What Are the Best AI Prompts for Cold Outreach?

Most cold outreach prompts work in the screenshot and fail on the hundredth lead. The line runs too long, a number appears from nowhere, or the email reads like everyone else's.

You only see it when you run the prompt on real leads and check every output.

This guide gives you eight AI prompts for cold outreach, from the first line to the objection reply, and shows how we tested each one on ten real sales leaders.

You can run the prompts from an AI Agent in Claude Code, Codex or Cursor, or give your team an n8n form that runs one prompt on one lead.

We ran both on real leads, and nothing was sent. Copy the prompts below, or use the kit to test your own.

We will walk you through the process step by step.

Everything is in one public GitHub repo: the skill, the eight prompts, your offer file, a small runner and the n8n form. Paste the prompt into your AI Agent, or pick the n8n tab for the form.

Agent kit Free · Public

GitHub n8n

Let your AI agent set it up

Paste this prompt into Claude Code, Codex or Cursor. Your agent clones the public repo, follows its instructions and asks before every change.

Prompt · GitHub Copy

Clone https://github.com/evaboot-labs/cold-outreach-prompts, read AGENTS.md
and follow the skill. Test the cold outreach prompts on my Evaboot export.

In the repo

• The cold-outreach-prompts skill: leads, run, check, fix one thing at a time
• prompts.py: runs Claude on each lead, checks every output, scores each prompt
• prompts/ and OFFER.md: the eight prompts with their limits, and your offer
• The n8n form, with its setup guide

Open the repo on GitHub

Set it up in n8n

Paste this prompt into your AI agent and it walks you through the import, one step at a time. Or open the workflow JSON on GitHub and import it yourself.

Prompt · n8n Copy

Set up the cold outreach prompts form in my n8n.
Clone https://github.com/evaboot-labs/cold-outreach-prompts and follow docs/n8n.md.
Walk me through building the workflow with my offer, importing it,
the Anthropic credential and publishing the form, one step at a time.

In the repo

• The workflow JSON: one prompt and one lead in, the message and its check out
• docs/n8n.md: the offer, the credential and the form, step by step
• The same limits and checks as the script

Open the n8n JSON on GitHub Open the repo on GitHub

We used Evaboot and Claude in our example, but with a few tweaks, the same prompts work with ChatGPT, Gemini and any lead source.

What Makes a Good AI Prompt for Cold Outreach?

A good cold outreach prompt does one job with real data and hard limits. Most prompts online skip at least one of the four parts.

Here is what each part does:

The last rule makes everything checkable. When the model lists the facts it used, a script can look for each one in the lead's data.

8 AI Prompts for Cold Outreach

Each prompt below ran on the same ten UK sales leaders from an Evaboot export of a Sales Navigator search. Under each one: its score and one real output, with the first name removed.

Every prompt also gets the same rules and the offer, added by the runner:

Hard limits: at most {max_words} words, at most {max_questions} question mark(s).
Count before you answer.

The offer: {your OFFER.md}

Use ONLY the facts in the lead JSON and the offer. Never invent a number,
a news item, a customer, a tool they use or anything you cannot see in the
data. Never guess anyone's gender. Return {"text": ..., "facts_used": [...]}

1. First Line From One Fact

The opening line decides if the rest gets read. It builds on one fact and says why you wrote, without a pitch. Limit: 25 words. Score: 9 of 10.

Write the opening line of a cold email to this person. Build it on ONE fact
from their data (their role, months in the job, company growth, company
description or headline). Say why that fact made you write. Do not pitch the
product in this line. Do not flatter.

2. Short Problem-First Email

A fact, the problem, one proof and one question, in that order. Limit: 90 words and one question. Score: 10 of 10, 70 words on average.

Write a cold email body (no subject, no signature) to this person. Structure:
one line on a fact from their data, one or two lines on the problem the offer
solves for someone in their role, one line of proof from the offer, then one
question as the call to action. Plain words, no buzzwords.

3. Three Subject Lines

Short and lowercase, like a note from a colleague. At least one names their company or role. Score: 10 of 10.

Write three subject lines for a cold email to this person, one per line, each
under six words, in lowercase like a note from a colleague. At least one
refers to their company or role. No clickbait, no emoji, no "re:" or "fwd:".

