Can you Vibe Code?

Vibe coding is a new approach to building software that uses AI to turn plain-language prompts into working code. The idea—popularized by Andrej Karpathy in 2025—shifts the process from writing code line by line to guiding an AI assistant that creates, tweaks, and debugs your app. Instead of obsessing over syntax, you think about what you want the app to do, and let the AI handle the details.

When I started, I didn’t know this term even existed. I was simply experimenting with Grok and tried building a Python app to parse my Google Location-history.json file. My goal was simple: figure out how much time I’d spent in different states during the year. (Google has been tracking me since 2009.)

Phase 1: Conversational Coding with Grok

I began with a conversational coding session. Within two days, Grok and I had a working desktop Python app that consumed my JSON file, calculated days spent in each state (or country, if outside the US), and exported a CSV.

The amazing part? I don’t know Python. I hadn’t coded in 20 years. Yet here I was, ‘coding’ my own app in Python.

Encouraged, I moved the project online with Flask. The early web version was crude—rough pages, lots of trial and error—but it could generate the CSV reports I wanted. That was enough to hook me.

Phase 2: Iterating with Claude

Next, I tried Claude.ai. It was far better at UX and could generate full .py files on demand. Sometimes everything worked; other times I had to splice pieces together, test, debug, and loop until stable. The process was iterative—Claude made human-like mistakes, even basic formatting errors—but progress was steady.

Eventually, Claude produced a massive Flask app: one 2,800-line .py file plus several helpers. The logic was scattered, making bugs painful to track down. So I asked Claude to refactor it into proper modules—file management, JSON parsing, reverse geocoding, and so on. To my surprise, it worked quickly. Within two days I had a well-structured web app that not only summarized yearly time by state but also listed cities visited and total miles traveled. Debugging became far easier. But progress had slowed ALOT and debugging was super painful.

Phase 3: No-Code with Replit

After weeks of pasting AI-written code into files, I was ready for a change. Enter Replit. This time I wrote no code at all.

Armed with a clear idea of what I wanted, I asked Replit to build the app from scratch. Because I’d already learned so much in earlier iterations, I could guide it precisely. In just two days, I had a polished web app with mapping, analysis, and yearly reports of where I’d spent my time.  A few weeks later it was stable with complete features.

Today, that app lives at www.ezpoint.com and does a solid job analyzing Google Location History.

Ten Things I Didn’t Expect

1. You can “code” without coding. Even pasting AI output into VS Code works if you understand basics from other languages.
2. AI sessions have memory limits. Claude often slowed down or failed mid-debug with “prompt too long.” It felt like your programmer quit mid-project and you had to onboard a new one.
3. AI can’t help itself. Claude once confessed, “I cannot help myself—I just added the extra code you told me not to.”  That was super odd to hear..
4. AI hallucinates. It sometimes suggested patching code that didn’t exist.
5. AI forgets. Claude repeatedly reintroduced old JSON parsing mistakes, following outdated internet advice. Even though I had it WRITE IT INTO THE NOTES.
6. AI rewrites its own work. Changes sometimes broke perfectly good code, that it had JUST written.
7. AI misguides. It told me to drop DB tables I was still using—until I showed it the resulting error. It was 'embarrassed' to have made that mistake.
8. You still need technical navigation. SQL, DNS, and command lines don’t disappear, though AI can usually guide you if its not your thing.
9. You must act as a GOOD product manager. Give precise direction, stop wasted effort early, and roll back when things derail.
10. AI is brilliant and clueless at once. It rarely thinks outside the box, so you’ll waste hours on problems it is sure is in one place, when it is somewhere else.

Conclusion

This experiment proved to me that it’s entirely possible to “code without coding.” The key is knowing exactly what you want and treating your AI like a brilliant but stubborn four-year-old. Clear instructions, constant guidance, and a willingness to iterate are essential.

Project Snapshot (3 Weeks of Effort)

• Files: 100 source files (85 frontend, 14 backend, 1 shared), 19 CSS
• Lines of Code: ~58,000 TS/JS, 4,000 CSS
• Structure:

The result is a substantial full-stack app: location data processing, user management, interactive maps, analytics, and admin tools—all built in weeks with AI. 🚀 And I'm still not a programmer.