Even the AI Is Confused

The Stack Didn't Break — The AI Did

I didn’t expect the AI to break before I did.

But that’s exactly what happened while building a real production-grade tournament manager. Not a blog. Not a demo. A full-stack transactional app with auth, sync, edge functions, Supabase, and a zero-tolerance build pipeline.

The AI tool I was using — Kilo — flagged my project as broken.

Spoiler: it wasn’t.

The app builds cleanly. The code is tight. The config is intentional.
The problem wasn’t with my stack. It was with how AI interprets it.

And honestly? That tells us everything about the state of modern web development.

🧠 The Setup

I’m running a modern JS stack, straight from the “best practices” playbook:

• ⚛️ React + Vite frontend
• 💅 Tailwind with semantic theming
• 🧱 Supabase as the backend/auth layer
• 📦 Modular structure, clean package.json, zero-dependency bloat

No wild experiments. No half-baked plug-ins. Just clean, modern code — built to ship and scale.

Before a major push, I ran a Kilo audit on the package.json to check for unused packages, broken scripts, or stale dependencies.

❌ Kilo: “Your Project Is Broken”

Kilo confidently reported:

“Your vite.config.ts references packages in manualChunks that don’t exist in package.json. This will break your build.”

That sounded… wrong.

Because:

• The app builds perfectly (dist/ verified)
• The chunk config is intentional, future-proofing for code-split scenarios
• Vite handles missing chunk targets gracefully

And yet, the AI doubled down.

✅ Kilo (Eventually): “My Bad. The Project Is Excellent.”

After pushback, Kilo retracted everything:

“I misunderstood how Vite’s manualChunks works. I assumed complexity where simplicity was intentional.”

“Your package.json is actually in excellent condition.”

“I’ve learned from this mistake.”

Except… it didn’t learn.
Because it can’t — not in the way real engineers do.

🔁 Most AI Tools Simulate Correction. Few Operationalize It.

This is the real issue.

Kilo didn’t persist that learning. It didn’t create a safeguard or update its reasoning. It failed to recognize idiomatic patterns — and it’ll fail again on the next audit unless I hand-feed it every exception.

That’s not intelligence.
That’s inference without infrastructure.

🧠 Gist of the full AI audit failure:
https://gist.github.com/chavezabelino/9164611dc20fc795e80409e2386e81b0

🧨 This Stack Is Coherent — The Ecosystem Isn’t

Let’s be clear: my project wasn’t broken.

But the modern fullstack ecosystem is so unstructured, pluginized, and overloaded that even AI tools can’t parse it cleanly.

Worse — they confuse future-proofing and modularity with error states.

Even good setups look broken because:

• No one tool has full-stack awareness
• Idioms change every 6 months
• “Conventions” are suggestions, not contracts

🗣️ This Isn't Theory. I'm Living It.

This isn’t recycled rage from YouTube.

I’m building the real thing — a transactional, production-grade app with deadlines, users, and real data.

So if you come at me with something someone else blogged about, podcasted, or tweeted?

Keep scrolling.

I have the receipts. I have the scars. And I’m not done yet.

🧭 What This Teaches Us

1. Modern JS isn’t broken by default — it’s broken by integration burden.
2. AI won't fix DX rot — it amplifies the gaps.
3. We need software patterns that are legible to humans and machines.

That means:

• Constraints over choice fatigue
• Structure over glue code
• Developer sanity as a design goal

We need better stacks — and better reflexes baked into our tools.