How I Use Claude to Create Winning Agency Ads (200+ Clients)

Cover image

Every ad I run is at least somewhat written by AI.

Not most of them.

I've generated over $2 million for my own agency, Paid House, using this exact method, and signed over 200 clients along the way.

Last year alone I made over $1 million from paid ads specifically, not counting referrals, and spent around $250,000 on ads to do it.

This is the exact system I use to turn Claude into an ad-writing machine.

I'm not going to pretend I'm an AI expert.

The truth is this is simple, and the results come from feeding the tool the right context, not from being clever with it.

Let's get into it.

Step 1: Build the Project and Feed It Context

The whole system lives in a Claude project.

A project holds knowledge files, which is how you give Claude the context it needs to write ads that sound like you and not like a robot.

So the first job is loading that context in.

Give Claude the following about your offer.

The exact offer you're running.

The exact price and payment structure.

The deliverables.

The outcome the client gets.

The timeline.

The qualification criteria, meaning who this offer is for and who it's not for.

Your best case studies.

Some successful old ads, if you have them.

If you don't have winning ads of your own yet, upload competitor ads instead.

You can use a tool like Super Scalar to download Facebook ads and transcribe them, then drop those into the project.

One more thing to include, and it matters: the reasons clients originally bought from you.

Step 2: The System Prompt That Stops Hallucination

This is the most important part, and skipping it is why most people's AI ads are useless.

Left alone, AI invents things.

It pulls out guarantees you never offered and proof that never happened, and drops them straight into your ad.

So the prompt has to lock that down.

This is the exact prompt I use:

"You are helping me research and write paid social ads for my agency. Do not invent facts, results, objections, customer language, deliverables, guarantees, or proof. Only use information contained in this project."

Then you paste in all the offer info from Step 1.

The prompt continues:

"Your role is to analyse customer language, organise advertising angles, and help me create variations. Your role is not to replace strategic judgment. Whenever information is missing, flag it rather than filling the gap with generic assumptions."

That last line is the safeguard.

Instead of guessing, Claude tells you what's missing, and you stay in control of what actually goes in the ad.

Step 3: Mine Your Sales Calls

Before you write a single ad, you mine your sales calls, because that's the highest-proximity access you have to your market.

I'd do this in a separate project from the ad writing, so keep one project for the calls and one for the ads.

Go to Fireflies, Fathom, or whatever you use to record calls.

If you're not recording your sales calls, you're leaving a lot of money on the table, so start.

Download your 20 to 30 most recent calls and throw them all into the project.

You can get a VA to do this or do it yourself.

Then you ask Claude to analyse them and build 6 sections.

The 10 most frequent problems prospects had before working with you.

The 10 circumstances they were in when they started looking for help.

The 10 outcomes they wanted most.

The exact phrases and vocabulary they repeatedly use.

The objections and fears stopping them from buying.

The events that create urgency and push them to look for a solution.

For every item, tell Claude to include supporting evidence from the material and not generalise beyond it.

Now you've got the raw material to write ads that use your market's actual words.

P.S. - If you want us to help you set this up for yourself + help launch paid ads so that you can hit $100k/mo in PROFIT...

- Book a call with us here -

(Scaled my agency Paid House $200k+/month & did $50M in adspend)

Step 4: Turn the Research Into Angles

I always write ads as angles.

Each problem, each circumstance, and each outcome is its own angle, and the ad gets built around it.

To see the difference, this is how they split in my niche.

Problems are the hard facts of their situation. The agency is stuck at $20k to $50k a month. The founder is doing all the selling. Referrals have plateaued.

Circumstances are what those problems feel like. They're working 10 hours a day, they haven't seen their family in weeks, their calendar is empty.

Outcomes are the inverse of the problems. 10 to 15 new clients a month, a predictable pipeline, the ability to spend more on ads to book more calls.

Then you feed the sales-call research into your ad project and prompt Claude to build angles from it.

