Skip to content

Web products

Vibe coding: what it is, common mistakes and why it is not for everyone

Understand vibe coding, common mistakes when building with AI, and when a business needs review, security and ongoing support.

Author: ZORR 9 min read
Vibe coding: what it is, common mistakes and why it is not for everyone
In this article

Vibe coding is a way to build software, a website or an app through a conversation with AI: a person describes the desired outcome, the tool proposes code, and further requests refine the result. It can produce a prototype quickly. A polished page or successful demo, however, does not establish whether data is protected, enquiries reach the team or someone can maintain the product after launch.

Prepared by the ZORR editorial team with AI assistance; technical statements were checked against the cited sources. The example below is illustrative, not a client project.

What vibe coding means in plain language

Instead of writing every line directly, a person describes a task, such as creating an enquiry form and a services page. AI proposes an implementation; the person opens it and asks for changes. The term became widely known after a 2025 post by researcher Andrej Karpathy and is now used in different ways. It can mean loose experimentation with little code review, or, more broadly, development assisted by AI. Researchers examine the practice in Vibe coding: programming through conversation.

Using AI does not automatically mean abandoning professional development. A developer can use it for drafts, tests or finding mistakes, then review changes and take responsibility for the result. Here, “vibe coding without oversight” means accepting an apparently working result without adequate technical checks.

When the approach can be useful

Experimentation makes sense when mistakes are inexpensive, the data is fictional and the result can be safely rebuilt or discarded. AI can help visualise an idea, compare interface options, prepare a demo or test a hypothesis before commissioning a full build. Speed depends on the task and experience: a quick draft does not guarantee a cheaper or faster final project.

TaskHow to use the approachCheck before launch
Early prototypeTest the flow with fictional dataNo real accounts, customer data or payments are connected
Website with an enquiry formBuild a draft and arrange a reviewMobile layout, delivery, storage and page access
Account area, CRM or paymentsUse AI as a supporting tool for the teamPermissions, data, payment flows, errors and recovery

Why ZORR does not recommend it to everyone

Our concern is uncontrolled launch, not the use of AI itself. The more a business depends on a product, the greater the potential cost of lost enquiries, exposed data or interrupted sales. An online store, account area or CRM integration needs clear requirements, access checks and someone responsible for support. The same applies to human-written code. We recommend choosing the depth of review based on the consequences of failure, not on who or what wrote the code.

OWASP discusses inappropriate trust in AI-generated code as a risk worth addressing and recommends review and testing: OWASP on trust in AI-generated code. This does not mean every AI-assisted project is insecure; the outcome depends on implementation and controls.

Common vibe-coding mistakes

  1. Treating an attractive screen as a finished product. Test the complete path: required fields, actual enquiry storage and notification of the responsible person, including temporary service failures.
  2. Hiding a button and assuming access is blocked. Permissions must be checked on the server: a customer should not be able to access another customer’s order by changing a request. See OWASP access-control guidance.
  3. Putting secrets or real customer data into prompts and tests. Use fictional data, restrict tool permissions and check that keys are not present in public code; revoke and replace a compromised key.
  4. Accepting every change and installing unknown dependencies without review. Inspect changes, package provenance and versions, licences and automated reports of known vulnerabilities.
  5. Testing only the happy path. Try invalid input, duplicate submissions, unavailable CRM or payment services, backup restoration and mobile screens.
  6. Leaving AI-invented prices, reviews or promises in place. Check claims with the business owner; review full Russian and Uzbek conditions, forms and contacts, not just headings.
  7. Ignoring ownership and support. Before launch, establish who controls the domain, code, accounts, database and backups, how to export the project and who handles failures.

Example: an enquiry appears successful but never reaches the team

Imagine a Tashkent business creating a website with AI. A visitor submits a form and sees a success message, but the CRM integration is temporarily unavailable. If the site neither stores the enquiry nor reports an error, the business may lose a potential customer. Acceptance should therefore check the agreed storage, notification, retry and responsible person as well as the on-screen message. This is an illustrative scenario, not a ZORR project.

What to check before publishing

  1. Write down the product purpose, audience and consequences of failure; list data, payments and external integrations separately.
  2. Ask a technical specialist to review the changes, permissions, dependencies, secret handling and environment configuration.
  3. Test normal and failure scenarios on a phone and computer; confirm that a test enquiry is actually delivered and stored.
  4. Verify domain and account ownership, backups, restoration, monitoring and the procedure for reporting a failure.
  5. Check content, language versions and terms with the business. Review indexability, titles and internal links: technical readiness does not guarantee Google rankings.
  6. Agree on a release decision: identify who approves launch and who supports the project afterwards.

For further technical checks, use OWASP guidance on secure coding with AI and W3C form validation guidance. Checklists help uncover issues; they do not replace a review of the specific project.

The real cost of a quick launch

Compare more than the initial screen. Include the tool subscription, hosting, domain, integrations, security review, content, fixes and support over the same period. Ask whether source code and data can be moved to another provider and what the move would involve. There is no universal price or timeline: both depend on scope and the responsibilities agreed.

Our approach: use AI and verify the result

We view AI as a tool that helps explore options and perform part of the work, while requirements, review and responsibility remain essential. A limited prototype may be enough to test an idea. A site receiving real enquiries needs agreed acceptance criteria; payments, accounts and personal data call for technical and legal review appropriate to the actual task.

Continue with our AI-built website checklist and website acceptance checklist. To discuss scope and implementation, see ZORR website development.

Frequently asked questions

What is vibe coding?

It is a way to create and change software through requests to AI and evaluation of the result. The term is used broadly; risk appears when code and product behaviour are accepted without enough review.

Can vibe coding be used for a business website?

Sometimes it is useful for a draft or idea test. A public site with forms, customer data or payments needs checks of its actual features, security and support before launch.

Why does ZORR not recommend vibe coding to everyone?

The cost of failure differs by project. A working demo is insufficient for real data or payments: technical review, clear ownership and a support plan are needed.

Do professional teams use AI in development?

AI can assist with drafts, analysis and tests. A professional process retains requirements, change review and human responsibility for the result.

How can I try vibe coding more safely?

Start with an isolated prototype without real customer data or payments. Before a public launch, agree on checks for workflows, permissions, ownership and recovery.

Web products

Next step: from answer to implementation

Pricing, agency selection, CMS, MVP and mobile products from the first decision to launch.

Let’s discuss
your project

Tell us what you want to do — we’ll suggest an approach and next steps.

+998 91 856 91 96