I Let AI Write 40% of My Code for 4 Months—Here's What Actually Happened

Last Wednesday, I was up until 1am messaging a backend engineer friend. He'd just spent three days building an internal data processing tool, only to discover the product manager from the neighbouring team had spun up something similar in 30 minutes using Cursor. With a decent UI, no less.

His soul left his body. I could practically hear it through the screen.

This is what everyone's calling "Vibe Coding" now. The term's been absolutely everywhere on Twitter—people showing off screenshots of apps they "built with one sentence." If you're still wondering what the fuss is about, here's the gist:

You describe what you want in plain English, AI generates the code, and you just... feel whether it's right.

No wrestling with syntax. No digging through documentation. You speak human, the machine writes code.

I've been using this approach heavily since January—about four months now. I've fallen into some spectacular traps, and I've also found scenarios where it's genuinely transformative. Let me walk you through what I've learned.

TL;DR for the Skimmers

• I use AI to write ~40% of my code now. It's brilliant for internal tools, frontend components, and prototyping.
• It nearly caused a production disaster when AI-generated payment code had a subtle concurrency bug that could've cost £3,000-5,000.
• Never trust AI with money, security, or complex state machines. Just don't.
• This isn't a passing fad—it's fundamentally changing what it means to be a developer.
• If you're a programmer today: learn to use AI tools, specialise deeply in a domain, and get good at "prompt engineering" (or whatever we end up calling it).

Not a Silver Bullet, But It's Eating Certain Scenarios Alive

Let me give you my actual numbers. Over the past three months, roughly 40% of my code has been AI-generated. It breaks down into three main buckets:

1. Internal Tools and Scripts

My team regularly pulls data from several sources, cleans it up, and generates reports. This sort of work used to sit in a backlog queue waiting for a backend engineer with spare cycles. Now? I open Cursor, describe what I need, and it's running in 10 minutes.

Last week I built a competitor price monitoring scraper with a notification bot. I didn't write a single line of code manually—well, I should correct that. I didn't hand-write any code. I was constantly telling the AI things like "add exception handling here" and "store this as a CSV."

The distinction matters. You're not passive in this process.

2. Frontend Pages and Components

This is where Vibe Coding absolutely shines. Tailwind components, form validation, responsive layouts—you describe what it should look like, and the AI spits out something mostly usable. I recently built a dashboard page where AI wrote about 80% of it. My contribution? Tweaking spacing, adjusting the colour scheme, the kind of micro-adjustments that just "felt off."

3. Prototype Validation

Here's the real game-changer. Previously, validating an idea meant sketching wireframes, then negotiating with developers for time estimates. Now I say "build me a landing page for collecting user feedback, with form submission and email notifications," and I've got something running in half an hour.

For indie developers, this speed is basically cheating. I mean that as a compliment.

When the "Vibe" Lied to Me

Right. Story time. And this one still makes me uncomfortable.

Last month I nearly caused a proper disaster. I'd used AI to generate code for a payment module. The logic looked correct, tests passed in the staging environment, everything seemed fine. We deployed it.

The next day, a colleague noticed that under concurrent requests, there was a risk of duplicate charges. The AI-generated code had a subtle flaw in transaction handling—it used READ COMMITTED isolation level, but that particular scenario needed SERIALIZABLE.

I felt physically ill when I realised. Genuinely.

If that bug hadn't been caught, we were looking at roughly £3,000-5,000 in potential losses. For a small team, that's not nothing.

After that near-miss, I established some hard rules about what AI should never touch unsupervised:

• Anything involving money or financial transactions: Concurrency, atomicity, rollback logic. AI is frighteningly good at writing code that looks correct but has hidden race conditions. That vulnerability I mentioned? It would've passed a casual code review.
• Security-critical functionality: Permission checks, data encryption, SQL injection prevention. I ran a security scan on some AI-generated code and found two medium-severity vulnerabilities—both were the "looks filtered but actually bypassable" variety.
• Complex state machines in core business logic: Multi-state transitions, edge cases. AI consistently misses unusual paths. An order going from "cancelled" back to "refunded"? The model rarely considers that flow.

Here's the uncomfortable truth about Vibe Coding: it tricks you into thinking you understand what's happening.

The code runs. The feature looks functional. You feel confident. But you might have zero understanding of why it works or when it'll break. That's genuinely dangerous.

