Teaching students how to prototype with AI tools instead of Figma revealed exciting possibilities and uncomfortable trade-offs. It also offered valuable insight into how non-tech people can create apps with generative AI.

I teach digital prototyping to fashion design undergraduates and master’s students at UPM’s Fashion School. My course focuses on exploring app concepts that can enrich the fashion experience. Past project ideas ranged from aesthetic affinity social communities (and dating) and live-action fashion-branded games to smart garment upcycling, VR fashion recommendations, and even a shared laundry app for roommates.

In previous semesters, I asked my students to prototype in Figma and standard design software. This past term, however, we switched to an entirely AI-based toolset, including Google Stitch, Google AI Studio, and Lovable.

The main reason for this change was, as Philip van Allen anticipated in his pre-LLM paper, that new AI design tools (such as Lovable or Bolt) would be much better suited to explore dynamic, interactive behaviors and flows than classic static screen-based design tools like Figma (in its 2025 version). Precious Ogar discusses the approach in more detail in his article AI-Focused Prototyping: A Practical Guide for UI/UX Designers.

The impact of this shift went beyond what I had anticipated and brought a new set of challenges. For starters, generative AI changed the very notion of prototyping: we were no longer building interconnected mockups to simulate an experience, but creating dynamic, functional applications that felt much closer to real products.

More importantly, it also changed how students practiced design skills and thought about product design — potentially for the worse. Greater reliance on tools meant less time spent conceptualizing. Endless prompting and refining replaced the hands-on practice required to master foundational design skills. And those skills are still highly relevant to creating production-quality experiences.

Yet, the final outcome gave my students the genuine confidence that they could build AI-enabled apps themselves — something I find deeply satisfying, and believe to be true.

In the end, the experience convinced me that prototyping with this new generation of tools demands a new kind of design literacy — one that puts more emphasis on thinking about our design than on the mechanics of building it.

Below are some of the key lessons I learned while teaching AI prototyping to fashion design students. They are presented in no particular order and vary in scope and relevance, but I believe each offers valuable insights for educators and designers.

What are we building?

Traditionally, a prototype was just a set of interactive mockups — a quick, low-effort way to test a concept without building an MVP or real app.

But GenAI has blurred that boundary. Today, anybody can create a partially working app in roughly the same time. And because these apps are made up of actual JavaScript and HTML code, they can include real features like live user authentication, persistent storage, and support all types of business logic. In other words, you are creating a conventional web app instead of a classic prototype.

But this impressive capability created a conceptual trap for my students. Many assumed the apps they were creating would inherently possess the same ‘magical’ AI capabilities as the tools used to build them. Some even saw these tools as oracles, as Olmanson et al. observed in their research with design students.

They expected their prototypes to be able to recognize images, perform semantic searches, or generate videos out of the box. They seemingly ignored that GenAI tools generate conventional web apps. Adding true AI functionality required significant setup work that was well beyond our class’s scope

I had to address this misconception directly in the classroom. I needed students to understand what they were building and where its limits were. When their concepts demanded complex AI capabilities, I guided them back to the fundamentals of prototyping: simulate the experience using mockups, predefined states, and pre-generated media.

Even in the age of AI, the golden rule of prototyping still holds: fake it, don’t make it.

The product vision is yours

Inevitably, most students began prototyping with high-level, generic prompts. And they were quickly disappointed by the results. It wasn’t just the visuals or interactions, but the product itself that didn’t match their expectations.

When I asked students to describe the product they wanted to build, it soon became clear that most couldn’t picture it themselves. So I asked them to develop a clearer vision of what they wanted to build. Starting from the home page, write down what information the product would show to the user, and what it would allow them to do. And then, determine the top sections of the app and repeat the same exercise for each.

Once they worked that out, they made much faster progress because they could communicate their intent much more clearly.

The conceptual model is yours too

This leads directly to the next challenge. When it came to specifying detailed app behavior and flows, a recurring challenge was mental model misalignment. Even when students had a clearer idea of how their product should work, AI tools frequently generated something completely different from their intent. Those alternatives could be perfectly reasonable. But they just weren’t right for the product the students were trying to build.

Help came when we discussed how to specify high-level design aspects for their apps. I introduced them to mental models and conceptual models. In a couple of cases, we applied very rudimentary conceptual grammars (mapping out objects, attributes, actions, and relationships). These allowed to map the users’ mental models into their apps more effectively and helped the students to prompt more confidently the detailed behaviour they expected from their products.

This lesson also extends to professional designers as well. Generative AI excels at providing average solutions for conventional problems. But if you are designing something complex, novel, or off the beaten path, AI will likely default to standard patterns and miss your intent. To design effectively with AI, having a clear conceptual model is non-negotiable.

