All posts
AI & Engineering

Is Vibe Coding Bad? A CTO's Honest Verdict for Founders

Vibe coding is fine for testing ideas. Whether it's risky depends on what your app touches and who answers for it when it breaks. Here's a CTO's honest verdict.

Nikhil Garg · CTO, Get Cover12 min read
Is Vibe Coding Bad? A CTO's Honest Verdict for Founders

Is vibe coding bad? No, not as a way to build. It becomes bad when the product handles other people's money or data and nobody understands the code the AI wrote. The typing method doesn't decide whether your product is safe. Two things do: what the app touches, and who answers for it when it breaks. Vibe coding isn't bad. Shipping something nobody owns is. Most of the noise online skips both. I'll give you my verdict as the person who would own the outcome, and then a way to apply it to your own product.

So, is vibe coding bad?

Andrej Karpathy coined the term in a post on X on February 2, 2025. He meant building by describing what you want to an AI, accepting what it produces, and not really looking at the code. By November, Collins named "vibe coding" its Word of the Year for 2025. Something that popular attracts a lot of opinions, and most of them are about something other than your business.

Here is my verdict. Vibe coding is a very good way to find out what to build. It is a risky way to run a product once real users trust it with their accounts, their payments or their personal information. The risk is not the AI. The risk is that the code has no owner: nobody who has read it, understands it and will answer for it at 2 AM.

So the useful question isn't whether vibe coding is good or bad. Ask what your app touches today, what it will touch in three months, and whether anyone can explain the parts that matter. I use AI every day in my own work, and a real share of that work is rescuing apps built with these tools. Everything below is how I tell founders where the line sits.

Why do developers hate vibe coding so much?

Read the Reddit and Hacker News threads and you'll notice most of the anger is about developers' careers, not your product. That's fair for them. But only some of it applies to you.

The complaints, sorted:

  • "Skills will fade and juniors won't learn." A real worry for the profession. It doesn't matter to you. You never planned to become an engineer.
  • "Vibe coders aren't real developers." This is about status. Ignore it. Your customers don't care how the code was typed.
  • "We're the ones who get called to clean it up." This one matters. It means a developer may one day inherit your code. How easy it is for them to take over decides what your next six months cost.
  • "Nobody understands what was shipped." This matters most, and it's the subject of the rest of this post.

To be fair to both sides, a lot of that anger comes from the people who get asked to fix this code later. And choosing AI tools to move faster is a business decision. It isn't a moral failing.

Is coding even worth it anymore?

For a developer, yes. The job is shifting from typing code to judging it, and judgment is the scarce part. For a founder, learning to code is rarely the best use of your time. Learning enough to ask the right questions is.

What are the critics right about?

The critics are right about the things that cost founders money.

AI writes insecure code more often than it looks

It is getting better, slowly. In Veracode's Fall 2026 GenAI Code Security report, published in October, the average security pass rate across 18 AI models rose to 62 percent, from 56 percent: the first real improvement in its series. That still means roughly 38 percent of the coding tasks produced code with a known security flaw, and the best model passed 73 percent of the time. The code worked. It just wasn't safe. A founder can't tell the difference by clicking around the app, because both versions look the same on screen.

It happens in public, too

In early 2026, the security firm Wiz looked at Moltbook, a social network for AI agents. Its founder had said publicly that he didn't write a single line of code for it. Wiz found the database key in the page's own JavaScript, with no access rules behind it, so anyone could read and write the production database. That exposed 1.5 million API keys, 35,000 email addresses and more than 4,000 private conversations. It was fixed within about three hours of the report, and Wiz traced it to a single database setting. Its own conclusion wasn't to stop vibe coding. It was to raise the standard.

The AI says "done" when it isn't

AI tools build the happy path well: sign up, pay, use the feature. They're weaker on the paths a business actually lives on. What happens when a card fails, a customer asks for a refund, a subscription lapses halfway through the month, or two people edit the same record at once. These aren't rare cases. Once you have paying customers they happen every week.

I saw it with Satsman, a Bitcoin education platform built with Replit's AI. In the demo, it was fine. Once real customers arrived, it started breaking far too soon, and nothing in Replit warned anyone.

Speed drops as the codebase grows

The first month feels magical. By the third, each fix breaks something else, because nobody designed how the pieces fit together. Every prompt patches what it can see. I've written about how to tell if AI generated code is any good if you want the warning signs in detail, and about what to do when you hit the ceiling on Bolt.

Nobody can explain it

This is the critics' best point. When the person who "built" the product can't explain how login or billing works, every incident becomes an investigation. Only the AI knows how the code works, and it doesn't keep that knowledge from one session to the next.

Where are the critics wrong?

They're wrong in three places, and these are the ones that matter to you.

First, they assume AI-built code is worthless. It isn't. Even when the code needs heavy work, you've already answered the expensive questions: which screens customers use, which flows they abandon and what they'll pay for. Those product decisions keep their value whether the code is fixed, refactored or rebuilt. A good engineer can rebuild from a working product much faster than from a document, because the product is the spec.

Second, they blame the tool. Human developers also ship insecure code, and plenty of agency code is worse than what a careful founder gets from an AI. A missing database permission is equally dangerous whether a person or a model left it out. The real gap is that no qualified person reviewed it. Experienced engineers use these tools every day. I do too. I use AI to research ideas, and it's a lot better at it than I am. But the final call is always mine.

