The weights don't change. The architecture doesn't change. The training data doesn't change. The only thing that changes between a useful response and a useless one is what you put into the context window. That's the whole game.
By the time you're talking to a language model, everything about it is locked in place. The billions of weights that encode its knowledge, personality, and capabilities were finalized before you arrived. You cannot change them. You cannot teach it something new mid-conversation. You cannot make it smarter or less capable than it is.
What you can change — the only thing you can change — is what you put in front of it. The prompt. The context. The input.
This sounds like a limitation. It's actually a profound lever. Because the frozen model contains an enormous range of capability, and what you put in the context window determines which part of that capability gets activated.
The same model that produces a vague, meandering non-answer to a poorly constructed prompt will produce a precise, structured, expert-level response to a well-constructed one. The model didn't get smarter. You gave it better context to work with. Prompting is the skill of constructing that context.
From the Overview track: the model generates text by producing a probability distribution over every possible next token, then sampling from it. The context window is what shapes that distribution.
Every word in your prompt shifts the probability landscape for everything that follows. A prompt that begins "Write a poem" opens up an enormous, diffuse distribution of poetic styles, topics, and forms. A prompt that begins "Write a four-line rhyming poem in the style of a Victorian newspaper headline, about the discovery of leftover pizza in a break room refrigerator" collapses that distribution toward a very specific target — one the model can hit with precision because you've told it exactly which part of its learned space to operate in.
You're not commanding the model. You're steering a probability distribution. The more precisely you construct the context, the more precisely you steer.
You've probably seen this: two people use the same AI tool, one raves about it, the other dismisses it as useless. Almost always, the difference is prompting — specifically, the difference in how much context each person provided.
The person who got a bad result asked a vague question and received a vague answer. The person who got a good result — often without realizing they were doing anything special — naturally provided background, constraints, examples, and a clear statement of what they wanted. The model responded to what it was given.
This isn't about tricks or magic words. It's about understanding that the model has no context about you, your situation, your goals, or your standards — unless you put that context into the prompt. Everything it doesn't know, it fills in with statistical averages. Statistical averages are rarely what you wanted.
Think of the model as a piece of highly capable infrastructure with no default configuration. It can handle almost any workload — but you have to configure it for yours. A prompt is a configuration file. A vague prompt is an empty config. You'll get defaults, and defaults are rarely optimal.
A multi-turn conversation with an AI isn't a dialogue in the human sense. It's a growing context window. Every message you send, and every response the model generates, gets appended to a single long document — and the model reads that entire document from the beginning every time it generates a response.
This means your earlier messages still matter. A clarification you provided in message three is still in context at message twenty. An instruction you gave at the start of a conversation shapes every response that follows, because it's still sitting at the top of the context the model reads.
It also means the model has no "memory" of the conversation beyond what's in the context window. If the conversation gets long enough that early messages fall outside the context limit, the model genuinely cannot see them. They're gone. It's not forgetting — it's that they were never in the document it's reading anymore.
Stop thinking of a conversation with an AI as a dialogue. Think of it as a document you're co-authoring — one that the model reads in full each time it contributes. The quality of what it writes next depends entirely on the quality of what's in the document so far.
Think about the last time you got a poor result from an AI. What was missing from your prompt? What context did you have in your head that the model had no way of knowing? How would the response have changed if you'd included that context explicitly?