My rule now: any AI-written code touching those sensitive areas gets a line-by-line review. Period. Cursor introduced a feature in March 2024 that automatically flags "potentially problematic code sections," but from what I've seen, it's maybe 70% accurate. You can't rely on it.

What's Actually Happening in the Industry

A few things I've been watching:

Cursor's growth is absurd. Anysphere went from zero to $100 million ARR in under two years. For comparison, Slack took three years, Figma took four. I personally know at least five developers who switched from VS Code to Cursor in Q1 2024. I've never seen a tool conversion happen this fast in tech.

Junior developer anxiety is real. Several major tech companies have noticeably reduced their graduate hiring this year—one department I heard about cut headcount by 40%. They're pushing AI-assisted development internally. The uncomfortable maths: work that used to require three junior engineers might now need one senior engineer plus AI. Junior roles won't disappear entirely, but if your entire skillset is "I can write CRUD endpoints," things look dicey. A former colleague of mine was laid off late last year. Took him three months to find a new role, and he ended up at a traditional company doing digital transformation. His own assessment: if he didn't understand the business domain, he couldn't compete on pure coding anymore.

Programming education is painfully behind. I know a few people who teach at coding bootcamps, and they're stressed. The curriculum—hand-writing sorting algorithms, memorising API parameters—is losing relevance fast. But nobody's figured out what the new curriculum should look like. How do you teach students to collaborate with AI? What's actually worth learning now?

I don't have clean answers here. But the fact that most bootcamp syllabi still look like they're from 2022 is... concerning.

My Take: This Isn't a Fad

Plenty of people dismiss Vibe Coding as hype—the next "low-code revolution" that'll fizzle out.

I disagree. Strongly.

Low-code's fundamental problem was that it tried to replace programming with visual drag-and-drop interfaces. But visual interfaces have limited expressive power. Complex logic still required actual code. I tried OutSystems back in 2019, and by the third screen I was stuck—you simply couldn't express the business logic I needed through dragging boxes around.

Vibe Coding is different because natural language is the programming interface. The expressive ceiling is dramatically higher than visual tools. Don't get me wrong—it's immature. Models still hallucinate wildly in complex scenarios. But the trajectory feels different from low-code.

Here's where I think we'll be in three years:

1. AI will shift from assistant to initiator. Right now, humans direct AI. Soon, AI will proactively suggest: "I noticed this module has a performance issue—want me to refactor it?"
2. The barrier to creating software will collapse. Just as you no longer need Photoshop skills to create decent images, you might not need HTML to build websites.
3. A developer's core competency will transform entirely. From "can write code" to "can describe requirements precisely, review AI output critically, and design robust architectures."

When I re-read this article in 2025, I'll probably think I was far too conservative.

What Should You Actually Do?

If you're a developer reading this, here's my practical advice:

Stop resisting and start using these tools. Whether it's Cursor, Copilot, or Windsurf—get AI into your daily workflow. You'll discover tasks where it's 10x faster than you, and tasks where it's utterly useless. Understanding that boundary is the skill. I now open Cursor the moment I start my computer. It's muscle memory.

Go deep in a specific domain. AI can write code, but it doesn't understand your business. If you have genuine depth in a vertical—if you know what breaks in specific scenarios—that judgement isn't getting automated soon. In the e-commerce work I do, things like inventory deduction timing or coupon stacking logic... these are traps you only recognise after stepping in them.

Learn to "train" the AI. The efficiency gap between someone skilled at using Cursor and a complete novice is maybe 5x. How you decompose requirements, how you write prompts, how you iteratively guide the AI—these skills are becoming a distinct competitive advantage. Honestly, I think this might be the most important technical skill of 2024. No exaggeration.

I've got mixed feelings writing this. On one hand, I genuinely enjoy the productivity gains. On the other... if writing code becomes this accessible, what exactly is the value of someone who's spent over a decade doing it?

I think the answer might be: writing code was never the point. Solving problems was. AI just changes how we go about solving them.

What percentage of your daily coding is AI-assisted now? A friend told me he's at 80%, which honestly surprised me—I'm hovering around 40%. Have you had any spectacular failures or surprising wins? I'd love to hear about it in the comments.

ai #programming #softwaredevelopment #cursor #vibecoding #developerexperience