3 Learnings from Building a Weekend App with AI

This past long weekend (thanks, Easter break) I dug into an idea I had knocking around. I read Hacker News religiously and the lengthy comments sections can be as insightful as the articles they reference.

However, the most interesting stories on Hacker News often have too many comments to digest. So I thought, why not build an AI app that summarises them for me?

Enter HN Summary. You can view a video demo on twitter or Linkedin.

The project started of as a rough vibe-coded throwaway app using Cursor. But after using the scrappy version of the product I realised that I actually enjoyed reading the summaries it generated. Maybe I could use this daily?

To build something I could rely on I had to go beyond "vibe coding". I had to work on secure APIs, cron jobs, caching, databases, user authentication and experiments with a variety of third party APIs.

Platforms like Lovable and Replit take care of all the backend plumbing so there's an option to skip the grunt work here. For this project though, I wanted to learn while building and I especially wanted to test the latest large language models given what people are saying about AI replacing software engineers.

What did I learn?

1. Vibe coding is great for prototypes but terrible for large long-term projects. If you keep building larger projects with AI doing all the work, you can end up with a large, buggy, unsecure, codebase that you don't understand. This makes it harder to iterate on projects and fix issues in the future.
2. Software engineers won't be replaced so soon. Yes, there's lots of hype about AI taking over jobs in software engineering, and some credible voices are saying that within 12 months AI might be writing all of the code. Even though I recognise my limited credibility here given I'm not a AI researcher, I'm confident this isn't happening within a year or even 3 years. Just speak to anyone who's building mission-critical software in a bank, healthcare setting, or some other high-stakes environment. They don't trust AI code generation systems yet and rightly so (see point 3). Even if AI does eventually write all the code, we will still need people that bring systems thinking, nuanced architectural decisions, and context-aware choices that you can't blindly delegate to a large language model.
3. Hallucinations are still an issue! I used Sonnet 3.7 and Gemini 2.5 in Cursor to help me build the app but even these state-of-the-art models kept making things up. They even went as far as reading files they shouldn't have had access to, and at times edited my codebase in areas that were outside the scope of the instructions I provided. I'm sure this will get better as new models are developed, but I'm surprised that after all the billions of dollars in investment in training LLMs, hallucinations are still an issue.

Building the app was still fun and I've made it public for now to see what people think. It's more sophisticated compared the first app I built with the help of AI. Maybe I'll keep using it personally or sunset it if I find something better. For now, its free to use with limited credits. Check it out at https://www.hnsummary.ai/.

The tech stack I used was: React 19, Tailwind CSS for styling, Supabase for data storage and user authentication, and deployment on Vercel. I'm also using the Hacker News public API, and large language models from OpenAI and Google.