learn this skill if you want to become irreplaceable

2 years ago, I was struggling to get AI to (actually) do what I wanted it to do.

I was 100% confident it was going to be the best thing I’d EVER used since Windows XP (remember the green and blue?)

Alas, AI was totally rubbish.

And trying to get it to do your bidding was like asking your 4 year old to cook you a steak.

Blue, or burnt to a cinder.

Side note - Ironically, recipes are actually an AI forte (you can boss gluten-free substitutions like nobody’s business).

Delicious food analogies aside.

I invested in paid versions of all of the tools, but would still hit brick walls.

I had to resign myself to the fact that some things just weren’t possible.

(until now)

The Dream

Ever since ChatGPT was old enough to talk, my dreams have been to get AI to:

• Write content like me (preferably better).

• Write cold e-mails like me (better doesn’t exist).

You will have your own Unicorns.

[Enter picture of Nicolas Cage in Gone in 60 seconds]

“Impossible”

“Can’t be done”

“You’re a maniac”

They (me) said.

And it did seem that way.

When I launched my ChatGPT guide in 2023. It was a complicated framework I named SPICES.

1. Scenario - The context

2. Profile - How you wanted it to act

3. Information - Data

4. Command - What you wanted it to do

5. Example - Examples

6. Structure - How you wanted it to output

All that, just to have a decent conversation with ChatGPT?

But then a new model was released.

And then another model was released.

Wave after wave of increased intelligence.

Well - that’s what the marketing would have you believe.

Either way…

• GPT3.5 - Sucked hard

• GPT4 - Was meant to take over the world (still sucked)

• GPT4o - Was a terrible name, but at least it had Scarlett Johanssons voice (and then it didn’t).

And then I found Claude.

Ah Claude, you quickly turned into my bezzie mate.

I finally felt like I’d found an AI that ‘got me’ you know?

Quickly skipping past the love story.

The breakthrough moment for me was the concept of ‘reasoning’.

This is where, instead of outputting instant bullsh**, a model will try to ‘think’ something through to more accurately produce what you’re asking of it.

Reasoning models include:

• OpenAI - o-series (o1, o3, o4-mini)

• Anthropic - Claude 3.7 Sonnet ← Yes!

• Google - Gemini 2.5

• xAI - Grok-2

And then a load of other's I don't really care about.

But why does all this matter?

1. Easier to train.

2. Less variability outputs.

3. More trust in the models.

This is all super important. Especially if you're using AI to reach out to thousands of cold prospects without being able to research all of the companies yourself.

Until recently, I'd still struggle with this 'me vs it' mentality. i.e, this is a me task, that is an AI task.

But what if AI was a true extension of yourself?

I'm not talking about neural implants.

But I am talking about a deeper level of Human to AI symbiosis.

AI Symbiosis

A mutual symbiotic relationship in biology is where two organism benefit from working together.

Example: Gut bacteria and humans - Bacteria get nutrients, humans get digestive assistance

(thank you for that one Claude - that can now be connected to the internet BTW)

A fundamental shift for me was moving from:

Do it for me.

to…

Do it with me.

So now, instead of trying to be clever, and tricking the AI into doing my bidding by loading it with a load of clever instructions.

I train it to act as I do, giving step step instructions on how I'd approach the goal if it were up to me.

This of course, relies on you having experience doing the thing that you want AI to do.

Sure, you can use AI to help you with stuff you don't have experience with, but that's not the approach we'll be looking at today.

Prompting to Achieve Impossible Results

Spoiler.

My dream of using AI to write content like me (or better than me) is on pause. I actually enjoy writing, and I find using AI to write sucks the joy out of it for me.

(there's something therapeutic about the physical act of typing your thoughts from start to end, and going through the pain of creating something from nothing)

But I have made leaps and bounds with cold outreach.

So…

Cold outreach, here we go.

[Note - These principles can be applied to every challenging task that you might have. I'm just talking through a recent project here.]

There are 2 main ways to use AI.

1. Directly via a chat window where you have a conversation (e.g ChatGPT)

2. Via an API, where you use another application to 'call' the AI.

The two main outreach platforms I've been using recently are:

1. Apollo

2. Clay

Both call AI models via API when you're carrying out your outreach.

