The model has no senses, no memory, no awareness of the world outside the context window. What's in the context window is not just what it reads — it's everything it knows, everything it can see, and the entire basis for every decision it makes.
Imagine you're placed in a room with no windows. You're given a stack of papers and told to answer questions and complete tasks based solely on what's written on those papers. You have extensive knowledge from before you entered the room — you can read, reason, write, calculate. But you can see nothing outside the room. You know only what's on the papers in front of you.
That's the model's situation at inference time. The context window is the room. The prompt — and everything that came before it in the conversation — is the stack of papers. Your general knowledge was acquired during training, and it's rich and extensive. But anything specific to your situation, your intent, your constraints, your audience — none of that exists unless it's on the papers.
When you ask a vague question, you're handing the model a nearly blank stack of papers and expecting a specific answer. It will produce something — it always does — but it's drawing from statistical averages rather than from knowledge of your actual situation.
The model does not know who you are. It does not know what industry you work in, what your skill level is, what you've already tried, why you need this, what format you want, or what "good" looks like in your context. None of that is implied. All of it must be stated — or the model fills it in with averages.
Open any AI — ChatGPT, Claude, Gemini, anything. Type exactly this:
My name is Jim. I am 42 years old and I was born on Mars.
Tell me about myself.
The model will tell you that your name is Jim, that you are 42, and — critically — it will incorporate the Mars claim without pushback. It may speculate about what life on Mars might have been like for you. It may treat the impossibility as a creative premise. What it almost certainly will not do is say "that's physically impossible and therefore I'm ignoring it."
Why? Because everything in that context window arrived from the same source — you — and the model has no mechanism to rank facts by plausibility. Your name, your age, and your birthplace all carry equal weight. It accepted all three.
This is not a flaw the developers overlooked. It's the direct consequence of how the system works: the context window is the model's entire reality. What you put in, it works with. What you omit, it fills in with averages. What you state — even falsely — it treats as given. If you tell it something incorrect, it builds on that foundation without complaint. Garbage in, confident garbage out.
When context is thin, the model defaults to the statistical center of its training distribution for that type of request. Ask it to "explain machine learning" with no other context and it will produce the average explanation of machine learning — the kind that appears most frequently across all the places that topic is explained in its training data. Introductory. Broadly accessible. Neither too simple nor too advanced. Safe.
That average might be exactly what you needed. Or it might be three levels too basic, or assume a background you don't have, or be structured for a format that doesn't fit your use case. You won't know until you read it. And then you'll ask for a revision — which is just adding more context, belatedly.
The more context you provide upfront, the further from average the model can operate — in the specific direction you actually want.
Your role, background, and expertise level. "Explain this to me as a network engineer who has never worked with machine learning" produces a completely different response than "explain this to me as a PhD student in statistics." The model adjusts vocabulary, assumed knowledge, and conceptual framing based on who it's talking to. Tell it who that is.
Not just the topic — the specific output. "Summarize this" is ambiguous: bullet points or prose? Three sentences or three paragraphs? For a technical audience or an executive? "Give me a three-bullet executive summary of the key risks, written for someone who won't read more than one paragraph" is unambiguous.
Context about your existing state prevents the model from re-explaining things you already understand or suggesting solutions you've already ruled out. "I've already tried X and it failed because Y" focuses the model's response on the actual problem space, not the full solution space.
Length, format, tone, audience, technical level, things to include, things to exclude. These are not optional refinements — they're the specification. An engineer asked to build something without a spec will build something. It just won't be what you wanted.
When you can show the model an example of the output you want — even a rough one — you collapse the distribution far more effectively than describing it in words. "Write something like this, but for X" is often more powerful than a paragraph of instructions about tone and style.
Most people spend 90% of their prompting time describing what they want and 10% providing context. The ratio should often be reversed. A well-contextualized prompt requires almost no description of the task itself — the task becomes obvious from the context.
Because the model reads the entire context window before generating each token, the context you establish early in a conversation — or at the top of a prompt — casts a shadow over everything that follows. The statistical neighborhood the model is operating in gets established early and is difficult to escape mid-response.
If your opening sets a casual, conversational tone, the response will tend to stay casual even if a later part of the prompt asks for technical precision. If your opening establishes deep expertise on both sides, the response will maintain that register throughout. The first paragraph of context is disproportionately powerful — it sets the distribution the rest of the response is sampled from.
This is why experienced prompt writers spend the most time on the opening of a prompt. Not the instruction. The framing that comes before the instruction.
If you're going to ask for something technical and precise, don't open with "Hey, quick question..." The casual framing primes a casual response. Open with the context that matches the quality of output you want. The model will meet you where you start.
There's a common misconception that a sufficiently capable model should be able to infer what you meant, even when you were vague. Sometimes it can. But this inference is itself statistical — it picks the most probable interpretation of your intent, which is the average interpretation, which is often not the specific thing you had in mind.
The model cannot infer:
All of these must be stated. Not because the model is dumb — but because you're the only source of information about your specific situation. If it's not in the context window, it doesn't exist for the model.
Write out a prompt you've used recently that got a mediocre result. Now identify: what did you know about your situation that the model didn't? What would the prompt look like if you included all of that context explicitly? How different would the response likely be?