What Is a Prompt, Really? — The Ilmverse Dialogue

01 / 11 · The message

You type a question.

You open the chat. You type one line. You press send. That line feels like the whole thing you handed the model. It is not.

> ▍

This is the part you see.

02 / 11 · No memory

It forgets you between messages.

The model has no memory of its own. When your message arrives, it does not remember you, your last chat, or even your last sentence. Every message starts from zero.

You What should I cook with rice and eggs?

LLM You could make egg fried rice.

empty mind

So how does it seem to remember? Keep scrolling.

03 / 11 · The chat, again

So it reads the whole chat again, every time.

Because it forgets, the app sends the entire conversation back with every new message. Every earlier line, yours and its own replies, is read again from the top before it answers. That is how it seems to remember.

The longer the chat, the more it re-reads each time.

04 / 11 · Hidden instructions

There are instructions you never see.

Before your first word, the app adds its own instructions. This is called the system prompt. It tells the model who it is, how to speak, what it may and may not do, even today's date. You never see it. The model reads it first.

This is why the same question can get different answers in different apps.

05 / 11 · Three things so far

So far, three things get sent.

Put it together. With every message, the model receives three things, in order. The hidden instructions. The whole chat so far. Then your new line.

And it is about to get bigger.

06 / 11 · It can act

It can do things, not just talk.

The tools people use now, like Claude, Gemini and ChatGPT, can do more than answer. They can search the web, open a file, or run code. A model that can act like this is called an agent.

But to act, it needs to read more first.

07 / 11 · The toolbox

First, it is handed a list of tools.

Before it can act, the model is given a list of its tools and how each one works. Those descriptions are text, and that text is added to the prompt too. Some tasks also load extra instructions, called skills. More text again.

It cannot use a tool it was not told about.

08 / 11 · The loop

It acts, then reads what came back.

Say you ask if it will rain tomorrow. The model cannot know, so it searches. The results do not go to you. They come back to the model, added to the prompt as more text. Then it reads everything again and decides what to do next. Act, read, decide, repeat. This is the loop an agent runs.

Each action makes the page longer.

09 / 11 · One page

To the model, it is all one page.

We have been drawing neat layers. The model does not see neat layers. To the model it is one long stream of text, read top to bottom. Instructions, chat, tools, results, your message, all of it, one page.

It is all just text. That is worth remembering.

10 / 11 · The window

It can only hold so much.

The page has an edge. The model can only hold so much text at once. This space is called the context window. When a chat grows past it, the oldest parts fall away to make room. That is why a very long chat slowly forgets how it began.

context window

A fresh chat is sometimes the kindest fix.

11 / 11 · The real skill

The skill is the context, not the words.

Now the useful part. People hunt for magic words to type. The real skill is quieter. Give the model the right material. The same question, with the right context around it, gives a far better answer. People who work with these tools well call this context engineering.

thin context drag to add context →

what you give it

what comes back

You are not commanding it. You are setting the table.

It only knows what you put in front of it.

The model is a mirror of its window, not a mind that knows the world. It reads the page you build, and answers from that. So build the page with care, and check what matters against real sources. The clearer the context you give, the more it can help you.

Take the AI Readiness Quiz →

■ end ■

You wrote one line.It read the whole page.