Can Developers Really Correct AI? A Story From My Own Experience

The other day, I had a conversation with a friend about the role of AI in software development. His argument was simple (and very confident):

“If you prompt well, AI generates perfect code. Developers can’t really correct it because it’s already beyond what they could do.”

At the time, I didn’t respond much. But after using AI tools every day for over two years, I knew this wasn’t true. I’ve seen AI do amazing things — but I’ve also seen it make mistakes that only experience could catch.

Let me share one story.

The Deployment Example

I was self-hosting a Next.js application. No Vercel. No Netlify. Just my own server and GitHub Actions for CI/CD.

The task? Deploy the same app to two different domains.

Instead of writing the workflow from scratch, I provided AI with all the details: my project structure, deployment requirements, server configuration, and asked it to configure GitHub Actions to deploy my app to two domains. It quickly generated a full YAML pipeline.

It looked perfect. I hit deploy.

And then… everything broke.

The Mistake AI Made

The problem wasn’t obvious in the code. The pipeline itself was valid. But here’s what the AI completely missed:

• Next.js environment variables (those starting with NEXT_PUBLIC_) must be available at build time, not runtime.
• The AI’s config didn’t include that, so the build failed.

Because I had dealt with this before, I knew exactly what was wrong. I pointed it out, asked AI to revise the config, and only then did it fix the pipeline.

The Takeaway

That day gave me the answer I wish I had told my friend earlier:

• AI can move fast.
• AI can generate working code.
• But AI doesn’t always understand context.

Developers aren’t just “prompt writers.” We bring the experience, intuition, and debugging skills to catch what AI misses.

AI doesn’t replace developers. It needs developers.

What do you think? Have you had moments where AI gave you code that looked fine but failed in practice?