Is Vibe Coding Bad? I Vibe Coded 170,000 Lines and Read Almost None of It

Is vibe coding bad: paper cutouts of a laptop, phone and smartwatch with dark screens and a coffee mug on a teal cutting mat

Is vibe coding bad? Not on its own. It’s bad when someone else has to live with code nobody read. I’ve vibe coded four apps, about 170,000 lines of Swift, and I’ve read almost none of it. At work I don’t vibe code at all: AI writes the first draft and I read every line before it goes in a PR. The rule is simple. If you’re the only one who pays when it breaks, vibe code away. If anyone else pays, read it.

I have four apps on my phone and Mac that I didn’t really write and mostly haven’t read. Chirp, Scraps, Rova and Goals add up to roughly 170,000 lines of Swift. Claude Code wrote nearly every one of them.

So when someone asks me whether vibe coding is bad, I’m not the guy who gets to say yes. I’m also not the guy who’d let it anywhere near my day job. Both of those are true, and the gap between them is the whole answer.

What Vibe Coding Actually Means

Andrej Karpathy coined the term in a February 2025 post on X: fully give in to the vibes and forget the code even exists. He also said it was fine for throwaway weekend projects. That second part got lost somewhere between his post and every vendor blog that now ranks for this question.

Martin Fowler draws the line I wish everyone used. In his definition, vibe coding means you never look at the code. When AI writes most of the code but you still review it and care about its structure, he calls that agentic engineering, a different practice with different consequences. Half the arguments online are two people using one word for those two jobs.

I do both. Just never in the same place.

The Four Apps I Vibe Coded

Every one of these exists because I didn’t want to keep paying a monthly fee for something I could build myself. Here’s what’s actually running on my devices:

  • Chirp: a Mac app for dictation, meeting notes, spelling and grammar checks, read aloud and summaries, all on device. About 61,000 lines. It’s free and open source on GitHub.
  • Goals: a goal tracker for iPhone, iPad, Mac and Apple Watch. About 51,000 lines, and it’s on the App Store with a paid tier.
  • Rova: my Strava replacement, a cardio journal that pulls my runs in from Apple Health. About 32,500 lines.
  • Scraps: Home Screen widgets and wallpapers that look like paper cutouts on a desk. About 24,700 lines.
Paper cutouts of worn running shoes, a phone, a smartwatch and a laptop with a microphone, standing for four vibe coded apps

My job on all four is product owner and QA. I decide what the app should do, I test it until it does, and I write down what went wrong. Every repo has a LESSONS file of gotchas the project already paid for, plus standards docs Claude reads before it touches anything. That’s how I steer code I don’t read. I don’t fix the code. I fix the instructions.

Is Vibe Coding Bad? Only When Someone Else Pays for It

Vibe coding isn’t bad on its own. It’s bad when someone else has to live with code nobody read. For a tool only you use, letting AI write code you never open is fine, because you’re the one who pays when it breaks. For code a team maintains or customers rely on, someone has to read and own every line.

The security numbers are why. Veracode’s 2026 GenAI Code Security Report found roughly 44% of AI code generation tasks introduced a risky vulnerability, and the average security pass rate sits at 56%, barely changed from 55% in its first report. The code compiles. It just isn’t safe by default, and it isn’t getting safer on its own.

Goals is my honest gray area, because strangers pay for it. My answer there is testing instead of reading. I use it every day and test it hard before anything ships. That’s not the same thing as a code review, and I know it.

The Bug That Showed Me What I Wasn’t Reading

The settings toggles in Goals started switching themselves off. Flip one on and it flipped back. Sometimes it reset a different toggle entirely. Tap a few times fast and the whole screen looked dead.

The cause was a pattern Claude had repeated across the app: each @AppStorage was built on its own inline UserDefaults(suiteName:), so every redraw made a fresh copy of the settings store. That fresh copy could read a stale value before my last write landed, then save the stale value right back over it. The fix was one shared instance used everywhere, plus an entry in the LESSONS file so it never happens twice.

That bug never got caught in code review, because there wasn’t one. In my own app, I’ll eat that. On a team that trusts my PRs, that’s exactly the bug I’m paid to stop.

Why I Don’t Vibe Code at Work

A magnifying glass cutout over a dark keyboard with reading glasses and coffee, standing for reviewing every line of code

When I started at my current company, I didn’t use Claude at all for the first three months. Not because it couldn’t write the code. Because I wasn’t learning the domain, and code I don’t understand is code I can’t defend in a review.

There’s research behind that gut feeling. In a randomized trial Anthropic ran with 52 mostly junior engineers, the group coding with AI averaged 50% on a follow-up quiz against 67% for the group coding by hand, and it wasn’t meaningfully faster. The people who scored well used AI to ask questions and explain code, not just to spit it out.

Now AI writes most of my first drafts at work too. The difference is that nothing goes into a PR I can’t explain line by line. My name is on it. If it breaks, nobody cares that Claude wrote it.

I made the bigger career case in Is Software Engineering Dead?, and the case for new developers in Will AI Replace Junior Developers? Same root idea: the job is understanding, not typing.

Where Vibe Coding Falls Apart

Testing only replaces reading if you know what right looks like. I’ve built software for more than ten years, and that’s doing a lot of quiet work every time I decide a feature is done. If you can’t tell when the app is wrong, you’re not vibe coding. You’re hoping.

It also piles up. Fowler’s warning is that unreviewed code gets harder to change, for people and for the AI. My LESSONS files exist for exactly that reason, and along with the tests, they’re what keeps 170,000 unread lines from turning into soup. If you’ve never coded and want to try this anyway, start with what Claude Code is actually good at for non coders.

Pick a Side of the Line

Here’s your challenge. Pick one subscription you’re paying for this month and vibe code a replacement this weekend. Don’t read the code. Just use it until it works.

Then open your last PR at work and ask whether you could explain every line out loud. If the answer is no, that’s the real problem. And AI didn’t create it.