The importance of prompts in your interaction with GPT-3

GPT-3 (Generative Pre-trained Transformer 3) is a language model released by OpenAI on June 11, 2020. Access was originally granted via a waiting list. The API was made generally available in November 2021. 

Since then, dozens of startups have launched AI-powered writing assistants. The most popular on the market at time of writing (July 2022) are Jasper (formerly Jarvis), Rytr and Copy.ai. 

The main use cases for AI writing assistants are advertising copywriting, email copywriting and long form blog creation. We also see the emergence of AI-assisted fiction writing, as reported by The Verge on July 20, 2022.

These are exciting times for creators, even if the technology also raises multiple ethical questions in terms of copyright and fact checking. 

One of the least discussed aspects of this AI revolution is the crucial importance of the quality of prompts when interacting with the AI language model. If, just like me, you’re a heavy Googler, you know that the relevance of the search engine results depends on the quality of your query. The better you phrase your question, the closer you’ll get to what you’re looking for. 

It’s exactly the same when you’re interacting with GPT-3: the relevance of the output depends on the accuracy of your prompt. 

How can you test GPT-3?

You can test GPT-3 on OpenAI’s website, which offers a GPT-3 playground. Access is free and you get a grant of $18 upon signup, to test the various types of models. 

If you want to experience the full power of GPT-3, I advise you to pick “text-davinci-003” in the drop down on the right hand side of the interface.

You’re presented with a blank canvas.

The answer you’ll get from the model now solely depends on the prompt you’ll write in the empty field. Let me show you a few examples. 

How to phrase your prompts to GPT-3?

You should use the prompt to phrase your question AND give formatting instructions.

Not only can you ask GPT-3 to answer a specific question but your prompt can also contain instructions about the parameters and the structure of the output. 

Let’s for instance ask GPT-3 to describe what a great CRM is. 

Our first prompt will be: Describe a great CRM.

Here’s the answer given by GPT-3: 

A great CRM will have features that allow you to track and manage your customer relationships, including their contact information, communication history, and purchase history. It will also have features that allow you to automate tasks like email marketing and follow-ups, and it will provide insights that help you improve your customer relationships.

It’s spot-on, nothing wrong about it. But you might want to get a list of features which describe a great CRM. So let’s rephrase our prompt: List the features of a great CRM.

Here’s GPT-3’s answer: 

Some features of a great CRM software include the ability to track and manage customer interactions, sales, and marketing data; automate repetitive tasks; and generate reports. Other features may include a contact management system, a customer service management system, and integration with other business software applications.

It’s actually quite similar in terms of structure to the first iteration. What if you want to get a list of bullet points? Easy, simply rephrase your prompt: 

List the features of a great CRM, presented in bullet points.

Here’s the output: 

• it should be easy and intuitive to use;
• it should offer a variety of features to suit the needs of different businesses;
• it should be affordable;
• it should be able to integrate with other business software;
• it should have good customer support.

Perfect. Now what if you want to get a list of 15 features instead of 5? 

Just ask: List 15 features of a great CRM.

1. Ease of use
2. Accessibility
3. Efficient
4. User-friendly
5. Customizable
6. Affordable
7. Scalable
8. Mobile-friendly
9. Integrations
10. Reporting and analytics
11. Lead management
12. Pipeline management
13. Sales forecasting
14. Customer support
15. Training and onboarding

Note: if you’re expecting a long output, it’s important to set the maximum length to 4,000 tokens. The token allocation includes the length of your prompt. So if your instructions are longer, you’ll have less available capacity for the reply. 

Can you set the tone of voice in your instructions to GPT-3? 

Yes, you can. Not only can you give formatting instructions as shown earlier, you can also ask the model to adopt a specific tone of voice. 

Let’s for instance ask GPT-3 to describe  the USP of a CRM. 

List 5 phrases to describe the USP of a CRM

Here’s the output:

-A CRM can help you keep track of your customer’s interactions with your company
-A CRM can help you manage your sales pipeline
-A CRM can help you track your marketing campaigns
-A CRM can help you monitor your customer service
-A CRM can help you build stronger relationships with your customers 

It’s fine but you might want something punchier. Let’s rephrase the prompt: 

List 5 punchy phrases which can be used to describe the USP of a great CRM

1. Keep your customer relationships in check
2. All your customer data in one place
3. The ultimate customer management tool
4. Simplify your customer interactions
5. Get the most out of your customer relationships

The output evolved from a purely descriptive list to a set of marketing punchlines. 

Instead of “punchy phrases”, you could also use “slogans”. 

List 5 slogans which can be used to describe the USP of a great CRM.

1. Keep your customers happy and coming back for more
2. Keep your customers’ information at your fingertips
3. Get to know your customers better
4. Keep your customers engaged
5. Give your customers what they want

How to add custom parameters to a text generated with GPT-3? 

Sometimes you might want GPT-3 to output specific parameters, either because you want to use them as sales arguments or simply to avoid a random mistake.  

