Skip to content

The practical limitations of vibe coding

An honest account of what AI-written code handles well, where it fails, and the habits that keep a project moving.

Draftship Editorial6 min read

Vibe coding limitations show up in a predictable order. The first small app works, the second feature works, and then a change breaks something nobody can explain. What follows is an honest account of what AI-written code handles well, where it fails, and the habits that keep a project moving when it does.

What the term means and where it came from

Vibe coding means describing software in plain language and letting a large language model write the code, often accepting the result without reading it line by line. Andrej Karpathy coined the phrase in February 2025, describing the practice as giving in to the vibes and forgetting that the code exists. The Wikipedia article on vibe coding traces how the definition, the reception and the criticism have developed since.

The term caught on quickly. Merriam-Webster listed it as slang and trending in March 2025, and Collins named it word of the year for 2025. By then it covered two very different activities: a weekend experiment where nobody cares what the code looks like, and a production system someone has to maintain.

That distinction matters more than any tool comparison. The real content of the definition is the review posture. You guide, you test, you give feedback, and you may not read the code. Every strength and every limitation follows from that.

What works well today

  • Small, self-contained features. A form, a filter, a settings page, a signup email. Requests that fit in one sentence and touch one or two files come back close to right.
  • Prototypes and internal tools. When the cost of a wrong screen is a shrug, generating it beats designing it.
  • Standard web work. Layout, routing, validation, database tables, login, uploads. These patterns are common enough that a model has seen them many times.
  • Glue code. Calling an outside API, reshaping JSON, moving data between two services.
  • Unfamiliar stacks. If you have never written TypeScript and need a working page, the model fills a gap that would otherwise take you days.

Adoption is no longer only hobbyist. Y Combinator reported in March 2025 that a quarter of the startups in its Winter 2025 batch had codebases that were 95% AI-generated. In Draftship, the requests that land reliably are the ones the platform is built for: pages, forms, database tables, sign-in, email, file uploads, and backend functions that call another service. The page on what the AI can and can't build is worth reading before you plan anything larger.

Where it breaks

Changes that spread across the app

A small request touches a page. A large one touches the data model, three pages, an email template and the permissions on a table. The model can produce all of it, but the further a change reaches, the more chances one part drifts out of sync. The failures are quiet: a field saved under a slightly different name, a permission that lets a logged-out visitor read a table, a link to a page that was renamed.

The fix is not a better model. It is smaller changes, applied in sequence, each checked before the next.

Requirements you never pinned down

The model fills gaps with plausible guesses, and a plausible guess is harder to spot than an obvious mistake. Ask for a booking feature and you get a booking feature, well built, without the cancellation rules or the reminders you had in mind, and you notice the missing parts when a customer does. Write the acceptance criteria before the prompt. A few sentences describing what a visitor does and what they should see afterwards is usually enough, and it doubles as a test script later.

Security you cannot see

This criticism has the most evidence behind it. Some developers commit generated code without understanding what it does, which is how undet bugs and security holes reach production. A Veracode study reported in October 2025 found that models had become much better at producing code that works while the security of that code had not improved much. Simon Willison's line, quoted by Ars Technica, is the practical version: vibe coding your way to a production codebase is risky because most engineering work is evolving existing systems, where the quality of the underlying code decides how fast you can move.

Failures in generated apps are rarely dramatic. They look like a table open to anyone, a key pasted into a client-side file, or an admin page that hides the button but not the route. None of that shows up in a screenshot.

Performance and platform limits

A model will happily write the version that works on ten records. Whether it works on ten thousand is a different question. Hosting platforms add ceilings of their own: in Draftship each app has 512 MB of memory and one CPU, schedules run at most every five minutes, and always-on background processes and WebSocket servers are not available. Those limits suit a web app and not video processing, and the model will not stop you from asking.

Anything you cannot verify

If you cannot check the output, you are trusting it. That is fine for a colour change. It is not fine for a payment calculation, a permissions rule or a data migration.

The review problem: code nobody has read

The defining risk is not that the AI writes bad code. It is that nobody reads it. A codebase can grow for months with its author having seen very little of it, and the cost arrives when something breaks and there is no mental model to debug against. Debugging turns into guessing, and guessing in a chat window costs another round of changes.

Two habits reduce this. Read the diff when the change matters, even if you skim. And keep a working version to return to. Draftship's version history saves a snapshot of every successful build and keeps the latest 50, so a bad change is a restore rather than an archaeology project.

Habits that help

  1. One change per message. Bundle three unrelated requests and you cannot tell which one broke the build.
  2. State acceptance criteria. Describe the flow a visitor goes through, not just the feature you want. The testing agent's own example, a visitor submitting an idea and then seeing it in the list, is about the right level of detail.
  3. Name the page or file. On a larger project the AI reads the files a request needs, so pointing at the right place improves the result.
  4. Test the flows, not the look. The testing agent writes tests as plain steps and runs them in a real browser, and can run after every change.
  5. Read the diff before you publish. If you use source control, review the change the way you would review a colleague's.
  6. Ask the AI to explain its security posture. Questions about who can read which table are reasonable, and often revealing.
  7. Paste errors instead of describing them. The console message contains more than your summary of it.

When to bring in a developer

Bring someone in when any of these is true:

  • You handle money, health data or other sensitive personal information.
  • You need something the platform does not do, such as a native mobile app, a long-running process, or a language other than the one the builder supports.
  • You cannot describe the problem well enough to prompt a fix, which usually means you need a mental model of the system first.
  • You are hiring, raising money, or selling, and someone will inspect the code.
  • The app is now load-bearing for a business, and downtime costs more than a review.

The useful framing is not "AI or developer". It is which parts of the app need someone who understands them. A marketing page and an invoice calculation deserve different levels of scrutiny.

Doing it with Draftship

Draftship is built around this way of working: you describe the app, it writes Next.js and TypeScript, and it hosts it with a database and sign-in included. After each build it opens your app in a real browser at desktop and phone widths and spends one round fixing what is clearly broken. Start with something small and specific, and keep the review habits above.

Primary sources

This article was drafted with AI assistance from the primary sources listed below and checked automatically against them. An editor hasn't reviewed it yet.