Building my first AI-powered app: what I learned as a non-ML developer

Full disclosure: the site you're reading right now is the app I'm referring to. viewsaremyown.com, a Next.js blog site deployed on AWS Amplify with a Neon PostgreSQL database and a custom admin panel. I built it with significant assistance from Claude.

I'm not a software developer. I'm an IT professional with 20+ years of experience who can write PowerShell, understands infrastructure, and has always been comfortable in technical environments. But building a full-stack web application was not something I'd done before in any serious way.

What I actually learned surprised me in a few ways.

What AI assistance actually looks like

The popular narrative about AI-assisted development tends toward one of two extremes: either "AI writes the code and humans just review it" or "it's all hype, you still need to know what you're doing." My experience was more nuanced than either.

Claude was extremely useful for scaffolding: generating the initial structure of components, API routes, database schema, and configuration files. Starting from scratch is the hardest part of building anything, and having a working starting point that I could understand, modify, and extend changed the experience completely.

But I had to understand what it was building. I couldn't just accept outputs blindly. When the generated code had bugs (which it did), I needed enough context to recognise the problem, describe it, and review the fix. When my requirements were ambiguous (which they were), I needed enough technical vocabulary to clarify what I actually wanted.

The AI didn't replace the need to understand technology. It changed the starting point and the velocity.

The surprising parts

Debugging was a genuine collaboration

When something broke (and things broke regularly), describing the error to Claude and having a conversation about what might be causing it was genuinely useful. Not because it always had the right answer, but because walking through the possibilities forced me to think more precisely about what I was observing.

Architecture decisions still required judgment

How to structure the database, how to handle authentication, whether to use server or client components for different pages: these required judgment calls that couldn't be delegated. The AI could explain trade-offs, but I had to make the decisions.

Documentation became essential

I wrote a CLAUDE.md file, a project document describing the site's goals, design system, tech stack, and content strategy, and referenced it at the start of each working session. Having that context available consistently improved the quality of assistance significantly. It's the closest analogue I've found to onboarding a new team member properly.

The learning was accelerated, not bypassed

Doing this project taught me more about Next.js, React, Prisma, and AWS in a few weeks than I would have learned in months of reading. The hands-on building, with AI assistance to get unstuck quickly, created a tighter feedback loop than traditional learning.

What I'd tell someone starting a similar project

1. Write your requirements down before you start. Ambiguity compounds in software development, and it compounds faster when you're delegating generation to AI.

2. Read and understand everything that gets generated before you move on. The code is yours; not understanding it is a debt you will pay later.

3. Test constantly and test with real inputs. Generated code often works for the happy path and fails for edge cases that weren't specified.

4. Don't be embarrassed about what you don't know. The whole point of using AI assistance is that it meets you where you are and helps you get somewhere new.

The project is live. It works. I built it. And I learned more in the process than I expected to. That seems like a reasonable outcome.