OpenAI has just revealed a huge update to its ChatGPT application. Instead of being limited to its training data, ChatGPT can now browse the internet. While this update is only available to paid users through ChatGPT Plus, it’s going to have major implications on what the technology is capable of doing.
With access to so much information, it’s even more important that the prompts we’re giving ChatGPT are detailed enough and specific enough to give us the results we’re looking for. To get to grips with prompts, you don’t need technical know-how. Instead, it’s just worth knowing the basics.
Let’s start with the term ‘prompt’. What does this actually mean?
In essence, the term ‘prompt’ refers to the input you type into ChatGPT to get it to generate a response. This could be a question or a set of instructions. Thanks to the latest developments in GPT-4, this could also include an image. Once you click enter, this input then ‘prompts’ the application to produce a relevant output.
“A prompt is something we use with large language models, like ChatGPT, to define the goals and boundaries of its interaction with users. You can think of a prompt as the brief or the set of instructions given to an interviewer. It tells them what the purpose of the interview is and how to engage, however they still have the freedom to choose their own questions based on the responses they get.
A prompt might give ChatGPT a persona to embody or a style of speech to follow. You might tell it through a prompt that you want it to discover certain things and to totally avoid other things. A large language model can then do what it does i.e. generate the relevant response, all within the limitations and objectives set out by the prompt.” shared our Head of Platforms, Gavin Harcourt.
Because ChatGPT is designed to work with natural language, anyone can write a prompt. You don’t need to know how to code or how to script. Instead, you can interact with the model using the simple language we all speak on a daily basis.
“One of the best things about prompts is that they can be written in natural language - no coding knowledge needed. So, even my 12 year-old sister could design a prompt. This means that a lot of things that were once very difficult to do and had a coding barrier are now much more accessible.” said our Product Manager of Machine Learning, Shafeeq Shahajan.
So what are the different prompt types and how do they work?
As we touched on earlier, a prompt could be a question you want ChatGPT to answer or a set of instructions you want the model to follow. Depending on the level of context you add to your prompt, it will fall into one of these three tiers.
A zero-shot prompt is the simplest type, as it only provides a description of the task.
“Write a short hypothesis about the meaning of the term ‘healthy’ to ice cream consumers.”
A one-shot prompt provides one example alongside the instruction that the model can then use to learn how to best complete a task.
“Provide a list of the top competitors in the healthy ice cream space. 1. Halo.”
A few-shot prompt provides multiple examples and this can be useful when you want the model to follow a certain pattern.
“This is a conversation between an interviewer and a healthy ice cream consumer. Please continue the conversation.
Interviewer: Why did you decide to choose this particular ice cream?
Consumer: Because I wanted to pick a healthy ice cream that I wouldn’t feel guilty eating.
Interviewer: What do you consider healthy about this ice cream?
Consumer: It’s low in calories and it uses natural ingredients instead of artificial flavours.
Interviewer:”
As you can see from the examples above, prompts can be tailored to any use case, including market research.
“You can use prompts to design surveys and chatbot interactions. You can use prompts to turn problem statements into research briefs, which is something we’ve been experimenting with here at Streetbees. You can also create prompts to validate consumer submissions in real time. Instead of waiting for a consumer’s survey submission to be approved manually, for example, we can now approve submissions in real time by pairing a specific prompt with the GPT technology.” said Shafeeq.
In the initial phases of your market research, you can also use prompts to gather publicly available information about your audience and your competitors. Although you would then need to validate this with actual data through market research, it’s a great starting point to help you frame your business problem and your research needs.
To allow a tool like ChatGPT to help you do your job better, you first need to help it do its job better. You can do this by giving it a clear and detailed prompt with enough context to help it generate a relevant output.
So how can you make sure your prompt is up to scratch?
- Be as specific as possible. Think about exactly what you want the model to generate and include as much context as you can to make this clear e.g. “write about what the term ‘healthy’ means” vs. “write a 300-word hypothesis about what the term ‘healthy’ means to ice cream consumers between the ages of 18-24’.
- Include helpful information. If you want the model to write a short bio for you, but you only give it your name, the chances are it won’t come up with much or it will mistake you with someone else. Instead, you should give it the important information you want it to include, so it has something to reference.
- Provide examples. By including relevant examples in your prompt, you can help the model understand what type of response you’re looking for. This is particularly useful when you want the model to follow a certain pattern or to generate a response in a specific format.
To summarise, a prompt:
- Refers to the input you type into a large language model, like ChatGPT, to get it to generate a response. This could be a question or a set of instructions. Thanks to the latest developments in GPT-4, this could also include an image.
- Is designed to be written in natural language, so anyone can create a prompt. You don’t need to know how to code or how to script. Instead, you can interact with the model using the simple language we all speak on a daily basis.
- Provides the best results when you’re specific about what you want the model to do, when you include helpful information it can directly reference and when you provide clear examples of the pattern or the format you want it to follow.
AI and Machine Learning is at the heart of what we do here at Streetbees and the introduction of these new large language models will allow us to get even closer to consumers and their everyday choices. We’re currently exploring new ways to incorporate the power of ChatGPT into our solution through prompts and our pilots are coming on leaps and bounds.
We’ll continue to keep you posted on these exciting new developments to bring you research that’s quicker, more accurate and more actionable. If you have any questions about prompts and how they can be tailored to market research, get in touch!
