ChatGPT Prompts for Cold Outreach That Don't Sound Like AI

Ask ChatGPT to "write a cold email for my product" and you'll get something grammatically perfect that nobody replies to. The model isn't the problem. The input is. With no context, it produces the average cold email, and the average cold email is exactly what your prospect deletes forty times a week.

This guide shows you how to give the model the few inputs that matter. It includes three prompts you can copy today and a short checklist to run before you hit send. It's written for solo founders and freelancers doing their own outreach, so there are no sales teams, sequencer software, or bought lists involved.

Why most AI cold emails fail

When you read the cold-email guides that rank for this topic, you see the same diagnosis again and again: generic inputs produce generic emails. One widely shared founder repo puts it bluntly: people "ask AI to 'write a cold email' with no context and send whatever comes out," so the prospect "gets the same email 40 other people sent that week" (sales-for-founders).

Three specific failure patterns show up:

1. No observed detail. The email could have been sent to anyone, so it reads as if it was.
2. Too many ideas. Features, benefits, a case study, and a calendar link, all in one message.
3. Invented proof. Models love to add "we've helped dozens of companies increase X by 40%." If that isn't true, it's a liability. And if you're new, it probably isn't true.

The fix is to change what you feed the model, not to hunt for a cleverer prompt.

The four inputs that do most of the work

Before you open ChatGPT, have these ready for each prospect:

• One observed detail. Something true and specific: a post they wrote, a job listing, a product launch, a review of their business, a tool they obviously use. Quote it exactly.
• One problem that detail suggests. Label it as a guess. You're asking, not asserting.
• Your offer in one sentence. What you do, for whom, and the result, without adjectives.
• One easy ask. "Worth a look?" or "Is this on your radar?" is easier to answer than "Can we book 30 minutes?"

If you can't find an observed detail in five minutes, that's a signal too. Either the person isn't a good fit, or your target group is too broad.

Prompt 1: Turn public info into a hook

Paste the prospect's public info (an about page, a recent post, a job listing) and let the model find the angle. It's grounded in what you pasted, so it's far less likely to invent anything.

My offer, in one sentence: [WHAT YOU DO, FOR WHOM, WHAT RESULT]
Public info about the prospect (copied from their site/profile/post):
[PASTE PUBLIC INFO]

Output:
- Fit (1–5) against my offer + 1-sentence reason
- The single most relevant OBSERVABLE detail (quote the exact phrase)
- A likely problem that detail suggests — label it [assumption]
- One opening line (max 25 words) that references the detail without flattery
- One easy question I could ask
If fit is 2 or lower, say "skip." Never claim I know them or read anything not in the pasted text.

What good output looks like: a fit score, a quote you can find on the page with Ctrl+F, and an opener that sounds as if a person noticed something. If the quoted detail isn't actually on the page, discard the whole output.

Prompt 2: The short first email (three angles)

This prompt bakes in the constraints experienced cold emailers keep recommending: short, one idea, and a low-friction question.

Write a first-touch cold email.
My offer: [ONE-SENTENCE OFFER]
Research on this person: [PASTE OUTPUT OF PROMPT 1]
Goal: [reply / short call / try a free sample]
Signature: [NAME, ONE-LINE DESCRIPTION]

Constraints:
- Under 90 words, 3–5 short lines, no bullets, no links
- Subject: 2–5 words, specific to them, no clickbait, no fake "Re:"
- Open with the observed detail — not "I hope this finds you well" or "My name is"
- One idea, one question. Interest-based ask ("worth a look?"), not a calendar demand
- Do not invent results, clients, or numbers. If proof would help and I haven't given any, leave it out.
Give 3 versions: (A) pain angle, (B) trigger angle, (C) give-first (offer something small and free).
After each, one line on the biggest risk of that version.

Why three angles: you learn faster. Send version A to ten people and version B to the next ten, and you have a real comparison instead of a hunch. Change one thing per batch.

Example of the kind of output to aim for (illustrative, not a real prospect):

Subject: launch week inbox Hi Priya, I saw your "sorry for slow replies, launch chaos" post. Congrats on the launch. Are most of those messages "where's my order?" I set up small Shopify stores with a help desk that shows the order right inside the ticket, so those take seconds. Worth a look before the next launch?

It quotes something real, asks one question, pitches in one sentence, and makes no claims it can't back up.

Prompt 3: Follow-ups that add something

A lot of replies come after a follow-up rather than the first touch. The follow-ups that work bring something new. "Just bumping this" doesn't.

Original email: [PASTE]
My offer: [ONE-SENTENCE OFFER]
One useful thing I can share: [RESOURCE / OBSERVATION / SMALL TIP]

Write follow-ups in the same thread:
- Day 3–4: share the useful thing (max 50 words)
- Day 8–10: a different angle or a simpler ask (max 50 words)
- Day 15–20: a polite close-the-loop that makes "no" easy (max 40 words)
Banned: "just bumping," "circling back," "did you see my email," guilt, fake deadlines.

The pre-send checklist

Run this on every template before it goes to more than five people:

• It mentions something specific and true about this person.
• It's under ~90 words (email) or 3 sentences (DM).
• It has exactly one question or ask.
• It contains no invented results, clients, numbers, or familiarity.
• It has no fake urgency and no misleading subject line.
• You'd be comfortable if they screenshotted it publicly.

A useful last step: paste your draft back into ChatGPT with "Act as a skeptical [their role] who gets 20 cold emails a day. Would you reply? Line by line, what would you cut?" Then cut it.

Stay on the right side of the rules

This isn't legal advice. In the US, commercial email is covered by the CAN-SPAM Act. The FTC's compliance guide lists the basics: accurate header information, no deceptive subject lines, a clear disclosure that the message is an ad, a valid physical postal address, and a clear way to opt out that you honor (FTC CAN-SPAM guide). Contacting people in the EU or UK can bring stricter rules, and every social platform has its own DM and self-promotion policies.

Practical defaults: send from your real inbox, keep volume human (dozens a day, not thousands), never buy lists, and stop the moment someone says no.

The short version

1. Find one true detail per prospect.
2. Let ChatGPT turn it into a hook, then into a short email with a single question.
3. Delete anything the model invented.
4. Follow up with something useful, twice at most.
5. Change one variable per batch and keep count.

The prompts do the drafting. Picking the right person and the right detail is still your job, and it's most of what makes an email work.