What is a prompt

When you message a language model, it receives exactly one thing: a block of text. It has no memory of previous conversations, no access to your intent, and no way to ask what you meant. Everything it knows must be in that block. That block is the prompt.

A prompt is the complete text input sent to a language model. It is the only channel through which you can control the model’s output.

What a prompt is not

A prompt is not a search query. When you search, your words are matched against an index of existing documents. When you prompt a model, it generates a new response from scratch, conditioned on every word you gave it.

A prompt is also not code. It is not executed by an interpreter; the model reads it as text and produces text in return. There are no variables, no loops, and no guaranteed output type unless you ask for one. People sometimes expect prompt instructions to behave with the precision of code. They do not, and understanding why is most of what this course is about.

What a prompt contains

At minimum, a prompt contains a task. “Summarise this article” is a valid prompt. So is a single question. But most real prompts contain more: context the model needs, constraints on how to respond, examples of good output, and instructions about format.

These ingredients interact. A prompt that mixes them without structure works sometimes. A prompt that places them deliberately works reliably. Closing that gap is what the rest of this course does.

How the model uses a prompt

The model reads the entire prompt before producing a single output token. It takes in the full input, builds an internal representation of what was asked, and only then begins the output.

The consequence: every word influences the output. A careless phrase early on can override a careful instruction later. An ambiguous word can send the model toward an interpretation you never intended, and it will not tell you that happened.

The prompt is also finite. Every model has a context window, a maximum number of tokens it can process in one call. Your prompt and the response together must fit inside this limit. A prompt that is too long gets truncated silently, and the model does not flag what it missed.

The model has no intent of its own

This is the most important thing to understand before learning any technique, and it takes the most adjustment.

The model has no goal. It does not want to help you or mislead you. It produces the statistically most likely continuation of the text you gave it, based on what it learned during training. That is the whole mechanism.

When a model gives a wrong or unhelpful answer, the usual cause is not a model failure but a prompt that left too much room to go in the wrong direction. The model went somewhere plausible, just not where you wanted. Better prompts close off the wrong directions before the model reaches them.

In the next lesson, we’ll look at the distinct slots that make up a well-structured prompt, and what each one does.