[Demo] Complexity in disguise: Crafting experiences for generative AI features
Summary
AI tools like ChatGPT have exploded in popularity with good reason: they allow users to draft, summarize, and edit content with unprecedented speed. While these generic tools can generate any type of content or perform any type of content task, the user needs to craft an effective prompt to get high-quality output, and often needs to exchange multiple messages with additional guidance and requirements in order to improve results. When you’re building an AI-powered text generation feature, such as a product description or email writer, you typically can’t expect users to craft their own prompts. And unless you’re building a chat interface, you’re unlikely to offer the ability to iteratively improve the output. Instead, your feature needs a robust prompt skeleton that combines with user input to produce high-quality output in a single response. For the designer, this means building an interface that helps users provide the exact information that creates a successful prompt. This process is more complex than simple form design or a mad-lib prompt completion tool. The user input, often including free form text fields, might be required to fill in prompt variables, but it also could change the prompt structure itself, or even override base instructions. The effectiveness of the user input significantly influences the quality of the output, underscoring the need for designers to be deeply familiar with the backend prompt architecture so they can design the frontend. Drawing on recent text generation projects, I'll demonstrate how the interface design can respond to and evolve with the prompt architecture. I’ll talk about how to determine which prompt components to make invisible to the user, which to provide as predefined options, and which should be authored by the user in free-form text fields. Takeaways How prompt structure can impact user interface design and conversely, how design can impact prompt structure Techniques to provide effective user guidance within AI generation contexts to ensure consistently high-quality output Real-world examples and learnings from recent generative AI projects in an e-commerce software product
Key Insights
-
•
Building specialized AI features requires predefining most of the prompt to guide the LLM effectively rather than having users write their own prompts.
-
•
Users refine prompts iteratively in chat interfaces, but specialized tools often allow only one shot with limited input variables.
-
•
Understanding the backend prompt structure is critical for designers to decide which parts are fixed and which are configurable by users.
-
•
Tone of voice is a complex variable that significantly impacts product description differentiation and brand personality.
-
•
Free-text inputs for tone proved difficult for users; predefined tone options simplify user choices while maintaining quality.
-
•
AI models inconsistently detect tone from existing text samples and report high confidence in contradictory results, making this approach unreliable.
-
•
Successful tone options must be sufficiently distinct to allow users to easily identify and select the best fit for their brand.
-
•
Embedding detailed tone attribute descriptions (e.g., vocabulary, pronouns, punctuation) in prompts improves AI output quality.
-
•
The default confident and positive tone of models like ChatGPT may not suit all contexts and needs to be explicitly adjusted in prompts.
-
•
Providing a minimal, intuitive UI that maps to rich, complex backend instructions balances user simplicity with output quality.
Notable Quotes
"In a chat interface, you’re able to fill in the information that you missed and iterate on the output as it’s being generated."
"When you’re building a specific AI feature, you don’t want to make your user write their own prompt."
"It becomes the designer’s responsibility to figure out what aspects of the instructions to the LLM should be configurable and what should be fixed."
"A prompt is just a set of instructions to an AI in the context of text generation."
"People get stuck during the sessions when asked to describe their brand voice in just one or two words."
"The LLM gave many different answers for the same passage, each time reporting it was one hundred percent confident."
"Tones need to be distinct enough so that merchants can spot the tone that best fits their brand."
"Even though we’ve got a single dropdown field, it maps to a far richer set of instructions in the prompt."
"The confident tone built into these tools may be one of the more problematic features of Gen AI."
"By default, you’re getting I am very confident of my answers."
Or choose a question:
More Videos
"Relinquishing control over implementation empowered community-driven solutions and fostered lasting trust."
Alexia Cohen Adriane AckermanIncreasing Health Equity and Improving the Service Experience for Under-Served Latine Communities in Arizona
December 4, 2024
"Building in interruptions exponentially increases the accuracy of the findings."
Marc Majers Tony TurnerInterrupted UX - Add A Dose of Reality To Usability Testing
March 11, 2022
"Design thinking has helped raise awareness but often lacks the depth of design practice which requires dedicated craft."
Jorge ArangoDesign as an Antidote to VUCA
May 9, 2019
"Design your processes around learning. Celebrate especially when you fail—you often learn the most then."
Brenna FallonLearning Over Outcomes
October 24, 2019
"We as humans value warmth information in others more than competence information."
Daniel GloydWarming the User Experience: Lessons from America's first and most radical human-centered designers
May 9, 2024
"Accessibility is a journey and a conversation—thanks for starting yours by being here today."
Sam ProulxOnline Shopping: Designing an Accessible Experience
October 3, 2023
"Learning should be an experience tailored around the learner, not the curriculum."
Kristin SkinnerFive Years of DesignOps
September 29, 2021
"I appreciate your allyship when I'm in the room, but I appreciate it more when I'm not in the room."
Denise Jacobs Nancy Douyon Renee Reid Lisa WelchmanInteractive Keynote: Social Change by Design
January 8, 2024
"If you're not doing recruitment in-house and have to outsource three different agencies, managing timelines becomes a nightmare."
Roberta Dombrowski Lianna Aduana5 Reasons to Bring your Recruiting in House
September 30, 2021
Latest Books All books
Dig deeper with the Rosenbot
How does Rally’s research infrastructure facilitate participant management and integration with enterprise data systems?
What are the risks of democratizing user research and how can quality be maintained?
What does the acquisition of User Interviews by UserTesting mean for participants and research ops teams?