Context = the Desk in Front of the AI
Imagine the AI is a temp worker who can't remember anything: every time you ask it to do something, its memory is only what's sitting on the desk right now. Whatever is on the desk is all it can use.
So when the AI answers poorly, the first question is usually not "is it dumb" but "is what it needs even on its desk."
Good Results Are a Byproduct of Good Context
Here's a key line: good results are a byproduct of good context. Same model—lay out the information well and the output is good; lay it out messily and the output is bad.
"Context engineering" is exactly this craft: bring what should be brought, take away what shouldn't, and arrange it the right way, so that what's on the AI's desk is exactly what it needs for this job.
A Fuller Desk Isn't Better
A common beginner mistake in the other direction: dump everything onto the desk at once, thinking more is safer.
Good context is carefully selected: keep what's relevant to the current task, take away what isn't. Few and accurate beats many and messy.
自测 · 学完检查一下
想真正动手做题、记进度、攒连胜？到互动课里练。
Using this lesson's metaphor, which is closest to "context"?
答案：All the information laid out on the AI's desk right now that it can see
Context = all the information the AI can see for this one answer, like what's on its desk. It can't remember anything else; it can only use what's on the desk.
Judge: the AI is like a temp worker who can't remember things—each time it works, it can only use "the information on its desk right now."
答案：True
This is the core of the "working memory / desk" metaphor: for one answer, it relies only on the context you give it (unless it's separately hooked to memory/retrieval).
You ask the AI to "shorten that report from yesterday," but it answers something irrelevant. What's the most likely cause?
答案：There's no "that report" content on its desk at all
"That report from yesterday" is in your head but wasn't put on the AI's desk. With the key info missing from context, even the smartest model can only guess.
When the AI answers absurdly, before calling it dumb, you should first check: does its ___ actually contain the key information it needs?
答案：context
The first step in troubleshooting is always to look at the context: is that piece of info you assumed it knew actually in front of it?
Judge: to be safe, you should dump all your materials into the AI at once—the more the better.
答案：False
More context isn't better. Irrelevant information dilutes attention and leads the model astray, burying the one thing that actually matters. Few and accurate is better.
Which practice best fits the standard of "good context"?
答案：Paste only the few key passages relevant to the current task
Good context is carefully selected: keep what's relevant, take away what isn't. Few and accurate beats many and messy.