Next: build your own GTM agent in minutesStart free ›

Claude Cold Email: Prompts, a Reusable Skill, and Getting It Delivered

Most Claude cold email sounds like Claude. Ours opened with a dated fact about each company, went only to checked addresses, and stopped at a draft. Here is the skill file, the run and the delivery rules.

How do you write cold email with Claude?

Give Claude cold email rules once, in a skill file, then give it facts for each person: a checked email address and something true and recent about their company, with a source. Claude is good at the writing. What it can't do alone is know who works where today, whether an address will bounce, or what a company announced last month. Connect tools for those, and the emails stop sounding generic.

We ran it on 12 October 2026 with Claude, Anthropic's AI assistant, connected to Layrcake over MCP (Model Context Protocol). This is the job we gave it:

The job we gave Claude
❯ Use the cold-email skill. Find 8 heads of marketing at UK B2B software companies with 50 to 200 staff. Check every address, research each company, and write a three-step sequence that opens with a fact about their company. We sell a tool that turns recorded customer calls into case-study drafts. Budget 80 credits. Stop at a draft.
8Heads of marketing found
8 of 8Work emails found
7Safe to send
78.75Credits used ($0.79)

Why does most Claude cold email sound like AI?

Because the prompt gives it nothing specific to say. Ask for "a cold email to a head of marketing" and Claude fills the gap with the average of every cold email it has read: a compliment, a pain point it guessed, a meeting request. The reader has seen that email a hundred times.

Our fix has two parts. The rules (length, tone, one ask, no invented facts) live in a skill file, so you write them once. The facts come from tools, per person, so every opener is about something that happened at the reader's company. The skill also tells Claude what to do when no fact turns up: write a plain role-based line and say so, instead of making one up.

Step 1: Put the rules in a skill file

A skill is a folder with a SKILL.md file in it. Claude Code reads the description at the top and loads the rest when a request matches, so "write cold emails for this list" picks it up without you pasting the rules again. Save it as ~/.claude/skills/cold-email/SKILL.md for every project, or in .claude/skills/ inside one project.

The rules that mattered most in our run: under 120 words for the first email, an opener built from one dated fact with a source, one yes/no question as the ask, and follow-ups in the same thread. The whole file is at the end of this guide, and the button under Share downloads it.

Step 2: Check every address before writing

Claude found 8 heads of marketing in one search (132 matched), for 3 credits. A work email came back for all 8, at 5 credits each. Then it checked every address a second time before writing anything.

Six came back valid. One was catch-all: the company's mail server accepts any address, so no check can confirm the mailbox. One came back unknown, and that check cost nothing. It was also the one address on a different domain from the company the search listed, which is a common sign that someone has moved jobs or the company has rebranded. Claude dropped it. That left 7 addresses we would send to.

Tool calls · what Claude didabout 60 s of tool time
  1. find_leadsHead of Marketing · United Kingdom · Software Development · 50 to 200 staff · limit 88 of 132 · 3 cr
  2. enrich × 8want: work_email8 found · 40 cr
  3. verify_email × 8one call per address6 valid · 1 catch-all · 1 unknown (free) · 8.75 cr
  4. research_company × 3depth: standard · goal: a reason to email the head of marketing3 researched · 27 cr

The second check is worth its 1.25 credits. A bounced first email costs you sending reputation you need for every email after it. The lead generation guide goes deeper on finding and checking people.

Step 3: Research each company for the opener

For each company, Claude ran one research call and got back a short list of facts. Each fact came with its source link, the date it was published and the line of the source that supports it. It picked the freshest fact that a head of marketing would care about and wrote one or two sentences from it.

Research · one fact per company

3 companies, 3 dated facts, 3 openers

From our run · 12 Oct
3Companies researched
12Facts returned
9Credits each
27Credits
01Launched a new AI agent for procurement teamsprocurement software · press release · March 2026Product
02Raised a large Series B; processes 220M invoices a yearaccounts-payable software · trade press · 2025–26Funding
03Acquired an automation company; serves 50+ housing providerssocial-housing software · January 2026Expansion
Company names and links are in the real answerEvery fact carries its source line
Opening line · the pattern (synthetic, company details removed)
❯ Saw the launch of [product] in March. [Category] is a crowded story to tell right now, and yours has real customers behind it.

