Prompts

Whenever you type into an AI chat box and press send, you have written a prompt. The word is used so often that it is rarely explained. This lesson explains it.

A model reads a lot more than your message. The lesson on context listed the parts of a request: hidden instructions, the conversation so far, tool descriptions, and more. The prompt is the one part of all that which you type yourself. In an ordinary chat app, it is also the only part you get to change directly.

This lesson covers what a prompt contains, how the meaning of the word has grown from "one line of text" to something larger, and where the prompt sits inside the full request. After this lesson you can say exactly which part of a model's input is your prompt, and why improving that part is the main way a user influences the answer.

1. A prompt has two halves

A prompt can hold two different kinds of thing.

The left column is the task. It is an instruction: summarise, rewrite, find. The right column is the material. It is what the instruction is about.

Look at the three tasks again. Each one contains a word that points at something else: "this", "it", "the policy". None of them can be carried out alone. "Summarise this" needs a "this". The material is what those words point at.

Both halves together are the prompt. Usually you send them in the same message.

A prompt can also be only a task, with no material. "Suggest five names for a bakery" is a complete prompt. In that case the model works only from what it learned in training. A prompt that is only material, such as a pasted paragraph with no instruction, leaves the model to guess what you want done with it, and the usual guess is a summary or a comment.

2. A prompt is more than one line in a box

The word "prompt" still suggests one line of text typed into a box. That was accurate for early text-only chat tools, where a typed line was all a user could send.

It stopped being only that some time ago. Many current models are multimodal, which means they accept more than text: images, documents, and in some products audio. Chat apps let you attach several files to one message. So a prompt today is often a bundle: some typed text plus one or more attachments.

The typed line is still there. It is one part of the prompt and no longer the whole prompt.

3. A prompt as it is sent

Here is a realistic prompt, as it would look in a chat app before pressing send:

new message

  supplier-terms-march.pdf         document
  supplier-terms-september.pdf     document
  delivery-note-scan.png           image

  Compare the delivery terms in these two, then check
  the scanned note against the newer one.

This message has four pieces: two PDF documents, one image, and two lines of typed text. The typed text is the task. The three attachments are the material.

The task refers to the material without naming the files. "These two" means the two documents with supplier terms. "The newer one" means the September document. "The scanned note" means the image. The model can resolve these references because all four pieces arrive together.

The three attachments and the question are sent as one thing, and that one thing is the prompt.

When the prompt reaches the model, there is no separate channel for files. As the lesson on context showed, the app places the contents of each attachment into the same input as the typed text. The text of the two documents becomes text tokens, and the scanned image becomes image tokens. So a prompt with three attachments is much larger in tokens than its two typed lines suggest.

4. Where the prompt sits in the request

The prompt is one piece of a larger whole. That whole is the context: everything the model receives in one request.

flowchart TB
    subgraph REQ["What reaches the model"]
        direction TB
        A["System instructions"]
        B["Tool descriptions and other additions"]
        C["The conversation so far"]
        D["Your prompt<br/>(you wrote this)"]
    end
    REQ --> M["The model"]

The diagram shows four parts going into the model in one request.

• System instructions are rules written by the company that runs the app, such as "Answer briefly".
• Tool descriptions and other additions are text that the app adds so the model knows what actions it may take, along with items such as saved memory.
• The conversation so far is every earlier message and answer in the chat, sent again each time.
• Your prompt is the new message with its attachments.

Three of the four parts are added around your prompt by the software, and they are mostly unseen. You wrote one part.

The diagram lists the parts in the order they are usually arranged in a real request: the instructions first, the history next, and your new prompt at the end.

An everyday comparison: you hand a note to a receptionist, who clips it to a folder that already holds the office rules and the earlier correspondence, and then passes the whole folder to the specialist. The note is your prompt, and the folder is the context. The comparison stops being accurate at one point: a specialist remembers the earlier correspondence, while a model has to be given the whole folder again on every turn.

5. The prompt is the part you can change

Of the four parts, the one you write is the one you can change. The other three arrive whether you want them or not.

• You cannot edit the system instructions of a chat app. The company wrote them. Some apps have a settings page where you can add standing instructions of your own, but the company's instructions stay in place.
• You cannot remove the tool descriptions. The app adds them.
• You cannot rewrite what was already said in the conversation, although starting a new chat does clear it.

This has a practical consequence. When an answer is not what you wanted, the place to act is the prompt. A clearer task, the right material, or both, is the direct way to get a different answer.

A small before-and-after shows the effect.

first prompt:
  What do you think of this?
  [ supplier-terms-september.pdf ]

second prompt:
  List every clause in this document that changes
  the delivery deadline, with the clause number.
  [ supplier-terms-september.pdf ]

The material is the same in both. The system instructions, the tools, and the model are the same too. Only the task changed. The first prompt lets the model choose what to comment on, so the answer could be about anything in the document. The second names what to find and how to present it, so the answer is a list of clauses about deadlines. Writing the task well is a skill of its own, called prompt engineering, and it is the subject of the next lesson.

6. Other uses of the word

The word "prompt" appears in a few related phrases, and it helps to keep them apart.

The statement that the prompt is the only part you can change is true for a person using a finished chat app. A developer who builds an app on top of a model is in a different position. The developer writes the system prompt, decides which tools are described, and decides what else is added to each request. Shaping all of those parts is called context engineering, and a later lesson covers it.

Interview points

• A prompt is what the user sends to a model in one turn. It is one part of the context, not the whole of it.
• A prompt has two possible halves: the task, which says what to do, and the material, which is the text, files, or images the task refers to.
• With multimodal models a prompt is often a bundle of typed text and several attachments, sent together as one message.
• Attachments are merged into the same input as the typed text, so they count toward the token total and the context window.
• A request typically contains system instructions, tool descriptions and other additions, the conversation so far, and the user's prompt, usually in that order with the new prompt last.
• In a chat app, the user controls only the prompt. The application adds the other parts, and they are mostly not visible.
• Because the prompt is the user's only direct control, changing the task wording or the supplied material is the main way a user changes the answer.
• "System prompt" refers to the instructions written by the company or developer. Developers control that part and the rest of the context, which users of a finished app do not.

Summary

• The prompt is the part of the model's input that you write.
• It can contain a task, material, or both.
• A prompt today often includes documents and images, not only a line of text.
• All pieces of a prompt are sent as one message and become one piece of the context.
• The app adds system instructions, tool descriptions, and the conversation history around the prompt.
• In a chat app, the prompt is the one part a user can change.