What Claude Got Wrong While Building My Next.js Blog

I recently built a blogging platform using Claude as my primary AI assistant.

Overall? It was impressive.

But it wasn’t perfect — and the imperfections were interesting.

Here’s what went wrong.

---

1. Version Awareness Issues

Modern JavaScript stacks move fast.

During development, I noticed:

• Confusion around the Next.js middleware → proxy shift
• Patterns that didn’t fully match Prisma v7
• Inconsistent alignment with NextAuth v5

None of this broke the project.

But it highlighted something important:

AI doesn’t automatically understand the current state of the ecosystem.

If you’re using bleeding-edge versions, you still need to guide it carefully.

---

2. SEO Was Almost Ignored

This was a blogging platform.

Yet:

• No proper static params generation
• Weak metadata setup
• No clear SEO-first structure

The UI looked good.
The routing worked.
The app ran fine.

But discoverability wasn’t prioritized.

AI optimized what was visible.
Not what was strategically important.

---

3. The JS Bundle Was Larger Than Necessary

The generated code produced a heavier-than-ideal bundle.

Interestingly, performance still felt fast.

But feeling fast and being optimized aren’t the same thing.

Bundle discipline still requires intentional review.

---

4. What It Did Surprisingly Well

To be fair, Claude handled changes cleanly.

When I asked it to:

• Refactor
• Add features
• Modify specific components

It rarely broke unrelated parts of the system.

That structural continuity was genuinely impressive.

---

What I Learned

AI is very good at:

• Local reasoning
• Feature iteration
• Maintaining structural consistency

It’s weaker at:

• Ecosystem nuance
• Strategic prioritization (like SEO for a blog)
• Long-term architectural thinking

AI accelerates execution.

But it doesn’t replace engineering judgment.

And honestly, that’s fine.

The real skill isn’t prompting.

It’s knowing what matters.