A good example is the foundation year of a company. Best Homes, a real estate company, was founded in 2015.

Let’s start with a prompt without the instruction.  

Describe why Best Homes is the perfect choice if you’re looking for a bespoke apartment.  

Best Homes was founded in 2006 with the aim of providing bespoke apartments that are designed to meet the specific needs of our clients. We work closely with our clients to ensure that their new home meets their every expectation, and our team of experienced architects and interior designers are on hand to create beautiful, unique spaces that are tailored to each individual. We believe that everyone deserves a home that is perfect for them, and we are dedicated to making that dream a reality for our clients.

I have underlined in the output the problematic section: GPT-3 randomly declared that Best Homes was created in 2006, but in reality it was founded in 2015.

To avoid this random mistake, let’s include a custom parameter in the prompt. 

Describe why Best Homes, founded in 2015, is the perfect choice if you’re looking for a bespoke apartment.

Best Homes was founded in 2015 with the aim of providing bespoke apartments that are designed to meet the specific needs of each individual client. The company has a team of experienced architects and interior designers who work closely with each client to ensure that their dream home becomes a reality.

Best Homes offers a wide range of services, from helping to find the perfect plot of land on which to build your home, to designing and constructing your bespoke apartment from scratch. The company also offers a full after-sales service, so you can be sure that your home will be maintained to the highest possible standard.

If you’re looking for a bespoke apartment that is designed specifically for you, then Best Homes is the perfect choice. Contact the company today to discuss your specific requirements.

Of course you might want to adapt the output, for instance if the real Best Homes doesn’t provide any maintenance services. 

👨‍💻 Can you generate code with GPT-3? 

Yes, you can. But here again the accuracy of your prompt will influence the quality of the output.

Let’s start with a basic prompt:

Generate the code for a signup form including an email field and a name field.

Here’s the code we get in reply:

<form>
  <div>
    <label>Email:</label>
    <input type="email" name="email">
  </div>
  <div>
    <label>Name:</label>
    <input type="text" name="name">
  </div>
  <div>
    <input type="submit" value="Submit">
  </div>
</form>

That’s what it looks like if tested on W3Schools.

Let’s add a few instructions: 

Generate the code for a signup form including an email field and a name field, presented on 2 different rows. Add a margin of 10px between each row. Center the form on screen using css. Include in the code a webhook to collect the email.

Here’s the output.

<form action="https://hooks.zapier.com/hooks/catch/3/o9nixj/" method="POST">
  <div style="margin-bottom:10px; text-align:center">
  <label>Email:</label>
  <input type="email" name="email">
  </div>
  <div style="margin-bottom:10px; text-align:center">
  <label>Name:</label>
  <input type="text" name="name">
  </div>
  <div style="text-align:center">
  <input type="submit" value="Submit">
  </div>
</form>

That’s what it looks like when implemented on a website.

🎙 Can you create a virtual podcast with GPT-3?

Within the current length limitations of the model, yes. 

Here’s a prompt: 

Create a dialogue between two startup founders (Jack and Fred) discussing the future of technology.

And here’s the dialogue generated by the AI: 

Jack: What do you think about the future of technology, Fred?

Fred: I think it’s looking pretty bright. A lot of new advancements are being made every day.

Jack: Yeah, I agree. I think we’re on the cusp of some really big breakthroughs.

Fred: I’m especially excited about the potential of artificial intelligence.

Jack: Yeah, that’s definitely one area that is rapidly evolving. I think it’s going to change the world in a lot of ways.

Fred: I think you’re right. It’s going to be interesting to see how everything unfolds.

In conclusion

GPT-3 is a powerful and advanced text autocomplete model, based on 175 billion parameters. But the accuracy of its output largely depends on the quality of the prompts it receives, just as Google can’t deliver any actionable results if you don’t properly phrase your search query. 

Very often, when someone asks me something, I notice that they can’t find the information because they don’t know how to phrase the question. 

Maybe we should develop online courses on how to formulate your requests.  

Final word to GPT-3

Prompt: Why is it important to properly phrase your prompts when interacting with GPT-3 ?

GPT-3 is a natural language processing model that is designed to interpret and predict the next word in a sentence. As such, it is important to properly phrase your prompts in order to ensure that GPT-3 can accurately interpret your intended meaning.

And here’s another question I’ve just asked to GPT Chat, the brand new release of OpenAI:

Why is a language model not as good as a human writer?

A language model, such as GPT-3, is trained to generate text that is coherent and grammatically correct, but it is not capable of understanding the meaning or context of the words it uses. This means that a language model may not be able to produce high-quality content that is engaging, informative, or interesting to readers. A human writer, on the other hand, has the ability to understand and interpret the meaning of the words they use, and can use this understanding to create content that is compelling and engaging. Additionally, a human writer has the ability to draw on their own experiences, knowledge, and creativity to produce unique and original content. Therefore, a human writer is generally considered to be superior to a language model when it comes to creating high-quality content.

What’s your opinion? Share your thoughts in the comment section below.