THE PROBLEM
Chat functionality relies on users prompting it with clear and specific questions or requests.
THE SOLUTION
Empower users with more control and assistance to craft and refine prompts to find information quicker.
THE PROCESS
Research, identify and leverage insights to build a prototype that improves the experience.
Constraining my Creativity.
To ensure my process and solution were robust and feasible, I applied several constraints to the project to ensure a successful outcome.
I also wanted to ensure my solution would blend seamlessly with ChatGPT's existing UI and feature set, so as to minimise design and development costs for a smoother and more sustainable integration.
Constraints:
Test only with existing users
Solution must be on brand
Easily implementable within existing design
Solution must be technically feasible
“We try to design systems that do not maximize for engagement. In fact, we’re so short on GPUs, the less people use our products, the better."
Sam Altman
CEO, OpenAI
Asking the Right Questions.
Considering Sam's vision and the stats, I sought to uncover:
Prompt behaviour
-
How are users phrasing their prompts to ChatGPT?
-
Does their behaviour differ depending on the task?
Perception and usage of AI
-
How do people perceive AI?
-
What AI tools do they use? Why?
Chat interactions
-
Do users start new conversations per query?
-
What behaviours do they exhibit during conversations?
Existing mental models
-
When searching for information, what mental models exist?
-
Why use ChatGPT over Google and vice-versa?
Methods Behind the Madness.
To help answer these questions, I used the following tools and methods:
.png)
Desk Research
Researched demographics, sustainability, GDPR, plugins, etc.
.png)
App Reviews
Analysed reviews on Reddit, Apple Store and Google Play.
.png)
User Interviews
Conducted interviews with 5 participants to uncover insights.
.png)
Usability Testing
Tested current app and my solution to improve the experience.
.png)
Product Versions
Compared v3.5 and v4.0 for insights on overall business strategy.
.png)
Affinity Mapping
Uncovered patterns and pain points in current app to address.
Connecting Dots with Data.
Analysing the data allowed me to identify gaps, barriers and behaviours in the current experience.
.png)
From Data to Decisions.
The affinity mapping process unveiled patterns that included:
01
Trust and reliability
Low levels of trust due to inability to verify, validate or view ChatGPT’s sources.
02
Poor search and findability
Users were frustrated that they could not easily search and find information within their own history.
Chosen Theme
03
Prompt control and accuracy
Satisfaction varied by person and topic. More assistance and control needed to help tailor responses.
04
Mental models: findability
ChatGPT is used for specific, niche or hard-to-find information or when users want quick, formulated results.
.png)
.png)
.png)
.png)
.png)
Pivot, PIVOT!
My project was initially titled centred around improving search functionality in ChatGPT3.5, with a possible solution of a vault of answers users could access to minimize cost and environmental impact.
Why Pivot?
Analysing insights uncovered, I considered OpenAI's overall strategy, partnerships and place in the market. Given GDPR issues and strategic partnerships with Microsoft (creator of Co-Pilot which already incorporates Bing search capabilities), I chose to focus on improving user control and prompting.
Goal Digger: Mapping Motivation with JTBD.
To better understand users and their goals, I used the following techniques and tools to more holistically understand users most pressing needs and their underlying motivators.
-
5 Whys
-
Job stories
-
Job statements
-
Job journeys
Chosen JTBD:
Generate content
Learn or gain knowledge
Find information
Verify information
Honing in on the problem.
To narrow my project scope, I decided to tackle the question...
How might we create an intuitive and customisable information discovery experience that reduces user effort and minimizes system strain?
Design Inspiration
Before ideating solutions, I sought inspiration and researched existing AI-integrated products, competitor solutions, and more general app and mobile solutions.
This process provided insights into applicable design patterns and helped me understand industry standards, which served as benchmarks for my solution.
What's in a prompt?
To design a holistic solution, I had to better understand prompt engineering and it's parameters. Here is an example from Google's Prompting Guide 101.
By analysing prompt engineering, I was able to incorporate it in into my solution, helping users get the most out of their conversations with ChatGPT.
Idea Evolution
After generating multiple ideas and solutions, I sought feedback on my initial sketches. Incorporating mentor feedback and considering my prioritisation matrix, I revised my designs and decided to proceed with two core features to address my HMW.
