Opinionated AI Development

Earlier this year, I started AI-assisted coding. It was the first time I felt the magic of software - a few hundred lines of code turning an empty canvas into a playground. As a product manager at a SaaS startup I had seen many ideas come to life, from idea to design to working software, but this was the first time I got to make all the decisions.

I took inspiration from this post from Marc Louvion, a prolific indie hacker, and chose Next.js, Tailwind, Vercel, and Supabase for my first project. But as I started more projects, I noticed something. AI wanted to use that exact same stack for everything - whether it made sense or not.

A few months in, I had the opportunity to help my friend set up his new company’s website. It was a super simple brief - a single page website that would showcase his videography agency and a contact form. I fired up a new project, and of course AI started building with Next.js.

But in my head I was wondering why we were using the same framework used to build applications for a single page website.

I decided to break the pattern and force AI to use Astro, a new but popular framework for content-heavy sites. I wasn’t trying to be contrarian - I wanted to get back to the problem I was solving: a simple single page site to showcase my friend’s business to potential customers. I found the code easier to read and understand, plus Astro is fast and built for content sites.

Astro has great documentation so getting AI to work with a different framework wasn’t impossible, but it would have been easier to go with Next.js and let AI make all the decisions.

But by consciously making a different choice, I ended up with a technical solution that better aligned with the need I was trying to fulfill.

The joy of any product development comes from solving needs, whether that is enterprise SaaS software, or a small personal project. With AI coding, everyone has become a builder, but not everyone is a problem solver.

The barrier to entry for web development has dropped significantly, which creates an arbitrage of software everywhere. And even though now I have the skills to create software, the skill of solving true needs doesn’t just come out of the box with AI. Yes of course you can prompt your way to finding problems and then trying to solve them… but the most useful things I’ve built have come from identifying real needs and solving them, not manufacturing problems with AI.

And solving those needs well requires more than just blind adherence to any suggestion AI makes, but an understanding of what you’re building and why.

After a year of “vibe coding” I’ve landed on my preferred stack. It’s not always the AI defaults, because it depends on the need I am trying to fulfill.

This is where skilled product managers will stand out in the AI era - understanding needs and making judgement calls on what needs to be done to solve them, using AI to accelerate each step. The advantage is understanding how to harness AI as a partner, not let it drive while falling asleep at the wheel.

This is what PMs have always done - understand tech enough to make good decisions, know what you actually need, prioritize, and drive projects forward. AI makes building easier, but it makes judgment even more important.

Thanks for reading!