Blog

Insights

Is Vibe Coding Bad? Not With the Right Guardrails

Is vibe coding bad? No. Most problems come from skipped steps, not the AI. See the data, where vibe coding works best and 7 guardrails that keep apps safe.

Writer

Nafis Amiri

Co-Founder of CatDoes

Is Vibe Coding Bad? Real Risks and When It Works — minimalist illustration featuring a coding laptop, red warning icon, and green checkmark representing the risks and benefits of AI-assisted coding on a white grid background.

Is vibe coding bad? No. Vibe coding is a fast, legitimate way to build apps, and AI now writes code for most developers at work. The problems people point to come from skipped steps, like database access rules or a rollback point, not from the AI. With a few guardrails, vibe coding works for real products, not just prototypes.

This post is for founders, non-developers and developers deciding whether to build a real project with an AI agent. It covers the common concerns with the data behind each one and the fix for each. We build an AI agent that writes apps, so the guardrails section comes from what we ship.

TL;DR: Vibe coding isn't bad. Stack Overflow's 2026 survey found 83% of developers use at least one AI tool at work. The concerns critics raise (security gaps, messy code, rising costs) are real but avoidable, and each has a simple fix. Vibe coding is a great fit for prototypes, internal tools and websites. For apps with logins, payments or personal data, add guardrails: a plan, database access rules, keys kept out of the code, a rollback point and a check before launch. Platforms like CatDoes build several of these in for you.

Table of Contents

  • What Vibe Coding Means, and Why the Definition Matters

  • Why Do People Say Vibe Coding Is Bad?

  • When Vibe Coding Works Best

  • How to Vibe Code Safely: 7 Guardrails

  • What to Do When Your App Gets Hard to Change

  • How CatDoes Builds the Guardrails In

  • Is Vibe Coding Bad for Beginners?

  • Is Vibe Coding the Future, or Is It Dead?

  • Common Questions About Vibe Coding

  • The Verdict: Vibe Coding Is Good With Guardrails

What Vibe Coding Means, and Why the Definition Matters

Andrej Karpathy coined the term in a post on X on February 2, 2025. He described a new way of building where you "fully give in to the vibes" and "forget that the code even exists." You describe what you want, and the AI writes the code.

The word spread fast. Collins Dictionary made "vibe coding" its Word of the Year for 2025, defined as using AI prompted by natural language to write computer code. That broad meaning covers almost all AI-assisted building, which is why people argue past each other.

Most "is vibe coding bad" debates are really about how the work gets checked:

Style

How the work gets checked

Risk for a real app

Unchecked vibe coding

Nobody checks. Whatever the AI writes ships

High

Vibe coding with review

AI writes it, a person reviews and tests it

Low

Vibe coding platform with guardrails

The platform handles logins, keys, rollback and checks before launch

Low for most apps. Add an expert review for payments or health data

Karpathy's own view has grown with the tools. In February 2026 he said his favorite name for it now is "agentic engineering."

Why Do People Say Vibe Coding Is Bad?

Critics point to five concerns. Each one has data behind it, and each one has a fix:

Concern

What the data shows

How to avoid it

Security gaps

About 44% of raw AI code generation tasks introduced a vulnerability (Veracode, 2026)

Access rules, keys out of the code, a security check

Messy code

10.83 issues per AI-co-authored pull request vs 6.45 for human-only ones (CodeRabbit, 2025)

Small steps, review, refactoring

Rising costs

Agentic tasks use about 1,000 times more tokens than code chat (Bai et al., 2026)

Plan first, match the model to the task

Changes to live data

Agents with write access can change or delete real data

Test data, approval before risky changes

Checking the result

48% of developers trust AI output when they can easily verify it (Stack Overflow, 2026)

Test flows in preview, ask the agent to explain

1. Security: AI code needs a check

Veracode tests AI models on coding tasks and scans their raw output. In its July 2026 report, about 44% of tasks introduced a risky vulnerability. That's the case for guardrails: raw model output needs a check before it ships.

Illustration of an app with its database left unlocked, the mistake that database access rules prevent

Real leaks often trace back to a single setting. In February 2026, security firm Wiz found the database behind the Moltbook app open to anyone, with 1.5 million API keys exposed. Row Level Security, a database rule that limits each user to their own data, was turned off. Our vibe coding security checklist shows how to check your own app.

