We stopped paying for Clay. Here's how we build cold email campaigns in Claude Code.
One of the emails our engine wrote opened on a single fact: the company had 12 roles open, including a senior data engineer. Then it said what that probably meant, offered a case study, and stopped at 64 words. Nobody typed it. Here's the engine that did, step by step.
- We run every outbound campaign, ours and our clients', inside
Claude Code. We stopped paying for
Clay in spring 2026. - A campaign starts as a crib sheet and a spoken brief. The engine builds the list, checks every company, picks the right people, verifies emails, researches each company and writes the copy.
- Every email has to pass 20 rules, every date and claim is checked against its source, and a person approves the copy before anything sends.
- In the demo, one brief produced 18 people, each with their own sequence. The email above came in at 64 words.
Why we left Clay
Clay sat at the heart of our work for a long time. The OneID campaigns that produced 85 sales-qualified leads in four months were built in it.
Then we took on our tenth client, and working inside
Clay got hard. To be fair to
Clay, it has improved a lot and it's still a good tool. It just wasn't built around the way we work.
So we built our own engine, and in spring 2026 we stopped paying for
Clay. Every campaign we run, for ourselves and for clients, now runs in the
Claude Code terminal.
What is a Claude Code outbound engine?
Claude Code is Anthropic's AI coding tool, and it runs in a terminal. You talk to it in plain English and it does the work: it runs the tools, reads the results and decides what to do next.
Our engine is three things working together:
- a rulebook: one file of rules every campaign follows
- a crib sheet for each client
- the data and research tools it connects to
To show how it works, I built a campaign live on camera. Here are the steps, in order.
1. Start with a crib sheet
Every client fills in an onboarding form, and it becomes their crib sheet. It asks things like:
- What do you sell? Tell us the outcomes, not the features.
- What proof do you have? Results, case studies, named clients.
- Who are your clients today? We leave them out of every campaign, and we can find lookalikes from them.
- Can you share recordings from your AI note-taker? Real discovery calls show us the pain points in your buyers' own words.
- What sales material do you have? It helps the engine match your tone, so the copy sounds like you.
Every decision about a campaign gets written into the crib sheet. Everything after this builds on it.
2. Brief it out loud
I don't type briefs. I use a voice dictation tool (
Superwhisper and
Wispr Flow both work well) and just talk. For the demo, the brief was a practice campaign for us:
- What we sell: done-for-you outbound. We build and run the whole thing: the list, the data, the copy and the sending, across several channels.
- The proof to lead with: OneID, 85 qualified leads in four months, including calls with OpenAI, Bending Spoons, MoonPay and Betfred.
- Who to target: custom cloud and software engineering consultancies in the UK, with 20 to 500 staff.
- Why them: they rely on referrals, nobody owns outbound or the pipeline number, the founder does the selling, and the focus is on delivering client work.
- Who to leave out: our current clients.
- Three emails per person, each with a different angle: offer the OneID case study; offer to build them a list of 200 to 500 leads at no cost; then ask for a 20-minute call.
- What to open with, in order: hiring for delivery roles, recent relevant news, someone new in their role or long in it, and clients named on their website.
And one instruction I give on every campaign: tie every observation back to what we sell. An observation that doesn't lead to the offer is just trivia.
3. Work out who to write to
In most companies, more than one person has a say in buying. Too many cooks in the kitchen. So we use our Multi-Persona Hypothesis to map who's involved.
We built it because so many clients, especially when they moved into a new market, didn't know which job title to target. It's based on evidence rather than guesses: competitors' case studies (who gave the quote, and their title), job adverts (what each role is measured on), and the client's own call recordings.
For this campaign, that gave us three people at each company: the founder, CEO or MD; the sales leader, such as a chief revenue officer or director of business development; and the marketing leader, such as a CMO or head of marketing.
4. Build and clean the list
We start with
AI Ark, which has the strongest LinkedIn data we've found. Then
Prospeo fills the gaps, and both have lookalike search for finding companies like your best clients. Using more than one source simply means we reach more of the market.
Every lead we find is stored in our own database (we use
Supabase), so we never pay twice for the same lead.
Then
Claude checks every company against the ideal customer profile in the crib sheet: who to target, and just as important, who to leave out. In
Clay this used to get technical. In
Claude Code it's simple, as long as the crib sheet is clear.
Then it picks the right people: usually two or three decision-makers per company, more for bigger companies. It leaves out assistants, investors and board-only roles. In our experience they do reply, but they're not the ones who decide or take the call.
It also tidies the data. If someone writes their name on LinkedIn in capitals, or a company name carries "Ltd" and a tagline, it gets cleaned up before it ever appears in an email.
