The becoming of UXfolio’s AI-powered UX case study generator

At UXfolio, AI came to us in an organic way. At first, our entire team was wary of adding AI-powered features to the product simply because it was the shiny new thing and everybody else was doing it for, seemingly, no good reason. But once we started tackling the question of our case study generator, we realized that AI can bring our users incredible value that nobody else offers yet.

From our continuous research since the inception of UXfolio, we always knew that the wow-moment in our product was the case study generator. The case study generator appears when a user adds a new case study to their portfolio. At this moment, a modal appears where the user can pick from an assortment of UX methods they used throughout the project they want to chronicle.

Seeing those UX methods laid out was the moment when most testers said, “wow, this is very cool”. This reaction was triggered by a mix of things:

• It was unexpected, since before this moment they were thinking that they’ll get a blank page where they can start writing or building. Exactly like most website builders do.
• Though this extra step was unexpected, this feeling was quickly followed by a sense of familiarity, due to the UX methods listed, that UXers are all familiar with (prototyping, user interviews, usability testing, surveys, etc.)
• This also provided a sense of accomplishment, since everybody could pick at least 2-3 methods from the list: “yes, I conducted interviews, I prototyped, I did surveys.”

After choosing the methods, we generated a case study draft, instead of an empty page. This draft contained sections filled with text ideas and guiding questions for the methods chosen by the user. This means that it was not simply a fun step, but it was also beneficial, since it combated the blank page issue, dreaded by so many creatives. 

For years, this case study generator worked for us; therefore, we never made any significant changes to it. We updated the icons, colors, and fonts when we updated our design system and such, but that was about it. But at one point, we heard enough constructive feedback and ideas to consider updating and modernizing the generator:

• Users loved the UX methods, but they were missing rhetorical sections, like introduction, learnings, next steps, and so on. These are the glue to the UX methods, and they are very much expected from a case study.
• Users were grateful for the text ideas and guiding questions in the generated case study, but they were expecting something more. Especially with the advent of AI. What’s more, a huge portion of users thought that the text ideas were already generated by AI, which wasn’t the case.

At the same time, our growth team’s Product Designer, Réka Nagy, was also constantly advocating for more personalization as an effective and well-known driver of user activation.

At one point in time, our case study generator was extremely innovative, but new technologies bring different attitudes and raise new expectations, which is a very natural process. Our job is to adapt before it was too late.

So, all the stars aligned:

• Users let us know what was missing from the generator.
• They also revealed that they expect more after the generation.
• We wanted to provide more personalization, in-line with product trends.
• And we had the technology to provide more personalization.

The scope was the following:

• The goal of the case study generator was to provide a case study draft that’s personalized enough to encourage our users to start working on it (activation).
• We had to keep this as a wow-moment, since it was at the perfect moment in the user journey, and it was our main wow-moment.

Based on all the above information, you’d think that it was a clean, straightforward task. But that’s far from the truth:

• This is our product’s wow-moment, and we cannot overcomplicate it. But personalization is only possible through asking for user input.
• Our team has never worked with AI before, so we had to learn everything on the go: prompt engineering, creating a scalable infrastructure that incorporates ChatGPT, our chosen LLM. These, however, were not as scary as the first point, since we have an extremely talented team of professionals.

We began a balancing act: adding just enough options, asking just enough input to provide something that’s satisfyingly personalized. We also had to balance input from continuous user testing and our own goals of activation.

3 steps instead of 1

Instead of a single step, we arrived to the conclusion that we needed three steps to provide a satisfying, personalized outcome with the case study generator:

1. Step 1 remained a multiple-choice window, where users can pick and choose from UX methods. However, we added rhetorical building blocks into the mix (Overview, Problem & solution, Learnings  & next steps, etc.), as per their requests. This was a safe decision, as we knew the effects of this screen; we just amended it with what our users wanted from us.
2. Step 2 provides the option for the users to upload a screen/screenshot from their project. We use this image for visual personalization: putting the image into a stunning mockup in the hero section of the generated case study.
3. Step 3 gathers the input needed for personalizing the content. In this step, we ask users to write a few sentences about the project. We’ve also included a prompt template as a crutch.

The technical part of the task was handled by our developer, Dániel Kincses. He was learning along the way and created something magical: an infrastructure for AI in UXfolio, not just for this single feature. Together with Réka, they’ve refined the system prompt to the point it can turn 2-3 vague sentences about a UX project into a personalized draft.

Breaking the prompt

First, we tested the feature as a team. We tried to break it, we tried to make it write inflammatory content, and so on. Meanwhile Réka and Dániel kept refining the system prompts based on our findings. Eventually, it started providing consistently great answers and so, the new case study generator was ready for release.

Right after, Réka started interviewing users, who had an overwhelming positive reaction. To sum it up, most of them could not believe that we generated all that content from the 2-3 sentences they provided.

From scepticism to total embrace

Now, I have a completely different outlook on the AI-features topic. Take for example the generic grammar fixer AI solutions that have become the norm in content-based products. The thing is, they’ve become the norm, because they bring value to users. They save time, clicks, brainpower, and embarrassment.

When such features become base-level in your niche, it’s useless, and maybe even stupid to resist them, since users get used to them being available. Your product not providing such features signals that you’re behind the curve, not hitting the base-level.

If you need help designing actually useful AI features, explore our AI design and research service.