What is a prompt?

You type one message. The model receives much more.

The text box only shows what you typed.

Before your message reaches the model, the app may add other information.

This can include instructions, documents, chat history, tools, and formatting rules.

Your message is usually only one part of the full prompt.

In real AI products, much of the prompt is created by the application.

Everything becomes one sequence.

The model does not receive each section as a completely separate input.

The text is turned into tokens and placed into one sequence.

The model then processes that sequence together.

This means instructions and outside content can sometimes affect each other.

This is one reason prompt injection can happen. Untrusted text can contain instructions too.

The model usually reads the conversation again.

A chatbot can feel like it remembers everything you said.

But the model itself usually does not keep the conversation between requests.

The app sends the earlier messages again when you send a new one.

Longer conversations therefore use more tokens.

This is why long chats can become slower and more expensive.

Clear instructions lead to better answers.

A vague request gives the model many possible answers.

For example, “Write about our product” does not explain what you actually want.

Add the audience, goal, format, and limits.

The request becomes much easier to follow.

Useful details usually matter more than phrases like “please” or “you are an expert”.

Sometimes it is easier to show the model.

You can explain the format you want using words.

But examples are often clearer.

Give the model one or two examples of a good answer.

It can then copy the structure, tone, and level of detail.

A few good examples are often enough. Too many examples also use more tokens.

Breaking a problem into steps can help.

Some questions are difficult to answer in one jump.

It can help to split the task into smaller steps.

Each step gives the model more useful information for the next one.

This can improve the final result.

Some newer models already spend extra time working through difficult problems internally.

The app may add its own system prompt.

Many AI products add instructions before your message.

These instructions are called the system prompt.

They can describe the assistant, its rules, its tools, and the format it should use.

You usually do not see these instructions.

The same model can behave very differently when the system prompt changes.

Treat prompts like code.

In a real product, prompts should not change randomly.

Save different versions of them.

Test them with the same set of questions.

Check that a change improves results instead of breaking something else.

Good AI products test prompts just like developers test other parts of software.

A better prompt cannot fix everything.

A prompt cannot give the model information it does not have.

It cannot completely stop incorrect answers.

It also cannot make the model perfect at every task.

Sometimes you need documents, tools, better data, or a different system.

Prompting is useful, but it is only one part of building a good AI system.

Your words are a fraction

The system prompt, retrieved documents and history are usually far larger than anything you typed.

There is no memory

Every turn re-sends the whole conversation. Nothing persists inside the model between messages.

Cost grows per turn

A long chat re-reads itself each time, which is why it gets slower and dearer as it goes.

Constraints beat adjectives

Audience, format, length and exclusions change the output. “Be creative” mostly does not.

Examples beat description

Two or three worked examples settle format and vocabulary faster than any paragraph of instruction.

One channel, no boundary

Instructions and data arrive as the same text, which is why a prompt can never be a security control.