Examples and Prompting by Example

LECTURE RESOURCES: Read this lecture in Team-GPT, it is interactive and contains notes and comments from the author.

Most of this course is about examples. Why?

Because learning from examples is comes naturally for us, humans.

How ChatGPT Learns from Examples

Guess what? ChatGPT is also great at learning from examples. In fact, training AI models means to show them as many examples as possible. We call these examples: training data.

Because of this, ChatGPT loves being prompted by example.

In fact, we’ve already seen this technique.

Let’s go back to the ‘unlimited exercises generator’.

Inside Team-GPT I can fork this chat. Forking means: ‘duplicating but only from a certain point’. You can fork a chat from any message.

Add it to the folder 6.8. Examples and Prompting by Example. Let’s also give it a good name.

Using Examples in Your Conversations

Here is the context:

I have just started talking to LLMs. I am trying to think of different questions I can ask and have meaningful conversations with the AI.

Here’s a list of the conversations I’ve had thus far:

1. Write an email
2. How to resolve a conflict with my colleague
3. Learn to code
4. Have better sleep
5. Fix frozen phone
6. How to sell my product better
7. What is string theory
8. Write a poem
9. Write an essay
10. Write a report

The Power of Seeds In AI Conversations

I asked ChatGPT to create a list of 10 more conversation topics, similar to these. (By the way, we are already incorporating the knowledge from the Prompting PDF)

This resulted in 10 more conversation topics you can explore.

Notice how similar they are in structure but different in terms of content.

Truth is that all of these are really, really good chats you can have with the AI.

Then we provided 20 more and then 50 more.

Essentially, we can do this forever.

But no matter how hard we try, all examples will be somewhat similar to the first 10 in terms of structure.

I would call the first 10 examples the ‘seed’.

ChatGPT has captured the essence of these 10 examples and is ready to generate INFINITELY MANY other topics that are similar.

Note that if you provide a different ‘seed’ you would get very different results.

No Examples

Just for comparison, let’s explore the other extreme: no examples.

Inside Team-GPT I will fork the conversation from the first prompt and rename the chat while forking it.

Now I’ll edit it to remove all examples.

I’ll also change the goal to: Please create a list of 10 conversational topics I can have.

The results are not bad per se (see the full conversation here). It’s just that they are all about AI and something.

This output is not very good because ChatGPT has no CONTEXT about what we really want to achieve.

In the same way that Personas provide a lot of CONTEXT with very few words, examples provide a lot of CONTEXT, too.

More specifically, examples provide the PATTERNS we are looking for.

It is hard to explain a pattern but it easy to SHOW it.

You don’t even need to understand the pattern. Prompt well and ChatGPT will ‘get it’.

Create Personas by Example

Let’s do another one before we go.

We discussed personas at length in one of the previous lectures.

Did you create your own persona? Don’t worry if you haven’t, we’ll quickly create one.

Here is the context:

Based on these, please create a business development persona.

There you have it – a new persona has arisen (see the full conversation here). Feel free to later save it to your prompts.

Note that you can do this in a single step, through the ‘Build your own persona’ prompt in Team-GPT.

Making the Most of Examples with ChatGPT

Great! Showing examples to ChatGPT is one of the most powerful techniques.

Remember, ChatGPT sucks at being original. However, it is amazing at replicating things that have been done before.

And examples are just this: things that have been done before.

Good luck!

Take ChatGPT for Work: The Interactive Course for FREE.