2. Code quality: review keeps it clean

CodeRabbit compared 470 open-source pull requests. The AI-co-authored ones had 10.83 issues each, against 6.45 for human-only ones. Readability issues were more than 3 times as common, the kind of problem a review pass is built to catch.

GitClear's June 2026 study looked at 623 million code changes. Copy-pasted lines rose to 15.7% in the first half of 2026, up from 9.4% in 2022, as AI tools spread. The fix is simple. Ask the agent to clean up and reuse code as you go, and build one feature at a time so each change stays easy to check.

3. Cost: planning keeps it predictable

Agents use more tokens than chat because they do more of the work. A study by Bai and colleagues (April 2026) found agentic tasks use about 1,000 times more tokens than code chat. Runs of the same task varied by up to 30 times.

You can keep that in check. Agree on a plan before the agent builds, because changing a plan is cheap and changing built features isn't. Use a lighter model or tier for small edits and save the strongest one for planning hard features.

4. Live data: agents should ask first

An agent that can write to your live database can also change or delete what's in it. The fix is to build against test data and approve anything destructive before it runs. Good platforms make the agent ask for permission before it touches your backend.

5. Checking work you didn't write

You don't have to read code to check an app. Stack Overflow's 2026 survey found 48% of developers trust AI output "when I can easily verify the output," and you can verify a lot without reading code.

Click through every main flow in the preview. Log in as two test users and make sure neither can see the other's data. Ask the agent to explain each change in plain English. For apps that take payments or store health data, add an expert review before launch.

When Vibe Coding Works Best

Vibe coding fits almost any project. What changes is how many guardrails you add before launch:

Illustration of a vibe-coded prototype in a sandbox next to the same app as a secured production building with a lock, shield and database

Project

Vibe code it?

What to add before launch

Prototype or demo

Yes

Nothing. Build, test the idea, iterate

Personal tool

Yes

Keep API keys out of the code

Landing page or marketing site

Yes

Check forms and any keys

Internal tool for your team

Yes

Login and access rules

Public app with user accounts

Yes, with guardrails

Auth, database access rules, a rollback point, a security check

App Store or Google Play app

Yes, with guardrails

The above, plus a check against the store review guidelines

Payments, health or children's data

Yes, with an expert review

A payment provider like Stripe, a review before launch, privacy compliance

Prototypes are the easiest win. You can test whether anyone wants an idea in an afternoon instead of a month. If they do, the same project can grow into the real product once you add the guardrails.

If your project is a mobile app, our guide to vibe coding a mobile app without security holes covers the store-specific checks.

How to Vibe Code Safely: 7 Guardrails

None of these needs you to be an engineer:

  1. Plan before you build. Ask the agent for a written plan and approve it before any code exists. Changing a plan costs nothing, while changing built features costs time and credits.

  2. Work in small steps. Ask for one feature per prompt, and keep a checkpoint or Git commit you can return to.

  3. Add login and access rules before real data. Ask the agent to make sure each user can only see their own data, then test it with two accounts.

  4. Keep keys out of the code. Paste API keys into your platform's secure key fields or environment variables, never into the chat or the app's front end.

  5. Keep agents away from live data. Build against test data, and approve anything that deletes or changes real records.

  6. Check before launch. Test the main user flows yourself and ask the agent for a security review. Our security review prompt is a good place to start.

  7. Reset after three tries. Engineer Waleed Kadous uses this rule: if the AI hasn't fixed a bug in 3 attempts, change the approach. Roll back, describe the problem differently, or ask for a plan first.

If you only do two, do 3 and 5. They prevent the two most damaging mistakes: other people's data exposed, and real data lost.

What to Do When Your App Gets Hard to Change

Every growing app reaches a point where new features start to break old ones, whether people or AI wrote the code. With vibe coding, the signs are easy to spot: the agent keeps working on the same bug, or a fix in one place breaks another.