5. Find and verify every email
We use an email waterfall: try one source, and if it can't find a verified address, try the next.
AI Ark first. It has
BounceBan verification built in.
Prospeo to fill most of the gaps.
BetterContact for whatever's left.
Every address gets verified, including the tricky "catch-all" ones. Bounced emails damage your sending domains, and getting into inboxes is only getting harder, especially at enterprise level. We also use
Exa to check on LinkedIn that each person still works at the company.
6. Research every company
This is where the copy gets its material:
Exa reads each company's website: what they really sell, how they position themselves, their open roles, their sector and their case studies.
Parallel fills the gaps
Exa can't reach.
PredictLeads finds signals: job adverts, recent news, funding and the tools a company uses.
Every fact gets stored next to the lead. That's why nothing we send at scale is generic.
7. Choose what to open with
For every person, the engine runs our observation waterfall. It ranks everything it found by how relevant it is, and picks the best true thing to open with. If there's nothing relevant and public, it never makes something up. It falls back to a pain point that comes with that person's role, or with their industry.
That choice also decides the channel. If someone has a strong observation and a verified email, they get email and LinkedIn. If all we have is the fallback, they get email only. LinkedIn limits how many connection requests and messages you can send each day, so we save it for the people we have the best reason to talk to. We've paid for that research, so we use it where it counts.
8. Write the copy
Every first email has the same shape: an observation, why it matters for what we're offering, the cost of doing nothing if it fits, social proof if we have it, and a soft call to action that offers something useful.
Most teams skip the second step. You get "noticed you liked so-and-so's post", then a list of features, then "can I have 20 minutes?". It doesn't connect the observation to anything.
Emails 2 and 3 each open with a new angle, never "just following up". Every email stays under 80 words, and only the last one asks for a call.
The 20 commandments of AI copywriting
Every email the engine writes has to pass all twenty. Break one and it gets flagged and fixed.
- Start with a manual template. Never invent a new structure.
- Never invent a fact. Every claim needs a source behind it.
- No spammy words: free, guarantee, act now, limited time.
- Lead with an observation: something real and publicly observable.
- Tie it to the offer. Don't mention a signal for the sake of it.
- Say "you", never "they": "your team", not "the company's team".
- Speak to the pain points of their role. What are they measured on?
- Keep it under 80 words. 60 is best, so it's easy to read on a phone.
- Short paragraphs, three sentences at most, with a blank line between them.
- Use words a 10-year-old understands.
- Ban the salesy buzzwords. They make it look like AI wrote it.
- Write like you talk.
- Don't overdo the flattery. Never "impressive" or "congrats".
- No em dashes. They're the quickest sign that AI wrote it.
- End with a soft call to action, and ask no other question in the email.
- Offer, don't ask: "Want the one-pager?" beats "Got 30 mins?".
- Keep the tone respectful. Never criticise the prospect.
- No "just following up". Every follow-up opens with a new angle.
- Two-word subject lines, in lowercase, so it looks like an internal email.
- No links, attachments or images, and never rely on open tracking.
9. Check everything, then a person approves it
Before anyone sees the copy, the engine checks it three ways:
- against the 20 commandments
- against an audit of every mistake it has made before. Whenever we find one, we feed it back as a new rule, so it never happens twice.
- by fact-checking every date and claim against a source we can click on and read. Data sources sometimes get dates wrong, and a wrong fact in an email is worse than no fact at all.
Then a person reads it. Nothing sends until we approve it, and on client campaigns the client checks the tone and brand too. On a campaign of 5,000 people, we read 20 to 30% of the messages ourselves before we're happy.
What came out the other end
While I talked through these steps on camera, the engine built the campaign in the background. When it finished, it sent me a review page: every person it found, the research behind them, and every message, each one ready to approve, edit or reject.
It found 18 people. Anyone with a strong observation and a verified email got three emails and two LinkedIn messages. Anyone without an email got LinkedIn only. The email I described at the top used the hiring fact, tied it to the offer, added the OneID proof, and offered the case study. It came in at 64 words.
What changed when we left Clay
The tools underneath didn't change much. We still use
AI Ark,
Prospeo,
Exa and the rest.
What changed is where the thinking lives. Before, our method lived in our heads and in table settings. Now it's written down in one rulebook. Every mistake becomes a new rule, and every new campaign starts with everything we've learned so far, not a blank table.
If you want to see what a campaign built like this can bring in, read how one campaign closed $26,000.
Claude Code for Cold Email: Our Full Campaign Build
Watch on YouTube: youtu.be/rAnJtbvq6xo →