● open to work
LET'S TALK
Home
Articles
Opinion • AI Era
Opinion
AI & Dev
• Updated August 2026
Originally published March 2026
⏱ 10 min read
September 12, 2025

I DON'T

WRITE

CODE
ANYMORE.

and that's the point.

After 15+ years of typing every line myself from Manipur to Pune, client to client I had to confront a quietly uncomfortable truth: the best version of my work no longer starts with a blank file.

The Numbers Behind This Article

15+
Years coding
2025/26
The shift point
50+
Client projects
~80%
Code now AI-assisted

"There's a new kind of programming emerging not quite programming as we know it. You describe intent, and AI figures out the implementation."

— Andrej Karpathy, 2024 (still true in 2026)
~80%
Of my code is now AI-assisted
15+
Years writing code the old way
Still

Reviews every line before it ships

Better

Outcomes than ever

Claude Code✦ Vibe Coding✦ Cursor✦ AI Orchestration✦ WordPress + AI✦ Developer Identity✦ Front-End in 2026✦ Windsurf✦ v0.dev✦ Bolt.new✦ Tabnine✦ Replit Agent✦ Agentic AI✦ AI-Native Dev✦ Claude Code✦ GitHub Copilot✦ Cursor✦ Windsurf✦ v0.dev✦ Bolt.new✦ Tabnine✦ Replit Agent✦ Agentic AI✦ AI-Native Dev✦
August 2026 Update
This article was written in March 2026 as a personal reflection. We’ve expanded it significantly with new sections on agentic AI, Claude Code’s role in daily work, and the “vibe coding” conversation that’s dominated developer discourse this year.

How I Used to Work — VS Code, Every Line

There was a time when my entire working day lived inside VS Code. I would open a blank HTML file and start typing. Not copying from StackOverflow actually typing. Crafting the structure element by element, writing CSS from memory, wiring up JavaScript logic character by character.
I was proud of that. Growing up in Manipur, coding felt like a superpower I’d earned. Each function I could write from memory, each layout bug I could chase and kill these were proof that I belonged. That I could compete with anyone, from anywhere.
That identity was real. And it lasted about nine years.

Personal Context — Why This Matters to Me

I moved from Manipur to Pune to build a career in web development. In a field where most opportunities were in major metros and most clients assumed you needed to be in Bangalore or Mumbai, technical skill was my only equaliser.
The ability to write clean, fast, correct code faster than the next freelancer was how I got work, kept clients, and built a reputation across 50+ projects spanning India, the US, the Caribbean, and beyond.
So when AI started being good enough to do what I'd spent years mastering, it wasn't just a professional question. It was personal.

“

What Actually Changed — And When

The shift didn’t happen overnight. It accumulated. First it was GitHub Copilot for boilerplate saving me from typing the same React component scaffold for the hundredth time. Then Cursor, with its multi-file reasoning that started understanding my codebases holistically. Then in 2026, Claude Code and that’s when the relationship changed fundamentally.

Claude Code isn’t an autocomplete tool. It’s an agent. I give it a task “Refactor this WordPress template to be more semantic and add proper Schema markup” and it reasons across files, makes decisions, runs tests, and reports back. I review. I adjust. I approve or reject. But I’m no longer the person writing every line.

The first time I gave Claude Code a complex multi-file refactoring task and watched it work through the problem questioning its own assumptions, catching its own errors, asking me for clarification on genuinely ambiguous decisions I felt two things simultaneously: impressed and quietly unsettled.

Before AI Agents (2024 and earlier)

  • Open VS Code. Start typing HTML from scratch
  • Write CSS rule by rule, tweak breakpoints manually
  • Wire up JavaScript interactions character by character
  • Debug by reading stack traces and searching docs
  • Write boilerplate that's nearly identical every project
  • Refactor = rewrite manually, file by file

After Claude Code & Cursor (2026)

  • Describe the component and its constraints in plain English
  • Review and refine the AI's generated CSS improve, don't recreate
  • Verify the AI's logic meets performance and accessibility requirements
  • Debug by asking the agent to reason through the error with me
  • Skip boilerplate entirely focus on the non-obvious decisions
  • Refactor = describe the goal, review the diff, approve or iterate

On "Vibe Coding" — What Karpathy Got Right

In early 2025, Andrej Karpathy coined the term “vibe coding”  the idea of describing what you want and letting AI figure out the implementation while you focus on the overall feel and direction. It became one of the most discussed (and debated) developer concepts of the year.

Some developers hated the framing. It felt like it was dismissing the craft. Others embraced it uncritically treating AI output as production-ready without real review. Both camps missed the point.
"Vibe coding is not abdication of skill. It's a redirection of skill toward judgment, taste, and systems thinking, and away from implementation mechanics."