For more on this step, see our guide to B2B cold email subject lines.

4. Follow-Up With a New Angle

Never "just following up". Bring one new angle: a different problem, a different proof or a different fact. Limit: 60 words. Score: 10 of 10.

Write the second email of a sequence to this person, sent four days after a
first email with no reply. Bring ONE new angle (a different problem, a
different proof, or a different fact from their data). Never say you are
following up or bumping the thread. End with one easy question.

5. Last Email That Gives an Easy Out

Short, polite, and easy to answer with a no. Limit: 45 words. Score: 10 of 10.

Write the last email of a sequence to this person, after two emails with no
reply. Be short and polite, give them an easy way to say no or to point you
to someone else, and do not guilt them. Mention their company once.

6. LinkedIn Connection Note

No pitch, no link, no meeting request: under 280 characters, one fact and a reason to connect. Score: 10 of 10.

Write a LinkedIn connection request note to this person, under 280
characters. Build it on one fact from their data and say why you want to
connect. No pitch, no link, no meeting request.

7. Reply to "We Already Have a Provider"

Agree first, then give one reason to compare. Never criticise their provider. Limit: 70 words. Score: 10 of 10.

This person replied to your cold email: "Thanks, but we already have a data
provider." Write your reply. Agree first, then offer one specific reason to
compare (from the offer, not invented), and ask one low-effort question.
Never argue, never criticise their provider.

8. Ask for the Right Person

When the lead may not own the problem, ask who does. One fact shows you looked. Limit: 50 words. Score: 9 of 10.

Write a short email asking this person who on their team owns outbound list
building, in case it is not them. Use one fact from their data to show you
looked. One question only.

How to Test Your Prompts on Real Leads

A prompt that reads well on one example can still fail on half your list. Test it on real leads before any of it reaches a prospect, in four steps:

1. Export Ten Real Leads

prompts.py facts keeps the people with a safe email from an Evaboot export of your Sales Navigator search. Ten is enough to see a pattern.

Use the people you will really write to. We used UK sales leaders at software companies of 51 to 500 employees, the buyers of our example offer.

2. Run Every Prompt on Every Lead

prompts.py run calls Claude once per lead and per prompt, and saves the text with the facts it used. Eight prompts on ten leads made 80 calls.

Our first run of the first-line prompt failed on every lead. The prompt file did not state the 25-word limit, so every line ran 30 to 37 words:

The fix was one line in the runner: "Hard limits: at most 25 words. Count before you answer." The next run passed 9 of 10.

3. Check Every Output Against the Data

prompts.py check rejects five kinds of failure: a fact or a number that is not in the data or the offer, a line over the limit, an extra question, a spam word, and a guess at someone's gender.

prompts.py score sums it up per prompt. On our ten leads, 78 of 80 outputs passed:

Read the failures, not only the score. One referral email asked two questions instead of one:

The check also catches your own mistakes. Our example offer said "free trial", and "free" is on the spam-word list, so three prompts failed until we changed it to "trial account".

4. Fix One Thing at a Time

Change one line of one prompt, then run it again on the same ten leads. If you change three things at once, you will not know which one worked.

Here is what a passing short email looks like, straight from the output file:

Read a sample yourself before you send anything. A check catches invented facts and long lines, not a tone that does not fit your brand.

How to Give Your Team the Prompts as an n8n Form

The same prompts run as an n8n form: pick a prompt, paste what you know about one lead, and get the message with its check.

The form asks for the prompt and nine lead fields, three of them required:

We ran the LinkedIn note on one real lead. The last page shows the note, the fact it used and the check:

The form sends nothing. Your team copies what passes into their own email or LinkedIn tool.

Which AI Is Best for Cold Outreach?

Any strong model writes a decent cold email when you give it the right data and limits. ChatGPT, Claude and Gemini all follow the prompts above.

The data and the checks matter more than the model. A model with no facts invents them; a model with good facts and no limits writes too long.

If you want the whole flow automated, from the list to the send, see how to automate cold emails with AI.

Common Mistakes With AI Prompts for Cold Outreach

These make a prompt fail on real leads:

Test before you scale. Ten leads and a check take minutes; a thousand emails with an invented number take a domain.

For the rest of the email, from timing to volume, read our cold email best practices.

Frequently asked questions