Check where each fact came from before you use it. Press releases and trade press are the safest openers. One of our third company's facts came from a company-profile directory page, so the skill sends a fact like that back for a second source before it goes in an email.

Step 4: Write the email once, with the opener as a variable

Claude wrote one template and passed each person's opener in as a custom field. The template stays the same for everyone; only the first line changes. That keeps the email short and makes the opener easy to read and fix before anything goes out.

Email 1 · as drafted (the product is a placeholder for the test)
❯ Subject: Customer proof at {{company_name}}

Hi {{first_name}},

{{opener}}

The hardest part of that story is usually proof. Customers say great things on calls, and almost none of it makes it into a case study.

We built a tool that pulls the best customer moments out of recorded calls and turns them into case-study drafts your team can edit in an afternoon.

Worth a look at what it found in one week of calls?

Sam
Steps 2 and 3
❯ Step 2 · day 3, same thread
Hi {{first_name}},

One example: a team your size turned 12 sales calls into three case studies in a week, without a single customer interview.

Happy to show you how, if it is useful.

Sam

Step 3 · day 7, same thread
Hi {{first_name}},

I will leave it here. If customer proof moves up the list next quarter, reply and I will send the examples.

Sam

The first email is 88 words. It names one problem the reader owns, one result and one yes/no question. There is no meeting link. Our view: asking for 30 minutes in the first email asks for more than you have earned.

What decides whether a cold email gets delivered?

Mostly your domain's setup and your complaint rate, under rules Gmail, Yahoo and Outlook each publish. All three expect authenticated mail and a low spam complaint rate. The strictest rules apply to bulk senders (5,000 or more messages a day to that provider), but the basics apply to everyone.

RuleGmailYahooOutlook.com
Authentication, every senderSPF or DKIMSPF or DKIMSPF, DKIM and DMARC at 5,000+ a day
Bulk sendersSPF and DKIM, plus DMARCSPF and DKIM, plus DMARC (at least p=none)Non-compliant bulk mail sent to Junk since 5 May 2025
Spam complaint rateBelow 0.3%Below 0.3%Not stated as a number
UnsubscribeOne-click for bulk marketing mailOne-click for bulk, honoured within 2 daysRecommended

Sources: Google's Email sender guidelines, Yahoo Sender Hub best practices and Microsoft's Outlook post on high-volume senders, checked 12 October 2026.

For cold email, that adds up to a few habits. Send from a separate domain with SPF, DKIM and DMARC set up, so your main domain's reputation is never at risk. Warm new mailboxes before you send from them. Keep each mailbox's volume low. Check every address, because bounces hurt the sender's standing too. And add a one-click unsubscribe even if you are far below 5,000 a day. It costs nothing and keeps complaints down.

The sources: Google's sender guidelines, Yahoo's best practices and Microsoft's Outlook requirements.

Step 5: Draft the campaign and approve it yourself

The last call is create_campaign: the template, the steps and each person with their opener. It creates a draft and never sends. It also reports anyone it left out, for example people who opted out or were emailed by another campaign recently.

Sending is a separate call, launch_campaign, which first answers with a preview, a fresh check of the recipients and an approval link. A signed-in person approves before a single email leaves. If you need sending mailboxes, setup_sending quotes warmed domains and mailboxes the same way: nothing is bought until you approve.

Nothing sends on its own

Claude stops at the draft. You read the openers, decide on the catch-all address, and approve the launch yourself.

What did it cost?

The run cost 78.75 credits, about $0.79 at $0.01 per credit, at the prices on 12 October 2026. Research was the biggest line, and it is the line that makes the emails worth reading.

StepAskedResultCredits
Find people883
Find work emails8840
Check addresses87 kept8.75
Research companies3327
Draft campaign0

Lookups that found nothing and checks that came back unknown were free. Every call reported its charge.