Prompting can be intimidating

Because everyone uses apps daily, at first, my students assumed that describing one to an AI tool would be easy. It seemed almost too simple. But after their first few rounds of prompting, that confidence quickly gave way to frustration. The results often fell short of their expectations, and in some cases the outputs were completely different from what they had intended. A similar perception was observed by Olmanson et al. in their research work.

To help, I suggested using LLMs to refine their prompts before feeding them into the prototyping tools. This worked well in some cases. However, a few students ended up endlessly polishing prompts in ChatGPT or Gemini instead of testing them in the actual vibe-coding tool. For some, prompt refinement became a way to avoid confronting the prototype itself.

At that point, I had to remind them that iteration is about improving and evaluating the prototype, the tool’s output, not endlessly perfecting the prompt.

Design vocabulary is essential

One thing that consistently improved the quality of the results was using standard design vocabulary. When students described interfaces in everyday language, the outputs were often mediocre or inconsistent. Once they started using established design terms — such as button group, navigation menu, card, accordion, modal dialog, or splash screen — the AI became much more precise. The same was true for common UI behaviour terms like transition, persistence, overlay, scroll, and animation. And in certain cases, some basic software development vocabulary also proved useful.

This makes a lot of sense since the internal currency of AI tools is language. The more precise our vocabulary, the better they can understand what elements make up an interface and how those elements should behave and respond to user interactions.

Azamnouri et al. (2026) set up a similar AI educational project and also found that their students needed to acquire certain technical capabilities to successfully prototype their app ideas.

English produced better results

Our classes took place in Madrid, and most students naturally wrote their prompts in Spanish. While this was not in any way a controlled study, we noticed that prompts written in English often appeared to produce more accurate results and soon became a practical consideration in class.

This may reflect an implicit bias in AI tools, partly because so much of the material used to train them is in English. And if we see this effect with Spanish, one of the world’s most widely spoken languages, the gap could be even wider for languages with far fewer online resources, such as Catalan or Macedonian. Language support raises a real inclusive design challenge for AI tools.

The value of element targeting

This might seem a minor point, but it offers a very practical lesson. Students often struggled to control the scope of their refinements. The tools would frequently apply modifications well beyond what the students had intended. This reflects how prompting, by its own descriptive nature, doesn’t make it easy to define a precise scope.

Luckily, most modern AI design tools allow you to select a specific on-screen element -a card, button, image, or section- as the target for your prompt.

That small interaction made a big difference. By narrowing the scope before prompting, students got more accurate results and ran into far fewer headaches. It also reinforced an important principle of working with AI: the clearer the context, the better the outcome.

Pricing is an educational problem

Pricing was an important constraint. While preparing the course, I contacted several AI companies, including Lovable, Base44, and Claude, to ask about academic pricing. There were no class-specific plans available, and their standard plans were both expensive and unpredictable for a course where students need to explore freely. Unfortunately, under these conditions, my university denied funding.

This is a serious limitation. Exploration is not an optional part of design education. Students need room to try ideas, make mistakes, discard approaches, and start again. Credit-based pricing in educational settings clearly penalizes that type of experimentation.

Google’s vibe suite

As a result, we ended up using Google’s AI design products extensively, as they offered the most generous and comprehensive free plans. We were going to use Google Stitch anyway. It is Google’s vibe design tool and is particularly valuable for the broad exploration of both visual and interaction concepts, which are essential in the early stages of rapid prototyping. Once a clear direction emerged, students moved on to Google AI Studio, a proper vibe coding tool, to build interactive apps and flesh out their ideas.

On the downside, we did run into some significant performance issues, particularly with Google AI Studio. Our classes took place in the evenings in Madrid (CET), which coincides with morning and central business hours in the US. I can only speculate that we were hitting a period of peak demand for these tools.

Vibe coding is a QA nightmare

Prototyping with AI tools pushed students to adopt a tactic more commonly associated with software development: regression testing.

Modifying a feature in Google AI Studio often alters another part of the app that has already been completed. A new addition could change a layout, break a flow, remove a feature, or modify something that did not seem connected with the prompt. After a few of these incidents, students learned that every important change required a broader review. They had to test the whole app again, not only the feature they had just modified. Also, version bookmarking and history were very helpful in handling some catastrophic updates smoothly.

While code dependencies can partly explain this behaviour, sometimes it seemed quite random. Either way, this is currently the rather frustrating reality of vibe coding that app creators have to deal with.

Is AI prototyping slower?

This was the most shocking outcome of the course. Overall, it took my students slightly longer to create a prototype with AI tools than with Figma.