This is the prompt:

"Based only on the research in this project, create 21 different advertising angles. 7 specific problem, 7 specific circumstance, and 7 specific outcome angles. For each angle, include the core claim, the customer insight behind it, the exact audience segment it will resonate with, the offer feature or proof that supports it, the primary objection we must answer, and why this angle is materially different from others."

There's a powerful extra step here.

Ask Claude to score every angle from 1 to 5 based on how often it comes up on your sales calls.

9 times out of 10, your best ads come from the angles that show up most on calls, because your calls are the closest you get to your market.

Step 5: Generate the Hooks

Angles and hooks are different things, and people confuse them constantly.

The angle is the macro direction. What pain, outcome, or circumstance is the ad built around.

The hook is the specific opening line, the entire ad idea condensed into 1 to 3 seconds.

The angle encapsulates the hook and the body, but you still have to write the hook itself.

When you launch, you're aiming for a hook rate above 20%, which is one of the key metrics to watch.

I use a specific framework for hooks called ABC.

A is audience. Call out exactly who you're talking to.

B is benefit. Name a benefit they genuinely want. This can also be a problem or a circumstance.

C is curiosity. Say something that makes them need to know more.

This is an ABC hook from my world.

"Do you want to get a 4x ROAS on cold traffic, without touching Meta ads?"

The audience is e-com brands, the benefit is a 4x ROAS, and the curiosity is the idea that you can get there without Meta ads, when they believe Meta is the only way.

Then you prompt Claude to write about 3 materially different hooks for each angle, and you pick the best one.

This is the prompt:

"Create 3 materially different hooks for each angle. Every hook must identify the exact audience, include the benefit, problem, or circumstance it's attached to, create curiosity without vague clickbait, be understandable when spoken aloud, avoid generic AI language, avoid claims unsupported by this project, and qualify the audience by revenue or business stage where appropriate. Keep each hook between 15 and 30 words. After writing them, explain the audience, buying reasoning, and curiosity device used in each."

Do 3 hooks across 21 angles and that's 63 unique ads.

You can see how quickly this scales.

The real challenge isn't quantity, it's quality, so you refine the output and train the project over time until it spits out genuinely strong ads.

Step 6: Write the Body With a Direct Offer

Once the hook lands, I build the body using what I call a direct offer ad body.

The idea is simple: as soon as the hook is set, get straight into the offer.

The structure runs like this.

You say "here's my offer to you," explain the offer in one sentence, then say "so that you can" and name a specific outcome they want.

Then you handle the main false belief, stack proof from case studies, tie in why it's working right now, and finish with the call to action.

This is the prompt I use:

"Using the selected angle and hook, write a 60 to 90 second direct offer Meta ad following this exact sequence. Use short spoken sentences, remove filler, do not invent results, do not use corporate language, and do not describe the offer as revolutionary, cutting edge, transformative, or game changing. After the first draft, provide a tighter 45-second version, a standard 75-second version, and a slower 90-second version."

That "do not describe it as revolutionary or game changing" line matters for a practical reason.

Those words get your ads flagged and disapproved, so cutting them protects your account.

And asking for 3 lengths gives you even more variations to test.

On length, people always ask how long an ad should be.

There's no set answer. I've seen 3-minute ads work and 30-second ads work, so test a range of styles.

Bringing It Together

1/ You build a Claude project and feed it your real offer, your case studies, and your proof.

2/ You lock the prompt so Claude never invents facts, results, or guarantees, and flags gaps instead of filling them.

3/ You mine 20 to 30 sales calls into problems, circumstances, outcomes, language, objections, and urgency.

4/ You turn that research into 21 angles, score them by how often they come up on calls, and generate 3 hooks each using the ABC framework.

5/ Then you write the body as a direct offer, in a few different lengths, and you refine the project over time until the quality is consistently high.

Do all of that, and you go from staring at a blank page to producing 63 grounded, on-message ads that actually sound like your market.

If you want help building the full paid ad system behind this for your agency, the link to book a call with my team is in my bio. It's a big part of what we do.

Otherwise, go and build your project. The ads write themselves once the context is right.