AI
I Built a Chrome Extension in 4 Hours Using Claude Opus 4.6 — A Real-World Test of Spec-Driven AI Development
Vlado Balko

I couldn’t find the extension I needed. So I built it in 4 hours. And I used it to stress test Claude Opus 4.6 in a real production scenario.
If you work with APIs long enough, you develop a quiet, daily annoyance. Mine was this: copy minified JSON from a <textarea> , paste into a formatter, copy it back. Repeat. Every. Single. Day.
So I did what any developer would do — I searched the Chrome Web Store.
Surely there’s an extension that formats JSON and XML directly inside textareas and code editors, right?
Wrong.
Everything I found either formatted raw responses only, forced copy paste, or ignored editors like Ace, CodeMirror, and Monaco.
That’s when this stopped being a search problem and became an experiment. I’ve been writing about how AI coding evolved from “vibe prompting” to spec driven engineering. When Claude Opus 4.6 was released, I wanted a real test.
Not a toy script. Not a hello world demo.
A production Chrome extension. Built properly. Shipped the same day.
The Build: 3–4 Hours of Actual Implementation
The full journey (idea → Web Store submission) took about 8 hours.
But the actual implementation time? Around 3–4 hours.
I started with a long initial prompt where I:
Described the problem
Outlined all feature ideas
Asked Claude to brainstorm and challenge assumptions
Then I switched to plan mode and asked it to do comprehensive research first. This is where Opus 4.6 surprised me.
It reasoned through:
Manifest V3 constraints
ISOLATED vs MAIN execution worlds
Shadow DOM isolation window.postMessage communication bridge
Editor API access patterns
Smoothly. Cleanly. Architecturally consistent.
This was easily the smoothest AI-assisted programming experience I’ve had so far. Was it truly one-shot?
Let’s say… 3–4 shots. But no architectural rewrites. No chaos.
Compared to my earlier “vibe coding” experiments — where prompts produced working but messy code — this felt fundamentally different. The model maintained structure across files and iterations. It behaved more like a senior engineer than an autocomplete engine.
That’s the biggest improvement.
What JSON Spot Does
The extension (JSON Spot) does one thing well:
It detects JSON or XML inside:
Textareas
Ace
CodeMirror 5/6
Monaco contenteditable elements
And lets you format/minify it in place.
Features:
Floating format/minify button
Keyboard shortcut (Alt+Shift+F)
Context menu integration
Element picker mode
Syntax highlighting
Clipboard copy
Dark mode popup
Badge count per page
~1,100 lines of JavaScript.
0 dependencies.
Pure Manifest V3.
And most importantly: no more copy paste.
What Improved Since “Vibe Coding”
In my earlier articles, AI coding felt like this:
Fast generation
Frequent refactoring
Architectural drift
Lots of cleanup
With Opus 4.6 + plan mode, this changed.
What improved:
Architectural stability — no structural rewrites.
Cross-file reasoning — consistent patterns across modules.
Better upfront thinking — planning actually reduced total iteration.
Less entropy — the codebase didn’t degrade over time.
The key difference wasn’t just the model. It was combining a stronger model with a spec-first workflow.
Specs > vibes.
And now it actually feels practical.
The Numbers
~3–4 hours implementation
~8 hours total (including packaging, store prep)
~1,400 LOC across 7 files
0 dependencies
24KB zipped v0.1.0 → v0.7.0 in one day
Try It
The extension will be available on the Chrome Web Store once it’s verified and public.
Chrome Web Store:
https://chromewebstore.google.com/detail/henkkoibcmmfphlpfmfnafejgalomhok
GitHub:
https://github.com/vbalko-claimate/json-spot-extension
If you’ve read my previous AI coding articles, this is the first real production case study.
If you haven’t, start here:
How AI Coding Evolved from Vibes to Specs
This extension is proof that the workflow described there now holds up in real shipping conditions.
AI
I Built a Chrome Extension in 4 Hours Using Claude Opus 4.6 — A Real-World Test of Spec-Driven AI Development