The reason probably lies in the fact that they were actually building something closer to a complete application rather than an app façade with some interactions, as is typically the case with traditional prototyping.

While vibe coding tools can produce incredible results in the matter of minutes, the process of creating a fully functional prototype clearly involves more decision-making, more debugging, and more testing. It also required students to develop a deeper understanding of the product they were trying to prototype.

AI and designer deskilling

By far, the most concerning lesson was related to the educational experience. Students using AI tools were able to create richer, more complete experiences. They explored more ambitious product concepts and even built working apps. Things that simply weren’t possible using Figma alone.

However, this also meant fewer opportunities to practise core UI design skills like arranging layouts, establishing visual hierarchy by means of size, color, and spacing, or grouping elements by applying the Gestalt principles. In comparison, my previous student cohorts, who used conventional prototyping tools, undoubtedly developed stronger competences in these key areas.

Perhaps, as Lee et al. (2025) evidenced in their work, relying on generative AI tools reduces users’ cognitive processes and critical thinking. In the end, this effect may contribute to achieving lower skill levels. Similar concerns were expressed by Slava Polonski in his article “AI skill atrophy: Are we forgetting how to think?”

There is an uncomfortable paradox here. We are currently witnessing how the product designer role is shifting towards a higher level of involvement in the design process, with practitioners potentially being displaced from more tedious UI work. Yet to effectively supervise a design, designers still need a strong grasp of the fundamentals to identify and correct any problems in AI-generated outputs, whether they are conceptual, visual, or related to interaction.

On the other hand, we are facing the possibility of degrading our students’ core design skills. These have traditionally been acquired through hands-on practice with tools such as Figma. The temptation now is to divert curricular time towards AI-based design practice. This creates a real risk of deskilling the very people who will soon be called upon to provide oversight of AI-generated output.

There is a clear implication for design education. We cannot replace core skill learning with AI-based tool practice. But neither can we disregard AI-based design practice because of its enormous potential. Our challenge is to integrate both types of learning into the curriculum to prepare our students for their future role as AI designers.

Final thoughts

Replacing Figma with AI tools turned out to be about more than just speeding up prototyping. It changed the nature of what students were actually building. Instead of creating only the façade of an experience, they were able to build functional applications that could look surprisingly close to real products. A key takeaway was that students need to understand what AI tools are actually creating, how and when to incorporate AI-related features, and when a convincing simulation is a better choice.

The students who made the most progress were the ones who developed a clear product vision and some form of a conceptual model. Being able to explain what their product should show and how it should behave turned them into more effective builders. Design vocabulary, prompting in English, targeting specific elements, and limited refinement of prompts in LLMs proved useful in most cases.

At the same time, this experience also revealed a less attractive side to vibe coding. Building a functional prototype requires more decisions, testing, and debugging than creating a conventional interactive mockup. That often means more effort and time. Credit-based pricing and unreliable performance also limited the exploratory work that our course relies on. And because AI tools handle much of the detailed UI work, students spent less time practising the visual and interaction-design fundamentals they may need in the future.

This is the core takeaway for design educators. We need to preserve the hands-on practice through which students develop design fundamentals, while teaching them how to use AI tools responsibly and effectively. We cannot treat these as competing approaches: students need both.

Despite all these drawbacks, I found the final outcome genuinely encouraging. My students left the course feeling capable of turning a product idea into an MVP app by themselves. I believe that confidence, colored by the understanding of the possibilities and limitations of these tools, was the most valuable thing they gained.

Resources

• Azamnouri, A., Hörauf, D., Schönberger, B., Bogner, J., & Wagner, S. (2026). Teaching AI competencies: Experiences from coaching interdisciplinary teams to develop AI prototypes.
• Lee, H.-P., Sarkar, A., Tankelevitch, L., Drosos, I., Rintel, S., Banks, R., & Wilson, N. (2025). The Impact of Generative AI on Critical Thinking: Self-Reported Reductions in Cognitive Effort and Confidence Effects From a Survey of Knowledge Workers.
• Nielsen Norman Group (2025). AI Prototyping in Real Design Contexts.
• Olmanson, J., Hassani, A., & Larsen, G. (2025). Learning educational technology prototyping with generative AI.
• Ogar, P. (2025). AI-Focused Prototyping: A Practical Guide for UI/UX Designers.
• Polonski, Slava (2026). AI skill atrophy: Are we forgetting how to think?
• Van Allen, P. (2018). Ways of prototyping AI. Interactions.

What I learned teaching AI prototyping at university was originally published in UX Collective on Medium, where people are continuing the conversation by highlighting and responding to this story.