Lessons from Vibe-Coding My First App


Hi Reader,

If everyone can do vibe-coding, it's tempting to think coding is officially DEAD.

But here's the truth: Vibe-coding is not a free lunch.

I recently listened to a Lex Fridman's podcast with DHH, a respected programmer. They had a very interesting discussion on vibe-coding, I'd encourage you to check it out.

In the past few months, some of you may have dabbled in vibe-coding.

It’s a pretty exciting time - anyone can build functional apps with AI-powered coding tools like Cursor, Replit, or Lovable just by describing what they want. Anyone can now build prototype and explore their new ideas.

But is vibe-coding for everyone?

My vibe coding experiment

Out of curiosity, I recently ran an experiment using AI coding agents to create a web app - and I break it all down in my new video:

video preview

With this experiment, my idea was to build an AI-powered web app that can extract actionable tasks from an email, screenshot or document, and even help me decide how important or urgent it is.

Here’s what I did:

  • I wrote a detailed doc outlining the concept and key features (see below).
  • I sent it to Manus agent to generate the first prototype.
  • When Manus got stuck, I moved the code to Cursor (an AI-powered code editor) to polish the UI and add more features.

I just ACCEPT ALL everything the agents did, without checking a single line of code.

Lessons I learned

After hours of vibe-coding with Manus and Cursor, here's what I learned:

1. AI will do the absolute minimum at first

Don't expect the AI coding agent to give you an amazing prototype out of the gate. Even if you clearly specify what you want, AI often simplifies, takes shortcuts, or skips essential logic.

👉 Vibe-coding is great for quick prototypes and pre-defined tasks, not for full-fledged apps with complex backend logic.

2. Trust but verify

A lot of times the coding agent will confidently tell you it’s successfully implemented something. But then you realize the feature didn’t work, and worse, something else breaks.

👉 You always need to test what the AI has done: run functional tests, do visual checks, or review the code yourself. I guess that is the human part of the workflow.

3. Starting is easy, finishing is brutal

With AI tools, starting is easy, but finishing is much harder.

In a traditional development workflow, you build things slowly at first, and then momentum builds over time.

In an AI-assisted workflow, starting feels magical - you get to 80% in no time. But it’s a bit of an illusion: Progress starts to stall, especially if you hit roadblocks and don’t know how to move forward.

4. You (still) need coding basics

Throughout my vibe-coding experiment, I had to debug, fix broken features, and figure out what went wrong.

And if you don’t have at least some coding skills, that’s nearly impossible.

👉 AI tools (still) don’t remove the need to understand what you’re building, or the need to learn some coding fundamentals and software engineering best practices. You don't need to be a coding pro, but just need to know enough to be dangerous.

5. The nuance of time-saving AI tools

A recent randomized controlled trial (RCT) by Meteor examined how frontier AI tools impacted the productivity of experienced open-source developers.

The results were surprising: AI slowed developers down by 19%, even though they thought they were saving time.

It's found that:

  • AI tools struggle with complex projects
  • Developers accept suggestions only ~50% of the time
  • Cleaning up AI code often takes more time

So instead of saving time, developers ended up spending more of it. That said, this study is conducted on experienced developers regularly working on complex projects. I'd imagine the impact on less experienced folks to be much more positive.


🔨 New tool I've tried

I used to hate the dictation feature on Mac. It was so bad, I figured it’s faster to just type 🥲.

But recently I discovered a free dictation tool called MacWhisper that works shockingly well. If you want to speed up things when writing emails, drafting documents or vibe-coding, this is an amazing tool (they didn't sponsor me to share this!). This tool is built only for Mac unfortunately, but I bet there are great alternatives for Windows users too.

I just downloaded the free version, chose a local AI transcription model (which is free to download), and set up a quick keyboard shortcut to turn dictation on whenever I need it.

It's been a game changer saving me hours of typing lately. Try it out, and you can thank me later 😉.

Wishing you a wonderful rest of the week,

Thu 🤗


P.S.: Work with me:

So if you’ve tried vibe coding and now feel the itch to really learn how to code, build AI applications from the ground up and understand what’s going on under the hood - I’ve built a program to teach you all the fundamentals of Python, Machine learning and AI👇.

You'll join a community of 300+ learners who are building their projects while getting direct access to me and supporting each other along the way.

🔗 Learn More


Thu Vu

Say hi 🙌 on Youtube, LinkedIn, or Medium

Thu Vu

Join 6,000+ data professionals who are advancing their careers by learning from real-world projects, gaining first-hand experience, and accessing top resources delivered to your inbox every week.

Read more from Thu Vu

Hi Reader,I’ve always loved making data visualizations. Partly because they look cool. But mostly because they make data and insights easy for anyone to understand. The problem? Coding up a visualization - or even better, an interactive, self-explanatory data story you can share - takes a ton of time. Recently, I had a lightbulb moment: What if I tried AI coding agents to build scrollytelling visualization apps? 🤔 My final data scrollytelling app created with Cursor coding agent. Not perfect,...

Hi Reader, Back in 2019, I decided to start a computer science bachelor degree, specializing in Machine Learning and AI. I was 27 at the time, and I felt lucky to find an online program that let me study while keeping my full-time job. I thought, “Great, I’ll finish it in 3 years.” Fast forward 6 years, with plenty of ups and downs in between, and I’m just now finishing the last module 🥲. If you asked me today whether I’d do it all over again, I’d hesitate. On one hand, the degree gave me...

Hi Reader, OpenAI dropped GPT-5 a couple of days ago, and the internet lit up. Some are calling it brilliant. Others are calling it a major let down after all the hype. After sifting through dozens of tweets and blog posts, here’s my takeaway: GPT-5 isn’t AGI - not even close. It’s more of an incremental upgrade than a godlike breakthrough over previous models. It still makes silly mistakes, has a glitchy “routing mechanism,” and struggles with understanding and generating images. Case in...