Core Features:
Prompt assistant
Prompt templates
.png)
Fine-Tuning the Solution.
Version 1
My initial wireframes included an additional side panel for users to control and edit their prompts easily.
Also included, template and persona features allowed users to select pre-defined or community submitted personas and templates.
Version 2
After feedback from mentors that the side panel idea would add too much cognitive load on the user and wasn't the best approach, I simplified my design to provide users assistance whilst building their prompt.
Final version (lo-fi)
After discussing ideas with a mentor, I chose to provide assistance post-prompt generation for a more seamless experience.
ChatGPT now asks clarifying questions to help users refine their prompts. This approach not only educates users but also increases the likelihood of them receiving a satisfying response from ChatGPT.
A Tale of Two Rounds (of testing).
Time came to get my solution in front of real testers. I conducted two rounds of moderated testing with 3 participants.
Three consistent issues emerged:
01
Identifying prompt assistant
02
Identifying prompt suggestions
03
Identifying templates
.png)
.png)
The Big Reveal.
Prompting just got easier.
With Prompt Wizard, you can easily fine-tune your prompt to receive the most relevant and personalised responses.
Simply turn on the toggle, enter your prompt and ChatGPT will ask a series of clarifying questions, based on memory and prompt engineering techniques, to provide you with tailored responses.
Benefit for OpenAI
Improvements to initial prompts will reduce need for follow up questions, thus reducing system strain and costs.
Benefit for OpenAI
Shortcuts will increase user satisfaction. They can help train the AI to anticipate user preferences, thus reducing costs.
Refine responses with shortcuts.
With Prompt Wizard enabled, you can access prompt shortcuts to refine ChatGPT's responses at the click of a button.
Shortcuts are generated based on the content provided and most likely user actions, as determined by ChatGPT.
Benefit for OpenAI
Increase in user satisfaction and retention by educating them on effective prompts resulting in reduction in strain and costs.
Prompt templates.
Simplify your experience by using Templates. Easily browse and customize templates to create precise and effective prompts. Save your most frequently used templates directly to your homepage for quick access.
This feature is perfect for both beginners and seasoned users, helping to craft clear and specific prompts to achieve desired responses.
Benefit for OpenAI
Shortcuts will increase user satisfaction and help train the AI to anticipate user preferences, thus reducing costs.
Prompt suggestions.
Generate intelligent follow-up prompts based on ChatGPT's responses. Suggestions help users seamlessly continue their conversations, ensuring deeper engagement and exploration of topics.
Suggestions help users seeking more detailed information by providing clarifying questions, or redirecting inquiries to the web for broader context, to elevate and enhance the experience.
Solution Evolution.
.png)
Lo-Fi: Prompt Assistant
Users felt they wouldn't use the additional control features and confused the new chat button with the prompt assistant.
Hi-Fi: Prompt Assistant
Users didn't associate the lightning bolt icon with assistance and one even reverted to the sidebar to find it.
.png)
.png)
Lo-Fi: Prompt Suggestions
Users didn't associate the lightbulb icon with help or suggestions and users assumed they should return to the lightning bolt again for help.
Hi-Fi: Prompt Suggestions
Users didn't immediately assume the lightning bolt would also help them again here and still reverted to the big lightning bolt button for help again.
.png)
Hindsight is 20/20.
I really enjoyed challenging myself and designing within the constraints of this project whilst also working full-time and attending classes for my Master's in UX at IADT.
The opportunity to get feedback and guidance at each step of the project from my mentor, and other UX Tree mentors, was invaluable. I improved my ability to defend my design decisions and consider factors that I might not have otherwise, such as GDPR considerations. This ultimately led to a more robust project overall.
Embrace new approaches
Using JTBD for the first time was daunting, but I learned the value of deeply understanding user motivations and desires before jumping to ideation.
Don't go down the rabbit hole
Keeping track of AI advancements was challenging but resulted in a fully iterative process. I was constantly integrating and analysing information.
AI mental models
Mental models on AI features are still in their infancy. Users don't have strong connections to AI yet and so need assistance to build models.