— Somorjit Konthoujam, August 2026

“

Claude Code — Real Session Example
Actual workflow
Me:This WordPress landing page has a hero section that breaks on iOS Safari 16. Find the issue and fix it without changing the visual design.
AI:Reading template files... found 3 CSS properties that cause issues on WebKit. The main culprit is a min-height: 100vh on a flex container — iOS Safari handles viewport units differently. Also found a gap property not supported in older WebKit. Proposing fixes:
AI:→ Replacing min-height: 100vh with min-height: -webkit-fill-available fallback
AI:→ Adding -webkit-gap polyfill for flex gap
AI:→ Adding @supports query so fix only applies when needed
AI:Applied to hero.css. Visual design unchanged. Want me to check other templates for the same pattern?
Me:Yes — check all page templates and fix wherever you find it.
// 4 minutes later: 6 files fixed, diff reviewed, approved, committed.
Claude Code WordPress Multi-file reasoning Human-approved every step
That session used to take me an hour of manual investigation, cross-referencing browser compatibility tables, and testing on BrowserStack. The AI did the investigation in 4 minutes. I spent the remaining 56 minutes on something that actually needed my specific judgement: restructuring the conversion flow on the same landing page.

The 8 Tools — Reviewed for 2026

The framing that resonates most with me and that I think describes where front-end development is heading is the concept of the orchestrator.

An orchestrator doesn’t write every part of the music. They understand every part deeply enough to know when it’s right, when it’s wrong, and how to direct each player toward the overall vision. They can pick up the violin if needed. But their highest-value contribution is the total picture not any single note.
A large, bold, lowercase red letter s followed by a red period on a light gray background.

SOMORJIT — ORCHESTRATOR

Front-End Engineer & WordPress Performance Expert

Architecture & Decisions

Structure, patterns, tech choices
Claude

Claude Code

Complex builds, refactoring, agents
Cursor

Cursor

Daily editing, multi-file context

v0.dev

UI component generation

Performance Review

PageSpeed, CWV, bundle size

Accessibility Check

What AI consistently misses

Final Review

Every line before production

Ship & Monitor

Launch + post-launch care
→ Every tool in this stack runs through human judgment before anything reaches a client or production environment.

What a Real Workday Looks Like in 2026

I want to be specific here, because vague claims about “AI-powered workflows” obscure more than they reveal. Here’s an honest breakdown of how I actually spend my time on a typical project day:

Architecture & Problem Definition

Deciding what to build and how. Reviewing the brief, scoping the problem, defining the component structure, identifying performance constraints before writing a line.
30%
Of daily time

Reviewing & Refining AI Output

Reading every generated file carefully. Checking semantics, accessibility, TypeScript types, CSS specificity, performance implications. This is skilled work not rubber-stamping.
35%
Of daily time

Hand-Writing Critical Code

Complex business logic, performance-sensitive sections, accessibility patterns AI gets wrong, security-critical code. I still write this myself AI doesn’t get to touch these.
15%
Of daily time

Testing, Performance & Communication

PageSpeed, GTmetrix, browser testing, client calls, written updates. These are non-negotiable and cannot be delegated to AI or anyone else.
20%
Of daily time
The honest number

Roughly 80% of the code that ships in my projects is now AI-assisted in some way generated, refactored, or significantly improved by AI. But 100% of that code passes through my judgment, review, and explicit approval before it reaches a client or production environment. The ratio has changed; the responsibility hasn’t.

What This Means for WordPress Work Specifically

WordPress is interesting in the context of AI-assisted development because the stack is genuinely complex: PHP, MySQL, the block editor, Elementor, WooCommerce, custom post types, hooks, filters, REST API, multisite it’s an ecosystem, not a framework.
In 2025, AI tools were mediocre at WordPress. They’d generate boilerplate PHP that looked correct but had subtle hook priority errors, or CSS that worked in isolation but broke inside Elementor’s widget rendering. You needed deep WordPress knowledge just to evaluate what the AI gave you.

In 2026, Claude Code changed this substantially. Its understanding of WordPress internals action hooks, filter chains, template hierarchy, WooCommerce’s cart/checkout flow is now genuinely useful. I use it daily on client sites. But I still catch errors that would have caused real problems: wrong hook priority, a function that fires on the wrong conditional, CSS specificity that looked fine in isolation but collided with a plugin’s output.

The lesson: AI is only as useful as the human reviewing it is skilled. A junior developer using Claude Code on a WordPress project is not the same as Senior Experienced Developer using Claude Code on a WordPress project. The tool is the same. The judgment layer is not.