There are AI prompt templates that you can use (but they're pretty rubbish)

So this is where we need to put our prompt engineering hats on.

In the below example I've split my prompt into user vs system. But they can be combined if the tool you're using won't allow you to insert a system prompt.

• System Prompt = The training and background knowledge the AI needs to carry out the task - Like a Director's instructions given before filming begins.

• User Prompt = The question asked by the user - Like the lines of dialogue from the actors in scene.

As a general rule:

I use system prompts to get the AI to assume a role, add context, describe how to handle data, give instructions, and specify the output..

I use the user prompts to give the data, and add any further instructions that aren't included in the system prompt.

User Prompt Screenshot

System Prompt Screenshot

Here's my old prompt structure:

—

User Prompt

Please review the below data and carry out your instructions

- Data Point 1 (e.g company name)
- Data Point 2 (e.g social media posts)
- Data Point 3 (e.g LinkedIn Profile info)

System Prompt

# Role

Act as a (e.g masterful SDR / copywriter)

# Context

You have been tasked with... (e.g writing a cold e-mail)

## Data

Here's a description of the sort of data you might receive from the user

- Data Description 1 (e.g company info)
- Data Description 2 (e.g social media posts)
- Data Description 3 (e.g LinkedIn profile info)

# Instructions

Using the information you've received from the user. Please {enter goal} (e.g generate a personalised cold e-mail)

Complete the below tasks before generating your output.

Take a deep breath, and work through the tasks step by step.

Let's begin:

## Task 1

Review the… (e.g cold e-mail structure)

## Task 2

Use the data to (e.g personalise the e-mail)

## Task 3

Check using these rules (e.g 100 words or less)

# Output

- Do this (e.g 10/10 cold e-mail)

- Don't do that (e.g output anything that isn't the e-mail body)

# Examples

- Example 1 (e.g successful cold e-mail template)

—

It was OK.

But there was still a wild variance in my output quality.

So, after an evening of chasing the kids around…

I suddenly thought.

Instead of giving tasks where the AI is trying it's best to complete each one separately.

What if I used steps that guided the AI towards an output based on how I'd do it in real life.

Sounds, so simple right?

But it worked:

Here's my new prompt structure:

—

# Instructions

Everything is the same, apart from the instructions.

Using the information you've received from the user. Please {enter goal} (e.g generate a personalised cold e-mail)

Work through the following steps before generating your output.

Take a deep breath, and let's begin:

## Step 1

Think about… (e.g your e-mail structure, and the placeholders you'll need to fill.)

##Step 2

Work through {first_step} (e.g first line personalisation)

Examples:

• {Example} (e.g looks like you're hiring BDMs right now)

##Step 2

Move onto {next step} {e.g third line personalisation.}

Example - {Example} (e.g It can be risky leaving SDRs to their own devices.)

## Task 3

Check using these rules (e.g 100 words or less)

—

It's up to you how explicitly you want it to follow a process, or how much flex you want to give the AI in carrying out the task.

Either way, the end result is symbiosis where the AI gets the benefit of your expertise and your thought process, and you get the benefit of something that can carry out tasks quicker and more efficiently than you'd be able to do on your own.

But wait,

I still wasn't quite there.

There was just something about the outputs that wasn't quite as natural as I'd like.

I looked at the model I was using:

OpenAI gpt 4o-mini

Ah…

So I switched to Claude 3.7 Sonnet.

It was like night and day.

The Result

130 personalised e-mails following a proven outreach structure.

That's 10 hours work saved compared to doing it manually.

But, it gets better than that.

Most people don't personalise follow-up e-mails in sequence.

But using the same method, I now can.

So that's 390 personalised e-mails.

I'm literally saving weeks worth of work.

Key Takeaways

1. Reasoning models tend to be more accurate than those that aren't.

2. When prompting, don't think in terms of tasks, think in terms of steps.

3. Accuracy increases with more expensive models (so don't skimp)

4. I'm in love with Claude 3.7 sonnet.

That's all for this week folks.

Catch up soon.

Adam

P.S - There're 2 ways I can help you.

1. Struggling with AI, sales or mindset? Book me for a 1:1 here.

2. Sign up to the waitlist for my new AI for Sales course here.

3. Join me in the salesman.com Academy - Find and close more business in 30 days or your money back (guaranteed) - Find out more here.