be to reply. Don't rewrite it.

```

Run the critic in a fresh chat, or on a second model if you have one. The model that wrote the draft has spent the whole chat agreeing with you. A new one hasn't. More outreach prompts, including a full sequence, are in the sales prompt pack.

Do AI cold emails work?

Yes, when the research behind them is real. Instantly puts the average reply rate at 3.43%, the top quarter of senders at 5.5% and the top tenth at 10.7% or more. Every one of those senders can reach the same models, so the list, the fact in line one and the size of the ask make the gap.

The page-one advice agrees on the same three moves: specific context in the prompt, a short email, and an edit before sending. Hunter's guide adds the catch that a model doesn't know your business, so you supply it. Nobody's data rewards a long email or a generic one.

Should you use an AI cold email tool or a chat model?

Use a sending tool to send and a chat model to write. Tools such as Instantly and Lemlist run the sequence, schedule the follow ups and track replies. A chat model writes from your research and sends nothing.

Whizi, which publishes this page, is a chat app and loses that comparison outright: it doesn't send email, warm a domain or track a reply. What it gives you is GPT, Claude, Gemini and 280+ other models in one chat. A project keeps your offer and your three best emails as instructions on every chat, and on Pro, side by side model chats put two models' drafts of the same email next to each other.

Common questions

Frequently asked questions

Can ChatGPT write cold emails?

Yes, and it writes good ones when you paste the research: one checked fact about the prospect, the problem it suggests, your offer and one small ask, under 80 words. Without a fact it writes a generic opener. Edit before you send, and see AI for email writing for the everyday replies that follow.

How long should a cold email be?

Under 80 words for the first email. Instantly's 2026 benchmark found that its elite senders average fewer than 80 words per first touch, and that 58% of replies come from step one. Keep follow ups shorter still, with one new fact each. The sales prompt pack has a short sequence ready to copy.

Is it legal to send cold emails written by AI?

In the US, yes, under the same CAN-SPAM rules as any commercial email: an accurate subject line, a valid postal address, a working opt-out honored within 10 business days. The FTC says there's no exception for business email. Other countries are stricter, and prospect lists hold personal data (what never to paste into AI).

Will AI cold emails go to spam?

Not for being AI written. Mail lands in spam when recipients report it or authentication fails. Gmail tells every sender to keep spam reports below 0.3%, and Gmail, Yahoo and Outlook require SPF, DKIM and DMARC from bulk senders. Specific, researched emails earn fewer reports (prompt engineering for beginners covers writing specific prompts).