I remember buying my first Cliff Notes in high school. It was for The Great Gatsby. 300 pages compressed into 30. Character breakdowns, chapter summaries, the thematic conclusions pre-packaged and ready to cite. I could talk about the green light and Daisy's voice and the valley of ashes like I had actually sat with the book.

I had not sat with the book.

What I told myself was efficiency. What it actually was: I traded the lesson for the appearance of the lesson. Fitzgerald was trying to teach me something about longing and class and the myth of reinvention. Cliff Notes told me what he concluded. I left knowing the ending and nothing about how to get there.

That shortcut followed me longer than I'd like to admit. Every time I reached for the summary instead of the source, I got faster at sounding right and slower at actually thinking.

AI is the new Cliff Notes. Except this time, the shortcut is so good most people don't notice they've stopped reading.

Most people are using AI to do their old job faster. That's not the shift.

If everyone has access to the same tool, access isn't the advantage. It's the floor. The prompt is that floor: the minimum, the level playing field. Prompt engineering was last year. Context engineering is this year. The people pulling ahead aren't prompting better. They're building systems that work for them, agents that run without them in the loop, workflows designed around how they think.

Most people will stay at the floor. That gap is widening every month.

The risk isn't that you'll use AI wrong. It's that you'll use it right. And let it do the thinking you should be doing yourself. That's cognitive atrophy. Faster output, softer mind. You stop bringing yourself to the work. And the work knows.

Csikszentmihalyi spent decades mapping the conditions for flow: the state of total absorption where performance peaks and time disappears. His finding was counterintuitive: flow doesn't come from ease. It comes from the tension between challenge and skill. Too easy, you disengage. The sweet spot is the edge of your current capacity, where effort is required.

When you outsource the thinking to AI, you eliminate the friction. Friction is not the enemy of good work. It is the condition for it.

Critical thinking sharpens through use. Insight emerges from wrestling, not from receiving. You find out what you actually believe when you are forced to write it, not when you read what the model generated. The question that took you 20 minutes to work through is yours. The answer handed to you in 20 seconds belongs to no one.

This is the mechanism. AI slop isn't just bad writing. It is the residue of thinking that never happened.

The distinction isn't complexity. It's position.

Most people who think they're building with AI are still prompting. They've learned to write longer prompts, chained a few together, and called it a system. That's not a criticism. It's a useful confusion. The shift between layers isn't technical. It's conceptual. And most people never see the line until they're already past it.

Here's what the line looks like in practice.

Prompt. You type "help me write a response to this client." You paste the email thread. You get a draft. You edit it. Tomorrow, same task. You start from scratch.

Context. The model already knows how you write. It knows your client's history, your communication principles, your standing priorities. You paste the email thread. You're not explaining yourself anymore. You're briefing a colleague who has worked with you for three years. The output isn't a generic draft. It's close to what you'd write — because the environment around the prompt is doing most of the work, and you built that environment once.

Agent. You're not in the loop for the routine. Weekly briefs assemble from your calendar. Follow-ups route to your queue. Work moves without you typing a single prompt. You show up for judgment and voice. The work that only you can do.

The cognitive load doesn't decrease as you move up these layers. It shifts. At the prompt layer, your thinking goes into the ask. At the context layer, it goes into the architecture: the documents, the principles, the self-knowledge required to represent yourself clearly. The model can only be as specific as what you've already worked out. That's harder, not easier. You have to know what you believe before you can build a system that acts on it.

That's why context engineering isn't technical. It's philosophical.

And it's why most people stall between the first and second layer. Not because they lack the skills. Because they haven't done the upstream thinking that makes the second layer possible. You can't codify yourself until you know yourself. You can't automate self-knowledge.

The model is only as sophisticated as the context you bring to it. If you've been outsourcing that thinking, there's nothing to build from.

Write the brief before you write the prompt.

Open a document. Write what you'd want any model to know about you before you ask it anything: how you communicate, what you're actually working on and why it matters to you, where you tend to drift from your best thinking. Not a bio. Not a list of preferences. A working brief: the kind you'd hand a sharp new colleague before their first week.

Then use it. Paste it as context before your next three AI interactions. Notice what shifts — not just in the outputs, but in how the outputs feel.

That document is the seed of your context layer. And the act of writing it will tell you more about how you think than anything the model generates for you.

The model will give you the ending. Only you can learn how to get there.

– Michael
Founder, The Drop In
& Author of 'Human Traits — a novel exploring humanity's relationship with AI'