Illustration of a person facing a wall of tangled code while an AI assistant holds a rewind arrow, showing how to get a vibe-coded app back on track

When that happens, pause new features for a moment and do this:

  1. Roll back to the last version that worked.

  2. Ask the agent to write a plain-English summary of how the app is built: screens, data, logins and outside services.

  3. Have it add tests for the 3 to 5 flows users rely on most.

  4. Ask it to clean up the parts that changed most, one area at a time.

  5. Keep the code in GitHub or export it, so a developer can review it if you ever need one.

How CatDoes Builds the Guardrails In

We build CatDoes, an AI agent that builds mobile apps and websites from plain English. Our goal is to make the guardrails above the default, so you don't have to set them up yourself. Here's what the product does today, per our docs:

CatDoes docs page explaining plan mode, where the agent plans a feature and waits for approval before building
  • Plan mode: tell the agent "Don't build anything yet" and it shares a plan, then waits for your go-ahead. As our docs put it, "Changing a plan is free. Changing built work costs credits."

  • CatDoes saves a checkpoint before the agent works on each message, so you can rewind any change. Data already saved in your database isn't reverted, though.

  • The agent asks for approval before it changes data in your database, deletes a function or disconnects your backend.

  • API keys live in encrypted environment variables. Values typed into secure fields are never visible to the AI.

  • Before you submit to the App Store or Google Play, the agent can check your app against the review guidelines and list blocking issues.

  • Watch groups runtime errors into issues, and "Fix now" sends one to the agent.

  • On the Plus plan and above you can export the code or sync it with GitHub, so you always own it.

CatDoes docs page explaining checkpoints, which save snapshots of a project so you can roll back changes

Together these cover several of the guardrails above without extra setup. For apps that take payments or store health data, a review before launch is still a good idea on any platform.

Is Vibe Coding Bad for Beginners?

No. Vibe coding is a great way for beginners to start, because you see a working app in minutes instead of weeks. It also teaches you how software fits together if you stay curious about what the agent builds.

To learn faster while you build:

  • Ask the agent to explain each change in plain English.

  • Ask why it chose one approach over another.

  • Look at how your logins and database rules work, even if you never write them yourself.

Is Vibe Coding the Future, or Is It Dead?

Vibe coding isn't dead. It's becoming the normal way a lot of software gets written.

The numbers point that way. Stack Overflow's 2026 survey found 83% of developers use at least one AI tool at work, and 65.9% use coding assistants or agents. Karpathy wrote in January 2026 that he went from about 20% agent coding in November 2025 to 80% in December.

The direction is AI writing more of the code inside guardrails: plans, tests, access rules and checks before launch. That's close to what Karpathy now calls "agentic engineering," and it's where vibe coding platforms are heading.

Common Questions About Vibe Coding

Is vibe coding a trap?

No. Vibe coding only gets frustrating when an app grows without a plan and every change touches everything. You avoid that by planning features before the agent builds them, working in small steps with rollback points, and changing your approach if a bug survives three attempts.

Why did vibe coding fail?

Vibe coding didn't fail. It became Collins Dictionary's Word of the Year for 2025, and 83% of developers now use AI tools at work. The vibe-coded apps that do run into trouble usually skipped the same steps: database access rules, keeping keys out of the code, or a rollback point.

Is vibe coding ethical?

Yes. Vibe coding is a way to write software, and the same responsibilities apply as with hand-written code. Protect your users' data and test what you ship. If you build for clients, tell them how you work. Used that way, it lets more people build useful software.

Do vibe coders get jobs?

Yes, and AI skills are in demand. Indeed found US software development job postings grew almost 15% from February 2025 to mid-2026, while all postings fell 7%. About 37% of that growth came from jobs with AI in the title. Employers want people who can build with AI and check the results.

Will vibe coding replace programmers?

No. It changes what programmers spend their time on: more planning, reviewing and testing, and less typing. It also lets people who don't code build their own apps. Demand for developers is holding up: Indeed found US software development postings grew almost 15% from February 2025 to mid-2026.

Is vibe coding bad for the environment?

A single AI prompt uses little energy. Google measured its median Gemini text prompt at 0.24 watt-hours. Agent sessions use more because they do more of the work. Planning first and using a lighter model for small edits cuts both energy use and cost.

The Verdict: Vibe Coding Is Good With Guardrails

So, is vibe coding good or bad? It's good, and it keeps getting better. It's the fastest way we know to turn an idea into a working app, and the concerns critics raise all come with a known fix.

Vibe code your ideas freely, and add the guardrails before real users arrive. If you want an agent that plans before it builds and keeps a checkpoint of every change, try building your app with CatDoes.

Writer

Nafis Amiri

Co-Founder of CatDoes