You.com custom prompt presets

Problem:

One of the initial problems I worked on solving in the AI space was personalization. I built a feature that would capture a users preference-type information into what was equivalently custom instructions. That was then used to prime each conversation a user had with the AI. I noticed however, that users were regularly adding lots of additional pre-contextual information with many conversations for certain intents. This was cumbersome for users to be constantly copying and pasting templates they had created. OpenAI had already released a GPT creator, but this was also large and a heavy feature that took a lot of cognitive load to use.

Solution:

The product manager I worked with suggested we recreate the GPT-creator feature that OpenAI had, but I didn’t feel that would adequately solve our users pain points. I knew we could make this feature more consumer friendly, especially since our target audience often doesn’t get as deep into AI and tends to live in the more web-search type space. Instead, I felt we could create a light weight prompt-preset creator. We already had AI modes and it would be easier on our backend to add a small preset creator which meant we could launch it sooner. To test it out, I created a prototype and then I was able to combine the saved preset text with the input text before making any API calls. Below is the prototype I created.

Intent classification and multimodal AI at You.com.

Overview:

With the various utilities of AI, it’s often difficult for users to get the exact types of results they want. Often, the AI will give large responses when they aren’t necessary or when no AI is necessary at all. I worked with a product manager to come up with various intents that would fulfill most users goals. I felt that a relatively low token count could classify the users intent without increasing costs too much. I decided to prototype the idea for user testing. I also wanted to ensure this was theoretically feasible before presenting the idea to the software engineers and it was massively helpful in explaining the method.

Here’s how I did it:

  • I did some trial and error to dial in a perfect prompt template that could get reliable intent classification.
  • I captured multiple intents and then sent API requests to separate endpoints.
  • I had various templates built out where I could display the different types of data that came back.
  • I also build various modules that could be combined if multiple intents were determined, such as creating an image and text.

AI null states at You.com

User problem:

The AI space is new for many users and has many limitations. It’s difficult for them to understand how it can be used to provide immediate value to them.

Solution:

From various user testing I did, I found that 2 patterns that were guiding users to submit more prompts to the AI but even more, they understood better the value it could give them. The concept below is a null state design that users would see their first time landing on the chat page.

  • Users could gain various ideas for how they could use the AI.
  • Concepts would be reinforced with imagery where appropriate.
  • Suggestions would update regularly to be more personalized.
  • This concept also supported our monetization efforts by explaining various AI modes and how they could be used.

Improve You.com chain of thought

Problem:

At you.com we created some unique modes that can crawl the internet in real-time and do complex computational tasks. These have come with a few problems though.

  • Sometimes these modes can get stuck midway through a process which confuses users.
  • During the mode processing, users were seeing full RAW AI breakdown processes which are often confusing or not helpful.
  • The AI processing feels like it takes a long time.

Solution:

I designed various templates to represent different types of chain-of-thought processing steps and AI was taking and I made them much more user friendly. Each step would still show some real data that was being processed without forcing the user to see everything in a large confusing list. There is a certain portion of our audience that did still want to see the full breakdown so I moved that into a modal that could be triggered after the generation had completed. When testing this prototype with users, I got feedback that it made the AI chain-of-thought process feel much more interesting and faster. The final responses were more clearly differentiated from the thinking process too.