The Importance of Context Engineering in Human Relationships

Hello.

Since the latter half of last year, the topic of context engineering has been gaining traction in the AI community.
Today, I want to talk about how the concept of "context engineering" actually applies directly to human relationships as well.

What is context engineering?

When generative AI first appeared, the skill associated with it was "prompt engineering"—the art of how to give instructions to an AI. Recently, this has shifted toward "context engineering," which is the design of what kind of context to provide to the AI.

This is actually quite straightforward: a general-purpose LLM knows nothing about the user.
While it possesses highly advanced intelligence, it won't return the intended results unless you share the context of "what is happening now, what the situation is, and what you want it to do."
It's not just about "what you want done," but also about the need to appropriately provide the information leading up to that point.

It feels like common sense, but since we love giving cool names to things like this, we're all excited about calling it context engineering. If a father of one of my daughter's friends at daycare asks me, "What do you do for a living?", I might say, "Lately, I've mainly been doing context engineering.". Just kidding.

Now, when you try to put this into practice, you realize you can't just blindly provide all the information. LLMs have something called a context window, which limits the amount of information they can handle.
Deciding which information to prioritize, in what order to provide it, and how to update that context—thinking through these things for each use case is what is broadly called "context engineering."

Isn't it the same for human relationships?

Now, this is the main point.

This topic of context engineering is often discussed as a tip for using AI, but if you think about it calmly, exactly the same thing can be said for communication between humans.

Just as an AI doesn't know your context, other people also don't know your context.
No matter how close you are to someone, or even if you see them every day, it is impossible to fully grasp what kind of life they have led and what they are thinking right now. Even if you limit the scope to work or "this project," communication rarely starts with both parties having an accurate grasp of what the other is thinking or the relevant information.
If that were possible, one of you would have to be a telepath or a "Newtype." Or perhaps it would be after humanity has undergone Instrumentality.

In the case of AI, it is intuitively understood as a physical fact that "it shouldn't know information you haven't taught it." If you haven't told an AI about your work, it's natural that the AI doesn't know it. No one complains about that.

However, when it comes to other people, for some reason, the following bias arises on its own.

"Doesn't everyone share the same context?"
"Doesn't the other person understand my situation?"

In psychology, there is a concept called the "false consensus effect." It is a bias where you mistakenly believe that your own common sense is the same as everyone else's.
"Your common sense is someone else's absurdity" is a common saying, but that's not just a philosophical idea; it's about the fact that such a bias actually exists in human cognitive structure.

In other words, in human-to-human communication, the premise that "context is not shared" is difficult to recognize in the first place. This is a major difference from communicating with AI.

The problem of thinking "the model is bad"

When you use AI and don't get the output you expected, what do you do?
While there are cases where the model's performance is insufficient, usually you would first try fixing the prompt, reviewing the design of the context you're providing, or making various adjustments.
In other words, the idea that "maybe there was a problem with how I communicated" comes naturally.

However, what about when it's with another person?

When communication doesn't go well, we tend to think "the model is bad," which is to say, "the other person is bad."
If it were an AI, everyone would understand that it's natural not to be able to understand each other without any context engineering, but when it's between humans, for some reason, we conclude that the other person's model performance is low.

I think this is a huge waste.

After all, the root of the problem might not even be there.
Or rather,human intelligence performance is basically a rounding error compared to our AI overlords. There is no major difference.
The problem is not the other person's comprehension, but the possibility that you haven't provided them with the appropriate context in the first place.
Perhaps the necessary background information wasn't shared, the order in which the information was delivered was poor, or it wasn't converted into a form that was meaningful to the other person.

The approach of "revisiting context design" that we do naturally with AIcompletely slips our minds when dealing with other humans.

First, recognize where the problem lies.

The idea that context engineering is also necessary for human-to-human communication is obvious once you think about it. There are many concrete methods for how to actually do this, and they change depending on the type of work and the relationship with the other person.

However, before looking at methods, the most important thing is torecognize that the essence of the problem lies there.

When communication isn't going well,instead of stopping at "the other person is at fault," try taking a step back and thinking, "maybe there is a problem with the context I am providing.". Just doing that should change quite a bit of what you see.

By the way, I wrote extensively about this in my upcoming book, specifically in the context of misunderstandings between engineers and sales staff, and communication gaps between different job rolesI wrote about it in depth in my upcoming book.
It covers how the "lack of shared context" between humans structurally causes problems, and how to overcome it.
If you are interested, please check it out (shameless plug).

The end.