How I Booked 120 Calls in 30 Days Using AI Personalization (The Exact Prompts)

Most cold email "AI personalization" is just mail merge with extra steps. What I'm doing is different, and it's producing results I couldn't get any other way: 120 booked calls in a single month, 1,200 new opportunities for clients, and over $8 million in sourced revenue. The whole system runs on Clay and GPT, and almost nobody is talking about it yet.

Here's the framework, broken into the exact layers we use.

Layer 1: Clean Up the Basics Before Anything Else

Before any clever personalization, you need clean data. If your email opens with "Hello ACME SOLUTIONS LLC," you've already lost. We run a normalized company name prompt in Clay as a baseline. It strips out "LLC," "Inc.," all-caps formatting, anything that makes the email feel like it was pulled from a spreadsheet.

From there we add industry classification. Instead of "we help software companies," we can now say "we help MarTech companies" or "bioplastics manufacturers." That specificity matters. It signals to the prospect that you actually know who they are, not just that you scraped their domain.

Both of these are straightforward Clay prompts using the 4.0 mini model. Copy, paste, run. They're the foundation everything else sits on.

Layer 2: Identify What the Prospect Actually Sells (And Who They Sell It To)

This is where it starts getting powerful. We pull the prospect's domain and LinkedIn profile into Clay and prompt the model to identify the specific type of work they consider highest value, or want more of.

For a law firm running Google ads, for example, the prompt returns case types: railroad injury claims, wrongful death, medical malpractice. Not "personal injury." Railroad injury claims. That level of specificity is something I could not have produced manually at scale, and it changes the entire tone of the outreach. The prospect reads it and thinks: this person actually looked at my business.

The prompt guardrails matter here. We tell the model to return only one case type, the one it believes is highest value, in as few words as possible. Without those constraints, it rambles.

Layer 3: The First-Line Compliment (Used Selectively)

This one works best for coaches, consultants, and service providers who publish content or case studies publicly. We browse their website for recent wins, testimonials, or press, then generate a genuine, specific compliment under 20 words, all lowercase, no exclamation marks, written like something you'd actually say to someone's face.

The prompt starts every compliment with "awesome how you" and explicitly tells the model not to mention the company name. It sounds small. It isn't.

One campaign we ran with this approach hit 4,300 contacts and pulled 11 positive replies on an offer that wasn't even that strong. The emails looked like this: "awesome how you helped Mary Ward connect the dots of her life and enabled her to truly live, not just get by." Then we tied in a pain point. Then a question-based CTA.

That's the structure: specific compliment, relevant pain point, soft call to action. It works because it feels human, because it is human, just produced at scale.

Layer 4: Identify the Gap and Poke the Bruise

This is the most technically involved prompt in the stack, and it's producing our best results right now. The idea, which Josh Brown talks about, is to use AI to identify an operational gap in the prospect's business and call it out in the first line.

Here's how we build it. First, we run a broad website research prompt that stocks the pond: news, case studies, products, anything noteworthy about the business. That output becomes the ammunition for the next prompt.

The gap-identification prompt then browses that research and the company domain, looking for bottlenecks that might be causing revenue loss, wasted spend, or drained internal resources. We arm the model with a list of common pain points for the industry (generated by just asking GPT "what challenges would make a business like this look for a solution?"), then use that as a baseline, not a ceiling.

The output looks like: "I was looking into your business and noticed how managing the high volume of booking inquiries might be overwhelming and slowing your response times." Specific. Relevant. Not generic.

From there, we run a second prompt that connects our offer to that identified pain point in one or two sentences, written at a fifth-grade reading level, starting with "we can help" or a variation. No quotation marks in the output. No room for ambiguity about exactly how we help. After the model drafts it, we tell it to ask itself "okay, but how?" and revise if there's still any vagueness.

The two pieces together read naturally. Pain point, solution, proof. That's the email.

📥 Cold Email Swipe File

Steal the cold email templates our clients used to generate $8M+ in revenue.

Get it here →

Layer 5: Keep Sequences Short and Stack Campaigns Instead

We used to run five or six-step sequences. Conversions were almost entirely happening on emails one and two, and after that, reply rates dropped and bounce rates climbed. So we cut it down to two or three steps.

Email one: the personalized pain point plus offer. Email two: "I reached out the other day and probably didn't explain myself well enough," then a cleaner, simpler explanation of exactly how we help, written at a third-to-fifth-grade reading level, broken into separate lines so it doesn't read like a wall of text.

If those two don't convert, we don't keep hammering the same pain point. We build a new two-step campaign targeting a different pain point for the same audience. Stack campaigns, don't extend sequences.

One tool worth mentioning: Hemingway. After writing any email, I paste it in and check the reading level. If it's above grade eight or nine, I go back and simplify. I use GPT to help with that too.

Layer 6: Content Ideas as a Cold Email Hook

For clients selling content writing, SEO, or anything tied to a prospect's public-facing presence, we run a prompt that visits their website and generates three to five specific content ideas based on what we find. For a print-on-demand brand, it returned: "optimize your print-on-demand process," "understanding variable data printing benefits," "best practices for publication printing success."

The email then presents those ideas directly: "I was back on your site and figured I'd share a few content ideas based on what I saw. Before I go further, would you be open to a quick conversation, or am I reaching the wrong person?"

That last line, "am I reaching the wrong person," consistently drives replies. It's low-pressure and it respects their time.

Key Takeaways

• Clean data first: normalize company names and classify industries before any personalization layer.

• Specificity beats volume. A prompt that returns "railroad injury claims" outperforms one that returns "personal injury law" every time.

• The gap-identification prompt (pain point plus offer connection) is the highest-leverage piece in the stack. Build it carefully and arm the model with real industry pain points.

• Write at a third-to-fifth-grade reading level. Use Hemingway to check. Simplify with GPT if needed.

• Run two-to-three-step sequences, not five or six. If they don't convert, build a new campaign with a different pain point rather than extending the sequence.

• Stack campaigns by pain point. One audience, multiple two-step campaigns targeting different problems, is more effective than one long sequence.

Frequently Asked Questions

What tools does this system actually require? The core stack is Clay (for data enrichment and running prompts at scale) and GPT-4o mini as the model. Troy also mentions Hemingway as a free reading-level checker for finished emails. No other tools are required to run the framework described here.

Why does writing at a fifth-grade reading level matter for cold email? Cold emails are read fast, usually on mobile, usually while the prospect is doing something else. Complex sentences create friction. When the email is easy to read, the argument lands. Troy's rule: if Hemingway scores it above grade eight or nine, rewrite it.

How do you handle prospects who don't respond to the first pain point? Don't extend the sequence. After two or three emails targeting one pain point, build a separate campaign that leads with a different pain point for the same audience. This keeps deliverability clean and lets you test which problem actually resonates.

What makes the first-line compliment work without sounding fake? Specificity and restraint. The prompt caps the compliment at 20 words, keeps it all lowercase, bans exclamation marks, and tells the model to write it like something you'd say to someone's face. It also pulls from real, published content on their site, so it references something the prospect actually did, not a generic observation about their industry.