A friend of mine told me that AI is not quite there yet. I asked why. She said the responses she received from ChatGPT are often too generic.
Similar to other text-based generative AI tools, ChatGPT is trained on a large-scale language model. This breadth has many advantages, but there's a catch—the quality and specificity of the input significantly affect the quality of the output.
In today's issue, I'll introduce a four-step framework to write better prompts.
I will use ChatGPT as an example, but it can be applied to a wide range of generative AI tools.
Note that I'll focus on front-end prompting today, as it is the simplest place to start. However, eventually, you'll want to learn back-end prompting to have finer control over the AI model's parameters. I'll talk about that in another issue.
What happens with a generic prompt?
Let’s say I am a product designer working on an AI Chatbot product for a home insurance company. I conducted user interviews and took some notes.
Now I want to summarize them into a simple table to share with other stakeholders.
Let’s say I want to ask ChatGPT to help me summarize.
This is what would happen if I provided a super generic prompt.
The summary is very hard for me to use.
So, below are 4 actionable steps that can come in handy.
Step 1. Give role.
By setting a role or perspective, you can frame ChatGPT’s response. Tell it to think like a particular professional or to adopt a specific viewpoint.
Example:
Imagine you're a product designer who conducted 5 user interviews and wrote down a lot of quotes.
Step 2. Provide context.
Establish the context of your request. Describe why you need to do this task and what you intend to achieve with it.
If you are not sure what to write, ask yourself three questions:
What the project is about (basic Info)?
What is the goal?
Who is the audience?
Example:
In the interviews you asked users to interact with an AI chatbot for a home insurance brokerage and share thoughts. You wrote down quotes from the interviews. Now you are summarizing user feedback. [Basic Info]
The goal is to share the summary with the stakeholders from different departments. It is for informational purposes only, without the need for immediate decision-making. [Goal]
The stakeholders only have basic knowledge of this project. [Audience]
Step 3. Outline a specific task.
Clarify your requirements and specify the details you're seeking.
Example:
Please provide a structured summary that captures key themes from the user feedback, focusing on areas of delight, friction, confusion, user recommendations, and other noteworthy comments.
Here are the notes:
"The AI chatbot was surprisingly intuitive, guiding me through finding the right insurance plan like a knowledgeable friend."
"I had to rephrase my questions a couple of times because the chatbot didn't understand me on the first try."
"It would be helpful if the chatbot could provide a summary comparison of the recommended plans, highlighting key differences."
"I loved the convenience of getting insurance recommendations at any time. The chatbot's availability is a game-changer."
"I was unsure how to proceed when the bot presented several options. Clearer direction would be helpful."
"The chatbot's personality was a nice touch, making it feel less robotic."
"The chatbot provided quick responses and concise information, which made the decision-making process smoother."
"The chatbot repeated the same message twice, which was a bit annoying."
"Adding user reviews or ratings for the insurance options presented would add more depth to the recommendations."
"I was impressed by the chatbot's ability to personalize recommendations based on my needs and preferences."
"Some of the insurance terminology used by the chatbot was confusing. Simplifying the language could help."
"The design of the chatbot was clean and appealing, enhancing the user experience."
"The AI chatbot saved me a lot of time compared to how long I would normally spend researching insurance options."
"The chatbot didn't recognize my input about a pre-existing condition, leading to irrelevant recommendations."
"Implementing a feature where the chatbot follows up with users to refine recommendations based on feedback would be beneficial.
Step 4. Specify the format.
Clearly communicate your desired format for the information presentation.
Example
Organize the summary into a table, with the key themes being the column names, and under each column, it is the categorizing the actionable insights and direct user quotes that illustrate each point.
Step 5. Ask ChatGPT.
Alright! Now is the time to put everything together and see the result!
🤩 Here is the result!
Happy with what I got. There are other advanced prompting techniques I could have used here but didn't think were necessary, such as providing examples and reverse prompting. I can even ask it to ask me questions to clarify my prompt.
Interested? I'll introduce those techniques another time. (I need to send this newsletter out before the kids wake up!)
Bonus Resources
[Youtube] I Discovered The Perfect ChatGPT Prompt Formula
[Youtube] Master the Perfect ChatGPT Prompt Formula (in just 8 minutes)
[Youtube] ChatGPT / GPT-4 System Prompt Engineering - Ultimate Guide
Thanks for reading. If you enjoyed this issue, please consider sharing this newsletter to a friend who might benefit from it.
Happy Tuesday!
—Xinran
Questions? Comments? Just reply to this newsletter.