There is a version of this problem everyone has already met. A capable new person joins on Monday. They are clever, they learned the theory well, and they want to be useful. But they do not know your filing system, they have never seen last year’s accounts, and they cannot yet tell which of the figures that land in the inbox matter and which are noise. None of that is a flaw in the person. It is simply the first day.
An AI is on its first day every time you open it. The productive afternoon you had with it yesterday is gone. The way you explained your month-end routine, the supplier you told it to watch, the document you taught it to reject: none of it carried over. It begins blank, and it will keep beginning blank until you give it somewhere to begin from.
Why this matters before anything else
You already have the recurring job this course builds around: a folder of receipts, invoices and statements you process accurately at the end of every month. Try this yourself. Open your folder for this month, find one invoice, and attach it to a fresh AI session, either by dragging it in or using the attach button.
Here is one invoice from this month’s folder. What percentage of this month’s total invoices does this one represent?
With one document attached, that question cannot be answered. There is no total to measure it against. The AI will either say so directly, or it will produce a number anyway, resting on a total it was never given. Either way, nothing it says can be a real percentage, because the fact the question depends on, the rest of this month’s invoices, was never in the room.
Now add the whole folder. Not one file at a time: the app can point at a folder directly, and reads everything in it in a couple of seconds. Ask the same question again, and this time it can work it out, because the fact it was missing is now in front of it.
Before you trust that number, notice what you do not yet know. You have not checked that every document in that folder was read correctly, only that an answer came back quickly and sounds plausible. An answer existing and an answer being correct are not the same thing, and the gap between them is exactly what the rest of this course closes.
The fix is not a better question
The instinct is to write a longer, more careful instruction each time: describe every document, every total, every exception, in the message you type. It works once. Then you close the session and all of it is gone, and next month you write it again.
What a director does instead is give the AI a place to know things from. Not a longer message: a standing source of background that it reads at the start of every job, so the knowledge you built up does not evaporate when the session ends. That source is what the next several lessons assemble, piece by piece, in plain conversation. Building that source is a conversation, not a technical task: you answer questions about your own work, and the knowledge accumulates somewhere it can be read again.
The term for giving an agent the background it needs to do a task is context. An AI agent begins with no prior knowledge of its task or what it has available to work with, and supplying that knowledge is the first of the five things a director controls. The widely used industry term for this work is context engineering, and it is what this section is about.