Reflections from building Mighty Veg — a vibe-coding experiment
I recently joined five designer-makers for a 4-day, AI-powered hack! Thanks for the invite, Tom Harman 😊
Working independently on our own projects, we had two goals:
- Use AI to build a product in 4 days.
- Share and learn along the way.
Introducing Mighty Veg
I made Mighty Veg, a mobile-optimised web app. It let’s people snap a photo of any recipe to be turned into a vegetarian version.
It exists to help meat reducers eat less meat. I often photograph and save recipes from the Guardian Feast magazine. I wondered if AI could convert meat-based recipes into veg-friendly dishes. I tested the idea with ChatGPT, and it worked.
Real-world problem, scratches an itch, simple interaction. I can use AI to help build the product and I can utilise AI in the product itself. Let’s go!
Main ways I used AI:
- Recipe conversion test — ChatGPT
- Copy — Co-created with ChatGPT
- Logo & graphics — Made by ChatGPT
- Code — Created by Replit
- Deployment — Handled by Replit
Reflections and notes
AI is superb at idea generation, but you’ll need a narrowing strategy
I’d already decided how Mighty Veg would work. But out of curiosity, I popped this prompt into ChatGPT.
I’m creating a mobile app that helps meat reducers to eat less meat. Generate a big list of ideas that might help meat reducers. Think of non-digital ideas, digital-led ideas, AI-inspired ideas, and any other type of idea you can think of. Be creative and divergent. Group ideas in a sensible way.
What came back? 48 ideas split across 10 categories (see below). A lot to process. You’ll need to think carefully about narrowing strategies, or a smarter initial prompt that returns a leaner list of ideas. Or a follow-up prompt that narrows the list. I tried this:
Please highlight the 5 that are most likely to gain adoption among meat reducers and list evidence-backed reasons why.
Check the chat for what AI suggested. To be fair, it does return a sensible list, backed by relevant behavioural models and academic studies.
Having lots of ideas is good. Just consider how AI is used to generate ideas, and more importantly, how human insight steers and makes the decisions.
AI claims text prompts are enough, but it really needs UX direction
Lovable and Replit make it sound so easy. Just enter a text prompt and your thing will get built. Magic, like rubbing a lamp!
My first prompt for Replit:
The app is called Mighty Veg. Its mission is to help people eat less meat and dairy products. It has only one main feature, allowing the user to take a photo of a recipe and then the app converts it into a vegan-friendly dish. The user can see a list of dishes they’ve saved and can remove the ones they don’t want.
Replit did an okay job. Covered the core interactions, and gave me something to nudge and build. We’re dealing with standard interaction patterns. Even so, it didn’t consider the overall user journey and finer details, even after a host prompts.
I wondered how the AI would react to UX direction in the form of good old-fashioned, hand-made wireframes (see below). It reacted well, and could see and understand what I wanted.
AI makes out it’s in control, but it means we can get lazy
This is linked to the reflection above. My initial text prompt generated all the code for the app. I know my way around front-end code and a bit of JavaScript. But I began to feel uneasy at not understanding the code.
In Co-Intelligence: Living and Working with AI, Ethan Mollick describes an experiment by Fabrizio Dell’Acqua:
When the AI is very good, humans have no reason to work hard and pay attention. They let the AI take over instead of using it as a tool, which can hurt human learning, skill development, and productivity. He called this “falling asleep at the wheel.”
Reflecting on this, I feel like I never even took control of the wheel. If we’re going to use tools to generate code, it’s important we at least have some understanding of how it works and what it’s doing.
AI is good at creating visuals, but it can’t determine good taste
I planned to play with Exactly and Visual Electric, but ChatGPT did a solid job creating a simple logo and loads of veggie graphics.
Below are some logo and graphics explorations, using source files as inspiration. I got permission from Carol Anne to use her Bremerton Farmers Market Logo, and from Jessica Lohman to use her Heirloom Tomatoes. The Brush Pen image from Pexels has a “free to use” license.
I’m impressed at what AI produced. However I’m not convinced AI can determine good taste, choose context-appropriate aesthetics or have a clear vision for a brand.
AI is always there for you, but doesn’t care about your wellbeing
By the morning of day 4, I was feeling tired and unappreciated. The problem is that AI is there for you 24/7. Always ready to move forward. It doesn’t pause, take stock, suggest we go for a walk, or have a cup of tea.
Can’t believe it took me 4 days to see that AI doesn’t care about wellbeing. Not like proper teammates. Working alone and plugging away with AI changes how we work. I’ve realised I need to be more disciplined than usual when it comes to self-care.
AI provides a productive team, but no chemistry
At times, I felt like I was rolling with my small team. Coder, visual designer, content creator. Me as the product strategist, knitting everything together.
But there was no real chemistry. No camaraderie. No sense of shared mission. No serendipitous side convos or satisfaction you get from solving a problem with your people.
I’m not sure what this means for working with AI, especially when working alone as a freelancer creator. I suspect that working in a small (real) team alongside AI tools will bring greater productivity and retain the chemistry.
AI aids production, but it isn’t going to craft a great product
AI is great at making things. Code, copy, graphics. But we still need human insight and ingenuity to create delightful, useful, and accessible products.
Sure, AI helped me build Mighty Veg fairly quickly. But in reality, it’s easy for people to substitute or omit non-veggie ingredients from recipes to make them veg-friendly. Plus, Mighty Veg hasn’t been well executed. It’s too slow at converting recipes and it doesn’t work any better than ChatGPT.
I can see the shortcomings of Mighty Veg. Testing would do the same. AI won’t reveal this level of insight and real-world understanding.
A clear benefit of using AI is shortening the development cycle. Getting something interactive, quickly, means we engage with the product and test a user experience as early as possible.
AI is changing how we work, but we need to learn alongside others
AI is moving fast. It’s having a profound impact on designing and making in the digital space. Keeping up is a challenge. I’m very much in learning and experimentation mode.
Working alongside others for 4 days provided structure, accountability, motivation, inspiration and a reflective space. Keen to do more of this.
For a list of the products built and to see who took part, see Tom’s excellent post: Proof of Vibes
Here’s our format:
- Week before — Share plans on a WhatsApp group
- Mon 10am — Meet online to say hello and share ideas and approaches
- Daily 10–10:30am — Meet to share progress and reflections
- Thursday 4:30pm — Demo our products and discuss learnings
- Throughout — General chit-chat and link-sharing on the WhatsApp
Finally, the app: Mighty Veg
If you’ve got down here, you might as well see the app. Works best on mobile. But be warned, it’s not that impressive. It might be slow to load, and even slower converting recipes!
🥦
