Vibe designing w/ Lovable and ChatGPT as my CTO

Aleksandra is a senior product designer in the banking world, where she spends her days solving complex design challenges. Outside of work, she likes to experiment with new tools and ideas, using AI as a creative partner rather than just a shortcut. In this episode, she talks about how she built an app to help women align their workouts with their natural energy cycles - a project born from her own experience of wanting to design with, not against, her body’s rhythm.

In this article, we highlight our biggest take aways from this chat.

---

What led you to start experimenting with AI tools outside of work?

I had been reading about new AI tools and wanted to experiment with them hands on instead of just watching from the sidelines. So I went actively looking for a real problem in my own life that I could build something around. After noticing how my energy levels and gym performance fluctuated with my menstrual cycle, I landed on the idea of cycle based workout tracking. “some months I can do a hundred-kilo hip thrust, and another part of the month I can’t even lift fifty… I feel dead, I feel drained.” I tried to find existing apps that did this but wasn’t satisfied with anything that helped me work with my natural rhythm rather than against it.

So, I decided to make something for myself - a tool that would help me align my workouts with my natural energy levels, and use this project as a sandbox to test AI tools I had been curious about.

I started by researching which ones could actually help me build a real mobile app and, ChatGPT, a subscription I already had, suggested Lovable. That became my base. From there I kept ChatGPT at the center as my CTO, and used Claude, Figma, and Supabase together, calling the whole process “vibe coding” as my way to learn by doing, testing, and playing.

---

You used ChatGPT, Claude, Lovable, Figma, and Supabase. How did each tool fit into your workflow?

Each tool was used for a different layer of the workflow, almost like having a small AI team. Here’s how I used each of them:

• ChatGPT, my main collaborator: I used it to frame the vision of the app, define jobs to be done, write user stories, catch edge cases, and also to scope and limit the MVP. I gave it detailed context upfront: things like what tech stack I was planning to use (“Lovable + Tailwind”) and what kind of app it was (“a mobile app for iOS that helps people align workouts with their menstrual cycle”).
• Figma, my visual playground: I collected a lot of nice benchmarks and then spent hours designing from scratch, which felt refreshing since in my day job I mostly work within a rigid design system that was set a while ago.
• Claude, my UX writing and design critique partner: I had heard Claude was very good at UX writing, so I used it to craft in-app copy and push notifications, and also asked it to critique versions of my design, giving it the goal first, example: “this is a calendar to tell a girl when she can push hard and when her energy might go down.”
• Lovable, the app builder: I chose it because ChatGPT told me it’s the best if you actually want to have it on the phone and have it as a real app. It let me generate working mobile code but often needs manual correction, in which case I would refer to GPT again.
• Supabase, the database layer: I didn’t know much about, but it integrates very well with Lovable, so I used it to store app data.

Overall, I stayed in the driver’s seat and used AI to accelerate my ideas.

---

You compared prompting to asking a designer for feedback. Can you talk more about that analogy?

Prompting an AI is no different from asking another designer for feedback. When you ask a colleague for a critique, you never just drop the screens on their desk. You always give context: what the project is about, what stage it’s in, what the goal is, and what kind of feedback you actually want. You might say, “Ignore the copy, focus on the information architecture or the visuals or the interactions.”

With my experiments, I realized that AI tools work the same way. The more context you provide, the better the output. If you just give them a vague command, they can’t give you meaningful results. But if you set the stage, the AI becomes a useful design partner. This is something people will learn the hard way, like I did, because prompting is not magic. It’s a skill that improves with practice and iteration. The more back-and-forth you have, the more the AI starts to act like a real teammate.

---

You mentioned how frustrating it was when Lovable didn’t follow your design exactly. What did you learn from that?

It was one of the most frustrating parts of the whole process. I spent a lot of time trying to get Lovable to copy my Figma design a hundred percent, writing what I thought was the perfect initial prompt. I even used ChatGPT and Claude together to help craft that perfect prompt, thinking that if I got the language right, Lovable would finally replicate her design pixel for pixel. But no matter how many times I tried, it never came out exactly as I wanted.

It taught me that AI tools still have limitations, and that even the best prompt can’t fully replace manual design refinement. Lovable is great for some initial prototyping, but unacceptable for launch at the moment. Building with AI isn’t about perfection, it’s about iteration. You have to be ready to prompt, test, correct, and even dive into the code yourself when the AI falls short. That patience was part of learning how to vibe code.

---

How did this project change your relationship with the product development process?

Working with Lovable gave me a real sense of what tech capacity actually means inside a company. While building I kept running into tiny issues like colors not updating properly across different states, or spacing and padding not behaving as expected. Each time I wanted to fix something, I had to decide whether it was worth spending credits and time on, or whether I should just move on to building and testing actual features.

That constant tradeoff between perfecting visual details and validating the core idea made me realize how engineering teams think about prioritization. You have to trade off what’s going to validate your hypothesis faster, and what’s a secondary priority you fix later on. There are always limited resources, competing priorities, and a need to focus on what matters most for learning or progress, not perfection.

---

What advice would you give other designers who want to start building with AI tools?

Approach AI playfully. It’s so much fun, you can just spend a whole weekend building and fixing and tweaking. Vibe coding isn’t about getting flawless results, it’s about giving yourself permission to build, test, and play. The key is to treat AI like a collaborator, not a command line. When you’re prompting, don’t just throw instructions, set the stage the same way you would for another designer. Give it context, goals, constraints, and what’s just a placeholder.

People often expect AI to “just do it,” but you learn (sometimes the hard way) that prompting is a skill you build over time. The more you do prompting, the better it becomes. AI tools are still messy, and you’ll spend time debugging, rewriting, and fixing code manually. But that’s normal, it’s how you learn how they actually work.