Stop Prompting, Start Building

Here is a thing you have done this week, probably more than once.

You opened a fresh chat with Claude. You typed something to set it up. Maybe it was "you are an expert marketer" or "act as my business analyst." Then you explained your situation, because it does not know your situation. You told it who your customers are. You told it what you sell and how you talk about it. You pasted in an example of the thing you wanted, because otherwise it writes in that flat AI voice that sounds like a press release from a company that does not exist. And then, finally, five or ten minutes in, you asked for the actual thing you wanted.

It gave you something decent. You nudged it twice, it got better, you copied the result out, and you closed the tab.

Tomorrow you will do all of that again. The setup, the explaining, the pasting, from zero. Because the moment you closed that tab, Claude forgot you entirely. It does not remember your customers. It does not remember your voice. It does not remember that yesterday you told it the exact same things. Every conversation starts from nothing, and you are the one paying to fill in the nothing, by hand, over and over.

That is the trap. And almost everyone is in it.

The industry has moved three times, and you may have only noticed the first

Let me name the three eras of working with AI, because the whole rest of this book is about moving you from the first one to the second, and showing you a glimpse of the third. Once you can see these three, you cannot unsee them, and you will spot exactly which one you are stuck in.

The first era is prompt engineering. This is the blank box. This is the belief that the skill is writing a clever enough instruction. Get the wording right, add the magic phrases, tell it to take a deep breath and think step by step, and you will unlock better output. There was a whole moment where people traded prompts like recipes and sold prompt packs like they were gold.

Prompt engineering is not wrong. Wording does matter. But it has a fatal flaw for anyone who does the same kind of work more than once: you pay the full cost every single time. The clever prompt lives in one conversation and dies with it. You are not building anything. You are performing a setup ritual, getting one result, and throwing the setup away. If your work is a series of one-off questions you will never ask again, fine. But your work is not that. Your work repeats. You write proposals every week. You check your market every Monday. You produce content on a rhythm. And for repeating work, the blank box is the most expensive way to operate that exists, because you rebuild the context from scratch on every lap.

The second era is context engineering. This is the shift the good operators already made, quietly, while everyone else was still hunting for the perfect prompt. The idea is almost embarrassingly simple. Instead of telling the AI who you are inside a conversation that will vanish, you write it down somewhere the AI reads automatically, every time, before it does anything. Who your customers are. How you talk. What you are working on. Your rates, your rules, your standards. You build the world the AI works inside, once, and then it stops asking you to fill in the nothing.

Context engineering is where the leverage actually is. It is the difference between an assistant who needs a full briefing every morning and one who already knows the company because they work here. The whole middle of this book is context engineering in practice, and by the end you will have a setup that reads you before every task instead of forgetting you after every one.

The third era is loop engineering. This one is further out, and I am only planting it here so you know the road continues. It is what happens when the AI stops waiting for you at every step. It drafts something, checks its own work against rules you set, fixes what it got wrong, and only then brings it to you. It runs on a schedule without being asked. Your job changes from doing the work with help to designing the work and inspecting the result. There is a bonus chapter at the end about this, once you have the foundation to appreciate why it matters. For now, just know that context is not the last stop.

Prompt, context, loop. That is the arc. Most people are living in prompt and do not know the other two exist. You are about to move.

What the trap actually costs, in two ordinary weeks

Let me make this concrete, because "you are re-explaining yourself" sounds mild until you count it.

The Monday competitor check. Say you keep an eye on three competitors. Every Monday you want a quick read on what they have been up to. So Monday morning you open a new chat and you type it all out. You are a market analyst. Here are my three competitors. Here is what we compete on. Here is what my buyers actually care about, so tell me what matters and skip the noise. You paste in some links. You get a summary. It is useful. You close the tab.

Next Monday, the tab is gone and so is everything you told it. Who the competitors are. What you compete on. What your buyers care about. All of it, retyped. You do this fifty-two times a year, and every time you are teaching the same class to a student who was wiped clean over the weekend. The summary at the end is the ten percent that is actual work. The other ninety percent is you doing the machine's filing for it, by hand, again.

The proposal that starts blank every time. Or say you write client proposals, maybe one a week. Each one, you open a fresh chat. You re-paste your positioning so it understands what you do. You describe your tone so it does not sound like a robot. You explain how you think about pricing. You paste in a past proposal or two so it has the shape. Twenty minutes of that before you have written a single line of the proposal you actually needed. And the worst part is the output still drifts off your voice, because you re-improvised the briefing from memory and left half of it out, like you do every week, because who can remember the whole thing at eight in the morning.

Neither of these is a hard problem. That is what makes the waste so quiet. Each individual re-briefing feels like nothing, two minutes, five minutes, no big deal. It is the repetition that eats you. It is the fact that you will do it again on Monday, and the Monday after, forever, because the setup lives in a conversation and conversations forget.

The fix is not a better prompt, and I am not going to teach it to you yet

Here is where I have to hold back, on purpose.

Both of those problems disappear the same way. You take the part you keep retyping, the competitors, the positioning, the voice, the rules, and you move it out of your typing fingers and into something Claude reads on its own, automatically, before every task. The Monday check becomes one line: this week's competitor read. The proposal becomes: draft one for this client, usual structure. The context is already there. You built it once. It works while you sleep.

But if I explain how to do that right now, in chapter one, you will build the wrong thing, because you do not yet know the ground it stands on. You do not yet know that Claude comes in three different forms and each one can see a different amount of your actual work. You do not yet know where this always-on context is supposed to live, or that there are several kinds of memory and they behave nothing alike. Pour the context in before you understand the container and it leaks.

So that is chapter two. We are going to look at the three surfaces of Claude, the chat, the desktop app, and the code one, and the single question that decides everything: what can each one actually see. Because the reason your AI keeps forgetting you is not that it is dumb. It is that you have been talking to the one part of it that was built to forget.

Let me show you the parts that were built to remember.