The Part AI Still Can't Do

I want to be clear about this because I’ve seen the opposite claim made that AI will eventually replace all of it. That’s not what I observe in my daily work.

Where AI Struggles

Contextual judgment about why a design choice exists. Knowing which performance tradeoff fits this specific client’s audience. Reading between the lines of a vague brief. Noticing that an accessibility pattern works technically but creates a confusing UX for this specific flow.

The Gap

Where Humans Remain Irreplaceable

Understanding client goals beyond the stated brief. Anticipating business context the code doesn’t show. Architectural taste: knowing not just what works but what will still make sense in two years when someone else maintains it.

The real skill gap in 2026

The developers who are struggling right now are not the ones being replaced by AI. They’re the ones who don’t have enough foundational knowledge to evaluate what AI gives them. You cannot review the code you don’t understand. You cannot catch AI’s mistakes in a domain where you have no expertise. The fundamentals matter more now, not less because they are the judgment layer that makes AI useful rather than dangerous.

Being Honest: The Uncomfortable Side

I would be dishonest if I only wrote about the productivity gains and the liberating shift in how I think about my work. There’s an uncomfortable side I’ve sat with too.
There are moments when Claude Code handles a complex multi-file refactoring in eight minutes that used to take me three hours where I feel a flicker of something I can only describe as irrelevance anxiety. If the tool can do this, what am I for?
That feeling passes. And when it does, I notice what’s left: the judgment call about whether the refactoring was actually the right architectural decision. The conversation with the client about why this approach fits their long-term goals. The performance audit that reveals the AI’s solution was technically correct but missed a critical Core Web Vitals impact. Those remain mine.

I’ve made peace with this framing: I am not the person who writes the code. I am the person responsible for the outcome. Those have always been different things. AI just made the gap between them more visible.

FAQs

If AI writes most of the code, what are clients actually paying for?
They’re paying for judgment, architecture, responsibility, and results. A fast WordPress site that converts not the specific act of typing the code that makes it so. This is the same reason you pay a surgeon rather than a typing speed champion. The skill is knowing what to do and taking responsibility for the outcome.
I review every generated file before it ships. I check semantic HTML structure, CSS specificity and cascade logic, TypeScript type correctness, accessibility patterns, and performance implications. I run PageSpeed and Core Web Vitals checks on every deliverable. The AI writes a first draft; I’m the editor and technical director who decides what actually ships.
No, but they should use them carefully. AI is an excellent learning tool when you use it to explain code, not just generate it. Ask it to explain what it wrote and why. Then verify that explanation against your own understanding. The danger is using AI to skip the fundamentals; the opportunity is using AI to accelerate through them with constant explanation.

Claude Code for agentic tasks and complex multi-file work (especially WordPress). Cursor for daily IDE editing and inline assistance. v0.dev for rapid UI component generation. GitHub Copilot occasionally for quick completions when I’m already in VS Code. I wrote a full breakdown in the 2026 AI tools article.

Yes. Complex business logic where clarity and auditability matter more than speed. Performance-critical sections where I need precise control. Security-sensitive code. Accessibility patterns that AI consistently gets subtly wrong. And sometimes just because I want to the craft still matters to me, even when the outcome doesn’t depend on it.

If this resonated
I write from experience, not theory. If you’re a business looking to understand how AI-assisted development can speed up your WordPress work or a developer curious about how I structure this workflow for client projects let’s have a conversation.
Share this Article
Facebook
X
LinkedIn
Email
WhatsApp
← Previous article
Best AI Tools for Front-End Devs (2026 Edition)

Journey

Freelancing

Next Article →
From Manipur to Pune — A Freelancing Journey

AI & Dev

Web Development

SOMORJIT KONTHOUJAM

WordPress Performance Expert & AI-Driven Developer with 15+ years of experience. Based in Pune, India. I help businesses get faster sites, better rankings, and launch-ready pages.

LinkedIn
Work with Me →
View All Articles →
Reading progress
- Key Takeaways
  • ~80% of shipped code is now AI-assisted but 100% passes through human review
  • Claude Code changed WordPress development in 2026 more than any previous tool
  • The orchestrator model: direct, review, and approve don't just generate
  • Fundamentals matter more in the AI era, not less they are the judgment layer
  • AI cannot replace contextual business judgment or responsibility for outcomes
AI in Practice
NEED AN AI-POWERED SITE?
I use Claude Code and Cursor on real WordPress projects. Want to see what AI-assisted development can do for your site?
Lets Talk →
Scroll to Top