PolyGlot AI
Vibe coding—or pair programming with an AI assistant—is here to stay.
My first experience with Cursor was… well, terrible. The very first prompts gave me clean, concise code, which lulled me into a false sense of confidence. I started making lots of changes, only to realize I couldn’t recover from the faulty logic it had produced. In the end, I rewrote the code myself—this time with some help from Claude (via its web app).
That experience left me wary. I stopped using Cursor for a long time. Meanwhile, I kept learning about LLM engineering, agentic AI, and other buzzwords being thrown around. Over time, I picked up a few tricks about prompts and context management.
When I eventually gave Cursor a second chance, it was a hesitant rendezvous. This time, I didn’t finish the project either—but unlike before, I made real progress. That felt like a win.
But then something else caught my attention: Claude Code. I decided to build an application with it.
My workflow had always been a patchwork of tools:
- For technical or coding tasks, I leaned on Claude.
- If Claude failed me, I turned to Google and used Gemini’s results.
- For everything else, I relied on ChatGPT.
- Occasionally, when I felt bold, I let ChatGPT handle code (though Claude usually came to the rescue when things broke).
At one point, I thought: What if I could talk to all three at once? What if they could each contribute, summarize their thoughts, and then I continue the conversation with that collective intelligence?
That was the seed of my iPhone app: PolyGlot AI.
Now, I’m not a mobile developer by training. I’ve only been learning iOS development on Udemy. I can navigate Xcode, I understand SwiftUI basics like HStack and VStack, but not nearly enough to build something as ambitious as PolyGlot AI without help.
So, I turned to Claude Code—not just for designing the UI, but also for writing the logic. This time, I took it slow. No more reckless vibe coding; just small, careful steps forward.
One of the fun parts was the back-and-forth with Claude. I often asked it to suggest different approaches, and it always obliged. To be polite, I usually agreed with what it recommended—but sometimes I nudged things in my own direction.
Of course, the journey wasn’t smooth. Claude occasionally defaulted to older models, or gave me incorrect API URLs. Maybe it could have fixed those issues if I had pressed harder, but I didn’t want to rely on it blindly. I double-checked the documentation myself, asked questions where needed, and patched the code. And yes—when I asked Claude directly to fix something, more often than not, it actually did.
What I Learned
- Claude Code is a great assistant, not a replacement. It shines when nudged in the right direction.
- My role is to provide the solutions, not just the code. The AI handles syntax; I handle clarity of thought.
- The future of programming might really be natural language. As someone once said, coding may eventually look more like conversation than typing symbols—and after this project, I believe that, at least to some extent.
What’s Next
- Multimodal Expansion: The next step is to add multimodal support—so text, images, maybe even speech can flow through the app.
- Smarter Coding Conversations: I’d like to refine how the app handles code-related prompts, letting it act not just as a “vote and summarize” system, but as a true collaborative partner for debugging and design.
PolyGlot AI started as a simple “what if” thought. Now, it feels like a glimpse into how we’ll all be coding—and thinking—tomorrow.
PolyGlot AI