Third, they treat "vibe coded" as a single category. A landing page with a waitlist and a platform holding customer payment records are not the same risk. The next section covers that difference.

Is vibe coding great?

For some jobs, honestly, yes. It's great for testing an idea in days, getting a clickable product in front of customers, building internal tools that only your team uses, and showing investors something real. I've written about when shipping the AI version is actually the right call, and sometimes it is. It is not great as the long-term foundation of something that holds money or data, unless someone takes ownership of it. If you're still choosing a tool, my ranking of the best vibe coding tools for founders scores each one on what you own once the demo works.

Is it bad for what I'm building?

Draw the line by what the app touches, not how it was made. I think of it in three levels.

Level 1: a prototype or demo

No real user accounts, no payments, no personal data beyond a waitlist email. Vibe code it. The worst case is that it breaks in a demo. You don't need me or anyone like me here, and paying for a review would waste your money.

Level 2: real user accounts

People sign up, log in and store things. Now the questions change. Can one user see another user's data by changing a number in the address bar? Does password reset actually check who's asking? Are the keys that unlock your database sitting in code that anyone's browser can read? These are the most common failures I see in AI-built apps, and they're invisible until someone goes looking. If you built on Lovable, I've covered this in whether your Lovable app is secure enough to launch.

Level 3: money, personal data, or several business customers

You take payments, you hold personal or regulated data (health, financial, identity), or you sell to businesses and keep several companies' data in one database. Here, vibe coding without an owner is a real business risk. One company seeing another company's records can end a contract and sometimes a business. Money that is charged and never recorded, or refunded twice, is a problem you have to explain to customers.

When I built the Get Cover warranty platform, the hardest work wasn't the screens. It was keeping each business customer's data separate inside one multi-tenant platform, and getting claims processing right. That platform has since processed over $1M in warranties and more than 1,000 claims. None of that work is visible in a demo, and it's exactly the work AI tools skip unless someone insists on it.

And here's the part nobody mentions: the level moves with your business, not your code. Your first payment, or your first business customer, moves the same app up a level overnight.

I already built mine with AI. What should I check?

You don't need to read code to run most of these. You need to ask direct questions and get direct answers.

  1. Do you own it? Is the code in a GitHub account you control? Are the hosting, database, domain and payment accounts in your name, with your email as the owner? If the answer is "it lives in the tool," find out how to export it before you need to.
  2. Can one user see another's data? Sign up as two test users, in two different browsers. Log in as one and try to reach the other's records by changing IDs in the address bar. If you sell to businesses, try it across two test companies.
  3. Are secret keys exposed? Ask your tool to list every key it uses and where each one lives, then have someone check that none of them sit in code that runs in the browser. Exposed keys mean strangers can run up your bills or read your data.
  4. Do failed, refunded and disputed payments work? In test mode, use a card that declines, issue a refund, then open a test dispute. Does your app show the right status afterwards, or does it still say "paid"?
  5. Has a backup ever been restored? Having backups is not the same as having tested them. If nobody has restored one, assume you can't.
  6. Can anyone besides the AI explain login and billing? If the only answer is "ask the AI again," then nobody owns the two most important parts of your product.

Then comes the decision the online threads never explain: fix, refactor or rebuild. Fix when the structure is sound and the problems are specific holes. Refactor when it works but each change takes longer than the last. Rebuild when the data model itself is wrong, for example when business customers' data was never separated properly. A rebuild still keeps everything you learned about your customers. This is most of what I do in AI codebase rescue: deciding what to keep, what to repair and what to replace.

When does my AI-built app need an owner?

The risk changes over time. It goes up at specific business moments, and those moments are your signal:

  • Your first payment. Money now moves through code nobody has checked.
  • Your first personal or regulated data. A breach now harms real people, and you're the one responsible.
  • A business customer's security questionnaire. They'll ask how you separate customer data, who has access and how you recover from failure. "The AI handled it" isn't an answer they'll accept.
  • Raising money or selling the company. Technical due diligence will look at the code, and surprises there lower the price.
  • Handing the code to your first developer. They need someone to tell them what's safe to change.
  • The fix-one-thing, break-another loop starts eating whole weeks. That's the codebase telling you nobody designed it.

An owner isn't a better prompt or a different tool. It's a person who has read the code that matters, makes the calls on what to keep and change, and answers for the result. Sometimes that's a strong developer you trust (I compared the two routes in vibe coding vs hiring a real developer). Sometimes it's a fractional CTO who leads your team, or does the build if you don't have one. When Satsman outgrew Replit, the job wasn't rewriting for the sake of it. It was moving the product onto production-grade AWS with proper deployment, scaling and security, and owning that move. It now serves over 40,000 users, and more than $43,000 has gone through it.

If you're at Level 1, keep building and enjoy it. If you're at Level 2 or 3 and can't answer the six questions above with confidence, have someone review it before it takes its first payment or holds personal data. That's what my Technical Health Check is for: a fixed fee, a written report within five business days of getting access, and a clear call on whether to fix, refactor or rebuild. If it's solid, the report says so.

Vibe code all you like. Just make sure somebody owns it.

  • Vibe Coding
  • AI & Engineering
  • Non-Technical Founders
  • Security
  • Technical Leadership

Get the next one in your inbox

Practical writing for non-technical founders building software. One post a week, no pitching, unsubscribe whenever.

No spam. Unsubscribe any time. Prefer RSS?