Prompts, Instructions, and Context

Work through the explanation, apply it to the example, and complete the quick check before continuing.

What the model can use now

Training happened earlier. A prompt is part of the information available now.

The full app may put several kinds of information together before the model creates a response.

These may include:

• System instructions set by the app maker.
• Other app rules or developer instructions.
• Your current message.
• Earlier messages in the chat.
• A file you uploaded.
• Passages found by a search or retrieval tool.
• Results from a calculator or another tool.

Together, this information forms the current context.

The prompt

A prompt is the message or input that asks the system to do something.

For example:

Write a short reminder for parents about Friday's school meeting. Use only the details below.

The details added after that sentence are also part of the current input.

System instructions

An app may give the model instructions before it receives your message.

These instructions may set:

• The role of the assistant.
• The type of task.
• The tone.
• Safety limits.
• The output format.
• Rules about tool use.

Different apps may use different instruction levels. Learners do not need to know every product's exact setup.

The important idea is that your visible prompt may not be the only instruction sent to the model.

The context window

A model can process only a limited number of tokens at one time. This limit is called the context window.

If the available material is too large, the app may need to:

• Leave some material out.
• Shorten it.
• Split the task.
• Find only the most useful passages.

A large context window does not mean that the model will use every detail correctly. Important information can still be missed or mixed up.

Context is not always memory

Chat history shown in the current context can feel like memory. But it is useful to keep two ideas separate.

Current context: information sent with this request.

Saved memory: information stored by the product for use in a later session.

A base model does not automatically keep a personal memory of every user. Memory depends on the product and its settings.

Worked example

Suppose the app sends this context:

Policy: Returns are accepted within 30 days if the item is unused and the customer has proof of purchase. Customer: I bought this 12 days ago and still have the receipt. Can I return it? Instruction: Give a short answer. Do not add rules that are not in the policy.

A good answer should mention:

• The purchase is within 30 days.
• Proof of purchase is available.
• The item must also be unused.

The model may still miss the unused condition. Context helps, but it does not guarantee perfect use of the information.

Remember

• A prompt is current input, not normal base-model training.
• The app may add instructions that you do not see.
• Context may include chat, files, found passages, and tool results.
• The context window limits how much can be processed together.
• Current context and saved memory are not the same thing.

Next, we will study why a model may give a clear but false answer.