A language model works with the text in front of it: your messages, anything you paste in, and its own earlier replies. That is its context. It doesn't know your school's schedule, which library you mean, what you can't do, or what you already tried. In 1.2.2 you saw that a language model writes what fits, not what it looked up. When your prompt leaves a gap, the model fills it with something plausible, and it usually doesn't tell you that it guessed.
Context engineering means deciding what to put in front of the model so it doesn't have to guess. It is the difference between "How do I get to the mall?" and a prompt with the time, the starting point and the real bus schedule.
Not every detail earns its place. Here is a rule that doesn't depend on opinion:
Try it. Everything in this lab is invented. Start with only the essentials and watch what the AI is forced to guess.
Switch details on and off. The prompt always asks: Which bus should I take tomorrow?
The AI would have to guess: when you can leave; where you are starting from; the real bus times (it may make some up).
Noise doesn't break anything, but it doesn't help either. It just makes the prompt longer. Try adding every detail, then take away one essential at a time.
Notice that the weather detail is noise here, because you've already decided to take the bus. In another prompt the same detail could be essential (for example, "Should I bike or take the bus?"). A detail's role depends on the question, not on the detail.
Inside one chat, the model rereads the whole conversation each time it replies. That is why a detail you gave earlier still shapes the next answer. The model does not rewrite itself while you chat. A new chat usually starts fresh, unless the tool has a memory feature that carries details over. Separately, a company may store conversations and, depending on its policy and your account settings, use them later to improve its products. What happens varies by tool and by account, and school accounts often work differently from personal ones.
A brief is a prompt that hands the AI what it needs, in an order it can use. Four parts cover most everyday prompts:
Then add one more line: "If anything essential is missing, ask me before you answer." A good AI will often ask. That is the AI telling you what you left out, instead of guessing.
An invented example. You start with: "Tell me how to get a library book." Build it into a brief.
Here is the finished brief. Compare it with the one-line version you started with.
Some AI tools can open links and files, and some can't. A tool that can't may still answer as if it had read the page, and it can make up what the page "says." If the details matter (a bus schedule, a menu, a set of rules), paste the text into the prompt. Then you know what the AI actually saw.
Context can hurt as well as help. Three ways it goes wrong:
A good brief makes the answer fit you: it takes the bus because you don't drive, and it picks a time after dinner. That is a real improvement. It is not the same as being true. The AI can follow every limit you gave it and still quote the wrong hours, add up the wrong total, or cite a study that doesn't exist. A well-briefed wrong answer is harder to catch than a vague one, because it looks exactly like what you asked for.
The fit check happens inside the answer: did it follow each limit I gave? Reread your brief and tick them off. It catches an answer that ignored you. It cannot catch an answer that is wrong. For that you need an outside check: something that comes from outside the answer, as you learned in 1.2.2. Asking the same bot "Are you sure?" isn't one.
| Check | What it covers | How to do it |
|---|---|---|
| Facts | Hours, prices, schedules, names, dates, places | The official source for that fact (the library's own site, the transit agency's schedule). Write down the date you checked. |
| Numbers | Totals, times, quantities, percentages | Do the arithmetic yourself from the numbers you gave it. Don't ask the same AI to "check its math"; it can repeat its own mistake. |
| Sources | Studies, quotes, "experts say", statistics | Find the original yourself and confirm it exists and says that. If you can't find it, don't use it. |
An invented example. You paste the brief from the last section but leave out the pasted hours. The AI replies: "Fairview Library is open until 9 p.m. on Saturdays, so go at 7:30."
Numbers are the easiest claims to check, so do it every time. If the AI says two $4 slices and a $2 drink come to $12 per person, you can redo it yourself: 2 × 4 + 2 = 10. It was off by two dollars, and you can see that without any outside website. For a claim like "studies show students remember 40% more," look for the study. If you can't find it, it's not evidence, no matter how specific the number sounds.