Your AI Doesn’t Know Your Situation
There’s a phrase moving through tech circles this year: context engineering.
Andrej Karpathy helped popularise it, Anthropic has since written about it formally, and the pitch is roughly this; prompting was never really about clever wording. It was about what the model has in front of it when it answers.
Most of what’s been written about that shift is aimed at companies building agents and pipelines. Fair enough, that’s who’s paying for it right now. But the same idea applies to something much smaller and much more common: you, typing a question into ChatGPT on your phone, getting an answer that’s fluent and confident and wrong for your situation, and not knowing why.
Here’s why.
The model answers the question you’d ask if you were nobody in particular
Ask a general question, you get a general answer. That’s not a flaw. It’s the only honest thing the model can do with what you’ve given it.
Say you ask for help writing a difficult email to a supplier. A generic prompt gets a generic email (professional, correctly punctuated, and built entirely out of assumptions). It assumes a size of business, a tone your actual relationship might not carry, an amount at stake it invented. None of that came from you. It came from whatever’s statistically typical for “email to a supplier,” because that’s all the model had to work with.
The email isn’t bad writing. It’s a good answer to a question nobody asked, built to fit an average of every similar request the model has ever seen.
What changes it isn’t a better sentence
Try telling it the real situation instead. Not more politely, more specifically. The actual amount, the actual relationship, what already happened, what you need to happen next.
The model doesn’t get smarter. It just stops guessing, because you took the guessing away from it.
This is the whole shift people are now calling context engineering, stripped of the enterprise framing: the model was never confused about how to write. It was confused about your situation, because you never told it what your situation was.
Rewording a vague prompt fixes vague wording. It does nothing for missing information, because there’s no phrasing clever enough to supply a fact you never mentioned.
Why this one is easy to miss
Because the model never says so. It doesn’t reply, “I don’t know your business size, so I’m guessing.” It writes the confident, generic version and hands it over as if it were exactly what you needed.
The tell isn’t in the answer. It’s in how well the answer actually fits, and that’s the one thing you have to check yourself, because the model isn’t going to flag it.
So the habit worth building isn’t “write better prompts.” It’s smaller and more useful than that: before you ask, ask yourself what the model would have to know to get this right that it couldn’t possibly guess.
Usually that’s a number, a name, a document, or a constraint you didn’t think to mention because it was obvious to you and invisible to everyone else, including the model.
This isn’t new. It’s just got a name now
None of this required the phrase “context engineering” to be true - it’s the same idea whether or not anyone’s writing enterprise blog posts about it.
What the term is really describing, once you take the agent pipelines and the job titles out of it, is something closer to how you’d brief a new colleague: the useful information is rarely the polish of the request. It’s whether the person on the other end actually has what they need to help you.
The tools have gotten good enough that most of us stopped noticing we were skipping that step. Getting it back is the difference between an answer you have to fix and one you can actually use.
When the extra attempt costs something
There’s a practical reason this matters beyond getting a nicer answer. If you’re working on a metered connection, an unreliable network or during a power outage, the distinction isn’t academic.
If the first answer is generic, the second is a correction, and the third finally gets the situation right, you’ve paid for three attempts at something that could have been useful much earlier.
That’s why context isn’t just a technique for getting better AI outputs. Sometimes it’s the simplest way to avoid paying repeatedly for an answer you could have made useful by explaining the situation once.
This is one of the ideas behind Stop Prompting Like It’s 2023, a practical guide to getting better AI answers in fewer attempts.