Prompting LLMs

You can get a lot of results with simple prompts, but the quality scales with how much information you provide and how well-crafted it is. A prompt can contain instructions, questions, context, input data, or examples — all of which help steer the model toward better results.

Here's a simple example:

Prompt:

The sky is

Output:

blue.

If you're using the OpenAI Playground or any other LLM Playground, you can prompt the model as shown in the screenshot below:

One thing to note: when using OpenAI's chat models like gpt-4 or gpt-3.5-turbo, you can structure prompts with three different roles: system, user, and assistant. The system role isn't required, but it helps set the overall behavior of the assistant — think of it as telling the model who it is and how it should respond. The example above only uses a user message, which works fine as a direct prompt. For simplicity, all examples in this chapter (unless explicitly stated) will just use user messages with the gpt-3.5-turbo model. The assistant message in the example above is the model's response. You can also define assistant messages to show the model examples of desired behavior. More on using chat models here: https://www.promptingguide.ai/models/chatgpt

The takeaway from that example: the language model completed the text based on the context "The sky is". The output might surprise you or miss the mark entirely. And that's the point — it highlights why you need more context or clearer instructions about what you actually want. That's what prompt engineering is all about.

Let's improve it:

Prompt:

Complete the sentence:

The sky is

Output:

blue during the day and dark at night.

Better, right? We told the model to complete the sentence, so the output makes more sense — it did exactly what we asked ("complete the sentence").

These examples show what today's LLMs can do at a basic level. They handle all sorts of advanced tasks: text summarization, math reasoning, code generation, and more.

Model settings

If you notice the same prompt giving wildly different outputs each time, check your model settings first (like temperature / top_p). Start with the next chapter:

• /learn/prompt-master/settings

Prompt formats

The examples above used pretty simple prompts. A standard prompt usually follows one of these formats:

<question>?

Or:

<instruction>

You can also use Q&A format (standard in many Q&A datasets):

Q: <question>?
A:

When you prompt like this without providing examples, it's called zero-shot prompting — you're asking the model to respond without any demonstrations of the task. Some LLMs can handle zero-shot prompting well, but it depends on task complexity, knowledge coverage, and how the model was aligned during training.

A zero-shot prompt example:

Prompt:

Q: What is prompt engineering?

With newer models, you can skip the "Q:" part entirely — the model figures out it's a Q&A task from context. So the prompt simplifies to:

Prompt:

What is prompt engineering?

Building on these standard formats, there's a popular and effective technique called few-shot prompting, where you provide examples (demonstrations). Here's the format:

<question>?
<answer>

<question>?
<answer>

<question>?
<answer>

<question>?

Q&A version:

Q: <question>?
A: <answer>

Q: <question>?
A: <answer>

Q: <question>?
A: <answer>

Q: <question>?
A:

Whether you use Q&A format depends on the task type. For instance, you can do a simple classification task with examples like this:

Prompt:

This is awesome! // Positive
This is bad! // Negative
Wow that movie was rad! // Positive
What a horrible show! //

Output:

Negative

LLMs can pick up tasks from just a few examples, and few-shot prompting enables this in-context learning ability. We'll cover zero-shot and few-shot prompting more extensively in later chapters.