The AI Mindset: how to talk to AI so it actually helps

Why you should care: two people can open the same assistant, ask the same question, and walk away with completely different results — one with a generic blob of text, one with a plan they actually understand. The tool was identical. What differed was how they talked to it.

Every article after this one teaches you a tool or a technique. This one decides whether any of them will help you. So before tools, before prompts, before anything technical, it is worth settling on your approach to working with AI. That approach determines how you interact with it, how you put questions to it, how you think — and how you perceive it.

Four words before we start

You will meet these four words on every page of this course, so let’s get them out of the way now. No tech background required:

• Chat or Session — one conversation with an assistant, like a messaging app. You type at the bottom, the answer appears above, and the whole thread stays there so you can continue it later.
• Prompt — simply what you type: your question, your instruction, and any details you add. A prompt is not code, it is just text.
• Model — the program that writes the answers. The assistant is the app (ChatGPT, Claude, Gemini); the model is the thing inside it that writes.
• Context — everything the model can see right now: your messages and its own replies in this chat. Context is not permanent memory, unless the tool has a memory feature that you switched on.

That’s the whole vocabulary. Everything below is built on it.

AI is an oracle, not an implementer

The approach I have settled on over years of working with AI is fairly simple: treat AI as an oracle — a mentor, not an implementer. Do that and you get the most out of it. There are certainly uses and practices, especially with the newer models, where you can delegate the work completely and never go into how the result was reached. But our task here is to learn — and for that we need to understand how the tool works.

The wrong way to use AI: treat it as an implementer. “Write me a website.” “Make this feature.” You paste a task, walk away, come back, accept the result. Sometimes it works. Often it doesn’t — and you can’t tell why.

The better way: treat it as an oracle — a well-read mentor who can talk you through most technologies, patterns and common mistakes: not because it knows everything, but because it has read an enormous amount about them. Someone who can guide you, lead you through unfamiliar territory, support you when you’re stuck. You consult it. You discuss with it. But you stay the one who makes decisions.

Here is the same task, asked both ways. Press Send through each one:

Look at who understands the result in each case. In A you get one artifact you have no way to judge — are the hours right, is the menu right, who is this page even for? Only the model knows. In B you answered three questions, chose between two options, and by the end you could explain the site to someone else. Same model, same task: the difference is not the quality of the answer, it is who ends up understanding it.

An implementer you pay and forget. An oracle you keep talking to — before the work, during the work, after the work.

Ask. All the time.

The people who grow fastest with AI are the ones who ask the most questions. Not “do X” but:

• “Why did you choose this approach over that one?”
• “What are the trade-offs here?”
• “What would a senior person do differently?”
• “Explain it to me like I’m five.”
• “What am I missing? What should I have asked you?”

Every answer is a free lesson from something that has read more on the subject than you will ever have time to. Asking is the new reading. You don’t need to study for years before you’re allowed to talk to the best experts in the world — they’re already sitting in your browser, waiting for questions.

One honest warning, though: asking and believing are two different things. Sometimes the model states something completely wrong in a perfectly calm voice. That has a name — a hallucination: a confident answer that isn’t true. So the habit is: ask freely, verify what matters.

Feed it context. Context is the key.

A model that knows nothing about you gives generic answers — the same ones it would give anyone on the planet. The magic starts when the model knows you: your project, your goal, your constraints, what you already tried.

Tell it what you’re working on. Tell it what failed and why. Keep a note with your context and paste it in at the start of a chat (or store it in a tool that does it for you). Correct it when it’s wrong — and watch how much better it gets at everything you do together.

This is not “training” in any scary sense, and nothing is stored forever: the context lives in the current conversation. It’s simply: the more the model knows about your situation, the more powerful it becomes. An oracle without a question is just text.

Context is the single most useful idea in this whole course — there is an entire lesson about it later: what the model actually sees.

And here is the practical consequence for you, right now. Everything an AI product does is one call: you send text, the model writes the most likely continuation, someone sends again. Settings like temperature or a system prompt do exist, but none of them will rescue a badly formulated request. So when an answer is bad, the fix is almost always in the text you sent — a clearer instruction, a missing fact, a shorter context — not in a different model. Better text in, better answers out: fix the text, not the model. That one idea has its own lesson: The one call: how every AI product actually works.

I'm going to work with you regularly, so here is my context. Use it for this whole conversation.

Who I am: [your job, your situation, your level]
What I'm working on right now: [project or task]
What I already tried: [what worked, what failed]
What "good" looks like for me: [short / detailed / with examples]
What annoys me in answers: [long preambles, buzzwords, code I can't read...]

Fill it in once, keep it in a note on your phone, and paste it whenever you start something new. Two minutes that change every answer you get afterwards.

What you should do

• Discuss before you build. Plan with it, poke holes in the plan, ask for alternatives — then act.
• Ask for options, not answers. “Give me 3 approaches and their trade-offs” beats “do it”.
• Explain your context. The more it knows, the better it guides.
• Verify what matters. It’s brilliant and it’s wrong sometimes. Both are true.
• Ask it to teach you, not just to deliver. “Explain why” is the most valuable prompt there is.

What you shouldn’t do

• Don’t outsource understanding. If it writes something you can’t explain — you didn’t gain, you lost.
• Don’t accept the first result. The first answer is a draft for a conversation, not a verdict.
• Don’t hide context and expect miracles. A vague question gets a generic answer.
• Don’t use it only as a search engine. Search gives you links; a conversation gives you options, trade-offs and explanations you can push back on.
• Don’t trust it blindly — and don’t distrust it blindly either. Learn where it’s strong and where it’s weak.

What this looks like in practice

Two conversations with the same goal — “a website for my photography”:

One message, one artifact, done. And it’s… fine. Generic. Not yours.

Same tool, completely different result. Not because the model got smarter — because the conversation got smarter.

Steal this prompt

That second conversation is not a talent, it’s a pattern — and you can copy the pattern. Paste this into ChatGPT and fill in the first two lines:

Before we start: I need you as an advisor, not as a contractor.

My goal: [what I want to achieve]

Do NOT write any final code or text yet. Instead:
1. Ask me 5 questions that would change your advice.
2. After I answer, suggest 2-3 approaches and their trade-offs.
3. Recommend one and explain why.
4. When I choose, work in small steps and explain each one — I want to understand, not just receive a result.
5. If you are not sure about something, say so instead of guessing.
6. Challenge me when my assumptions look wrong.

The first line is the whole trick. “Advisor, not contractor” flips the model into the oracle mode: it starts asking instead of dumping.

Learn faster than everyone else

Here is the compounding part: the more you ask, the faster you learn. And the more the AI knows about you, the better every single answer gets. It’s a loop:

1. Ask → get an explanation → understand something new
2. Feed back what you learned and what you’re doing
3. Get sharper answers → ask better questions

Learning that used to take years can go much faster. Not because the AI does the work for you — because you now have a guide who never gets tired of your questions.

When you want to go deeper: Prompt Anatomy breaks down what a prompt is made of, and When AI helps and when it doesn’t is the honest map of where these tools let you down.

Do this today

1. Open any assistant you already have — or make a free account in two minutes.
2. Paste the “advisor, not contractor” prompt above and fill in one real goal.
3. Answer its questions, then ask “why did you choose that?” about one thing it suggested.

Ten minutes. If you do only this, you are already ahead of most people who “tried AI once and it was fine”.