Google Dropped an Amazing 68-Page White Paper on Prompt Engineering.

Here's the best 7 insights from it

Google released a 68-page white paper on Prompt Engineering that is pretty great.

The problem?

Most people ain't gonna read a 68-page technical document.

But since I'm a psycho who's obsessed with giving you every advantage possible to win online...

I went through it in detail.

And I found some pretty interesting insights.

Here are the 7 most important takeaways...

(That can actually help you get better results from AI TODAY.)

The "Three Bears" Method of Example Loading

Remember Goldilocks?

This porridge was too hot...

This porridge was too cold...

This one was juuuust right.

Well, turns out AI prompting works the same way.

You got:

Zero-shot prompting (no examples)

One-shot prompting (single example)

Few-shot prompting (multiple examples)

And just like Goldilocks... that last one is usually "just right."

Especially for marketing tasks.

Here's my rule of thumb:

Give AI three different examples... from three different niches.

(E.g. Here is a great hook... Here is how I used it in the finance niche… here is how I used it in the dog training niche… and here is how I used it the parenting niche… Now use this to create me a hook for…)

That way it understands the PATTERN you want... without getting stuck copying a specific example too closely.

The "CRAFT" Framework Beats Google's Basic Bitch System

Google talks about System, Role & Context prompts.

But I actually prefer something better...

I call it the CRAFT framework:

Context (the situation)

Role (who's talking)

Action (what needs doing)

Format (how it looks)

Target (who it's for)

Think of it like directing a movie:

"In THIS setting... YOU play THIS character... performing THIS scene... shot THIS way... for THIS audience."

Makes it pretty damn hard for AI to screw up when you're that specific.

The "Step Back" Method

This one's interesting...

Instead of jumping straight to what you want...

You first ask AI some high-level questions about the topic.

Example:

Let's say you want a YouTube hook about procrastination.

Most people would just say:

"Write me a hook that makes people want to watch a video about procrastination."

Wrong move.

Instead, FIRST ask:

"What are the most emotionally frustrating moments people experience when procrastinating?"

Let AI spit out answers like:

• Watching deadlines creep closer while feeling paralyzed

• Endless scrolling while life passes by

• Self-loathing for not starting important stuff

• The "just 5 more minutes" lie you tell yourself

THEN hit it with:

"Using those emotional triggers... write a YouTube hook that makes procrastinators desperate to click."

See the difference?

You're giving AI emotional ammo BEFORE asking it to shoot.

The "Shotgun" Strategy

Most people treat AI like it's a monogamous relationship.

They find ONE model they like...

Write ONE prompt...

And expect magic to happen.

But different models have different personalities.

Different strengths.

Different ways of interpreting the same prompt.

So if you're only running your prompt ONCE...

Through ONE model...

You're leaving a TON of value on the table. So use the Shotgun Strategy instead.

Here's how it works:

1. Take your prompt and run it through multiple AI models:

   • Claude

   • Chat GPT

   • Anthropic

   • Gemini

   • Whatever else you got access to

2. Run it multiple times on EACH model

   (AI can be inconsistent... like that friend who's either the life of the party or a total buzzkill depending on their mood.)

3. Compare ALL the outputs

4. Cherry pick the best parts from each

But here's where it gets interesting...

Because after doing this a few times...

You start noticing PATTERNS.

Like...

Claude tends to be better at strategy and analysis.(That nerdy kid who always did the extra credit homework.)

GPT-4 usually crushes creative tasks. (The theater kid who's always "on.")

Anthropic is great at explaining complex shit simply. (That friend who can break down quantum physics using Pokemon references.)

And Gemini... well... (It's trying its best.)

So you start building a ROSTER.

You learn which models to use for specific tasks: • Sales copy • Technical writing • Creative brainstorming • Strategic planning • Content repurposing

And suddenly...

You're not just someone who "uses AI"...

You're an AI CONDUCTOR.

Chain of Thought Prompting (aka "Show Your Work")

Remember in math class when teachers made you show your work?

Same principle here.

Instead of just asking for the final answer...

You make AI walk you through its thinking process step by step.

This is clutch because AI kinda sucks at logic sometimes.

In fact LLMs don’t operate on logic. at. all.

They predict what “sounds right”.

But when you can SEE its thinking...

You can spot where it goes off the rails... and get it back on track.

Tree of Thought (The "Multiverse" Method)

You ever watch that Spider-Man movie with all the different Spider-People?

Where there's infinite universes...

And each one has a different version of events playing out?

That's basically what Tree of Thought prompting is.

But instead of different Spider-Men...

You're exploring different POSSIBILITIES for solving your problem.

Whereas Chain of Thought prompting is like following ONE path through the woods.

Step by step.

In a straight line.

Until you reach your destination.

Which is cool... if there's only ONE right way to go.

But what if there are multiple paths?

Multiple possibilities?

Multiple "right" answers?

That's where Tree of Thought comes in.

Instead of following ONE path...

You're exploring MULTIPLE paths simultaneously.

Like playing 5 different games of chess at once...

To see which strategy works best.

Here's a real example:

Let's say you're writing a hook for a weight loss product.

With Chain of Thought, you might prompt:

"Walk me through, step by step, how to write a hook about this weight loss supplement."

The AI will give you ONE linear path:

1. Start with the problem

2. Introduce the solution

3. Add proof

4. Call to action

But with Tree of Thought...

You'd prompt something like:

"Let's explore 3 different approaches for hooking weight loss prospects:

Path A: Focus on the emotional pain of being overweight

• Break down the specific moments that trigger shame

• Explore various emotional angles

• Consider different storytelling approaches

Path B: Focus on the science/mechanism

• Break down different scientific angles

• Explore various proof elements

• Consider different ways to make complex stuff simple

Path C: Focus on identity/transformation

• Break down different 'new you' angles

• Explore various aspirational approaches

• Consider different ways to paint that picture

For each path, give me 3 specific execution options..."

See the difference?

This works for:

• Hooks

• Headlines

• Email angles

• VSL openings

• Social posts

• Any creative task where there might be multiple "right" answers

The "Inception" Technique

Create prompts... that write better prompts.

I take it a step further. I have it create multiple prompts... tests them... then Frankenstein the best parts together into one SUPER prompt.

Pretty wild stuff.

( I'll break down in detail in a future substack.)

Look...

There's a lot more in that white paper.

But these 7 concepts alone can dramatically level up your AI game.

The key is actually USING them.

Because knowing this stuff and not applying it...

Is like having a Ferrari and never taking it out of the garage.

-Jason

P.S. If you got value from this... maybe share it with someone else who's trying to figure out this whole AI thing.