how to better use AI (before prompting):

→ before typing any prompt, turn this on:

Here’s how most people use AI:

They open a blank chat → They type a question → They get a mediocre answer → They rewrite the question → Still mediocre → They add more detail → Slightly better → They Google “best prompts for X.” → They copy-paste someone else’s framework → No results → Then they conclude: “AI just isn’t that good yet.”

No.

You aren’t that good yet, at setting up the conditions for AI to perform.

Think about it this way.

If you hired a brilliant consultant but gave them zero context about your business, no documents to review, no examples of what success looks like, and demanded answers in real-time without letting them think or check the internet → would you blame the consultant when the advice was generic?

That’s what you’re doing.

Every single time.

Your prompts fail before you even type them.

You do it wrong, exactly like this:

1. You open a new chat (not a Project).

2. You type a quick prompt (no Thinking, no Search activated).

3. Or worse, you copy a “super magic prompt” to finally do the work for you.

And then you don’t understand why it does not work.

ChatGPT (or Claude, or Grok) did not remember who you are (Project), it can’t follow your instructions (Thinking), and it hallucinates fake stuff (Search).

But this is how it should look like BEFORE you even prompt anything:

1. Extended thinking turned on, to follow instructions.

2. Search activated, to track sources and avoid hallucinations.

3. But more importantly, you are inside a Project for your task.

This step-by-step guide will teach you (exactly) how:

---

Stop introducing yourself.

Every new chat, you explain who you are. What you do. How you like things written. Your tone. Your audience.

Then you do it again. And again.

That’s insane.

It’s like meeting your coworker every Monday morning and saying: “Hi, I’m Ruben. I write a newsletter about AI. I like short sentences. I hate corporate speak.”

They’d think you lost your mind.

But that’s exactly what you do with ChatGPT.

Projects fix this permanently (& how to create one).

Projects are dedicated workspaces. (ChatGPT, Claude, and Grok have them.)

You set them up once. AI remembers forever.

Inside a Project, you can:

• Write custom instructions that apply to every conversation.

• Upload files AI will reference anytime you prompt it.

• Build context that compounds over time.

ChatGPT.

How to set up your Project in 1 minute 26 seconds:

Instructions:

✦ This is custom instructions to share your style & tone of voice.

✦ I also like to define the goal of the project, and what kind of files you added.

✦ I made a free GPT to quickly generate the right instructions for your projects.

Files:

✦ This is the most important part of your project.

✦ Don’t give too much text: it will struggle to know “which one is best”. Less is more. Don’t be lazy and expect AI to differentiate good information from bad.

✦ What I usually do: I create a google doc, give only the best information, and download it as a markdown file before I upload it to my project.

Chat with your project:

✦ Create a new chat inside your projects.

✦ Always use Extended thinking. Yes, it takes longer. Don’t be a baby.

✦ I often prompt “Only use my files as a source”. If not, I turn on search.

Claude.

How to set up your Claude Project in 55 seconds:

Same instructions as ChatGPT for custom instructions/ files.

But I much prefer Claude to write. I explain why here.

Grok.

How to set up your Grok Project in 1 minute 10 seconds:

Same instructions as ChatGPT for custom instructions/ files.

But I much prefer Grok to search. I explain why here.

---

Thinking mode is not “smarter mode”.

When you turn on extended thinking, AI is not exactly smarter.

You’re changing how it processes your prompt.

Without thinking mode, AI operates like a person in a job interview. The AI is giving you an answer because the social contract demands one, now. It pattern-matches to what sounds right and delivers it immediately.

It has to think and talk at the same time.

So you get surface level answers like this:

This is OK, but surface level. Let’s compare it to “Thinking” now.

With thinking mode, something different happens.

The model is allowed to be uncertain. To consider. To backtrack. To notice when something doesn’t fit. To build an actual argument rather than surface level information from pre-training.

Thinking mode changes the depth of solving a problem.

Same prompt. And finally a real insight. Turn on Thinking.

You can prompt questions that require reasoning. Questions where the answer isn’t obvious. Questions where you actually need the AI to figure something out rather than recall something it’s seen before.

Most people never ask those questions because they’ve been trained by bad outputs to expect bad outputs.

Turn on thinking. Then (finally) ask harder things.

Next stop: how to avoid AI from lying to your face.

---

Search to avoid lies.

Yes, turning on web search gives you access to recent data.

But that’s the obvious benefit.

Search is the antidote to hallucination.

Let me explain what’s actually happening when AI “makes things up.”

Before you ever typed your first prompt, the model was trained on a massive snapshot of the internet: books, articles, forums, websites, conversations.

Billions of words absorbed and compressed into patterns.

This is called pre-training. And it’s both the source of AI’s power and its fundamental limitation.

Here’s why:

The AI doesn’t know things the way you know things. It learned patterns of how true-sounding information is structured. It learned what confident answers look like. It learned the shape of expertise without necessarily having the substance.

So when you ask a question it doesn’t have real information about, it doesn’t say “I don’t know.”

It completes the pattern.

The sky is […]. The highest probability is blue, so blue.

It generates what a correct answer would sound like, because that’s what it was trained to do. Fill in the blank. Continue the sequence.

This is how hallucination can happen. Not lying. Not malfunction. Just pattern-completion running unsupervised.

Search breaks this loop.

When you turn on search, you’re forcing the AI to reach outside its frozen training data and touch something real. Something that exists right now. Something the AI has to cite rather than generate.

It becomes accountable to reality in a way that pure generation never is.

Think of it this way:

Pre-training gave the AI a memory of the world as it existed months or years ago, filtered through whatever happened to be on the internet.

Search gives it eyes.

And a model with eyes hallucinates far less than a model running blind on memory alone.

How do I know? Well, I read papers like this, so you don’t have to.

A paper on the lowest hallucination rates reported on HaluBench.

---

Too long, didn’t read.

Skip the guide, and do this:

1. Turn on Extended Thinking before prompting.

2. Turn on Search before prompting (if you need accuracy).

3. Use Projects for each task you want to complete over and over again.

It instantly puts you in the top 1% of AI users.

Now, how can you quickly go from top 1% to top 0.1%?

✦ Join a community of obsessed AI users (like my Slack with 900 professionals).

✦ Play with AI. You learned through games as a kid. You are still a kid.

✦ Pick one AI. Master it completely. Don’t be drawn into noise.

Master AI, before it masters you.

Humanly yours - Ruben.