The full skill file

Copy it, or use the download button. Change the length limits and sign-off to suit you; keep the rule about sources.

~/.claude/skills/cold-email/SKILL.md
---
name: cold-email
description: Write short, sourced cold emails and a three-step sequence for a list of people. Checks every address first, opens each email with a dated fact about the person's company, and stops at a draft for a human to approve. Use when asked to write cold emails, outreach or a sequence for specific people or companies.
---

# Cold email

You write cold emails that a busy person would read to the end. You never send email and never launch a campaign. You stop at a draft and show the user what you wrote.

## Before you start

Ask once for anything missing:
1. **The people:** a list (name, company website, work email if known), or a description of who to find.
2. **The offer:** what we sell, in one or two sentences, and one concrete result a customer got.
3. **The sender:** first name, and how they sign off.
4. **The budget** in credits, if the tools charge. Default to 100.

## Rules for every email

- **Under 120 words** for the first email. Under 60 for each follow-up.
- **Open with a fact about their company,** taken from a source you can link, with a date. A launch, a hire, a funding round, a customer, an expansion. No fact, no opener: use a plain role-based line instead and say so.
- **Never invent.** Every claim about the reader or their company traces to a source you were given or found. If a source is a data-broker profile or older than 12 months, don't use it for the opener.
- **One problem, one offer, one ask.** The problem is one the reader's role owns. The ask is a yes/no question, never a meeting link in the first email.
- **Write like a person:** short sentences, no exclamation marks, no "I hope this finds you well", no "quick question", no buzzwords. Reading level of a good newspaper.
- **Subject line:** three to six words, lower stakes than the email. No clickbait, no "Re:".
- **Follow-ups** stay in the same thread (no new subject). Step 2 adds one new piece of proof. Step 3 closes the loop politely.

## The job

1. **Check every address before writing a word.** Verify each email. Drop anything invalid. Keep catch-all addresses only if the user agrees, and say how many there are. A result you can't confirm is a drop, not a guess.
2. **Research each company.** One research call per company. Pick the single freshest fact that the reader's role would care about. Write it down with its source link and date.
3. **Write the opener per person** from that fact, in one or two sentences. Pass it to the template as a custom field (`opener`), so the template stays the same for everyone.
4. **Write the sequence once,** with `{{first_name}}`, `{{company_name}}` and `{{opener}}` variables: three steps over about a week (day 0, day 3, day 7).
5. **Create the campaign as a draft.** Never launch it. Show the user the draft, the openers with their sources, and what each step cost.

## What you hand back

A short table: person's role, company, the fact, its source and date, the opener. Then the three emails as they will read for the first person. Then the total credits used. Then one line: "Nothing has been sent. Approve the launch when you're ready."

To connect the tools it calls, see the Claude setup page. For a skill that runs the whole prospecting job, see the Claude Code sales agent guide.

Questions people ask

Can Claude write cold emails?

Yes, and it writes them well when it has facts to work with. Give it a checked address and a dated fact about each company, plus rules for length and tone, and the emails read as written for one person.

Is there a Claude skill for cold email?

You can write one: a SKILL.md file with your rules for length, tone, sourcing and follow-ups. Claude Code loads it whenever you ask for cold emails. Ours is free to download from this guide.

Will Claude send the emails for me?

Not in our setup. Claude creates a draft campaign. Launching it returns a preview and an approval link, and nothing sends until a signed-in person approves.

What are the Gmail and Yahoo rules for cold email?

Both expect SPF or DKIM on every message and a spam complaint rate below 0.3%. Bulk senders, at 5,000 or more messages a day, also need SPF and DKIM together, DMARC and one-click unsubscribe.

Your next move

Give your AI agent the key to go to market.

The right companies, people, details and signals, all available through one API key. Ready for your AI agent the moment you paste it. Nothing sends until you say so. Free to start, then pay as you go.

mcp url · https://mcp.layrcake.dev/mcp 300 free credits · No card required