Vibe Coding in Production: Where It Works and Where It Fails
Vibe coding in production carries real risks. Learn where it works, where it fails, and how to review, harden, or rebuild a prototype before real users arrive.
How this article was made: drafted with AI assistance, then reviewed and approved by the StackLift team before publishing.
Vibe coding, a term popularized by Andrej Karpathy in early 2025, describes building software by prompting an AI and accepting what it produces without closely reading the code. It is a real shift: people can now build a working demo in an afternoon. The trouble starts when that demo gets real customers. This article looks at vibe coding in production, where it holds up, where it breaks, and what a prototype needs before people depend on it.
What this article makes clear
- Vibe coding is well suited to prototypes and low-risk internal tools, where failure is cheap.
- The main production risk is that nobody understands the code well enough to secure, fix, or change it.
- Before launch, require human review, server-side security, tests, dependency checks, backups, and rollback.
Where vibe coding genuinely works
Vibe coding is excellent for answering questions quickly. Will this workflow make sense to users? Is this idea worth funding? What should the interface feel like? A clickable prototype answers those questions far better than a slide deck, and it costs little enough that you can throw it away without regret.
It also suits small internal tools with a limited blast radius: a script that formats a report, a dashboard over non-sensitive data, or a one-off migration helper that a technical person will watch while it runs. If the worst outcome of a failure is a mild inconvenience, moving fast with light review is often the right call.
Where it fails once real users arrive
The core problem is understanding. When nobody on the team knows how the code works, nobody can reliably fix it during an incident, estimate the next change, or tell whether a new feature has quietly broken an old one. Each prompt can also rewrite parts of the system that were working, and without tests, you find out from your customers.
Security failures tend to be the most expensive. Generated applications can ship with checks enforced only in the browser, missing authorization on server endpoints, API keys in client code, or database rules left wide open. The app looks finished and works in a demo, because these flaws do not show up on the happy path. A person has to go looking for them.
The risks that are easy to miss in a demo
Dependencies deserve particular attention. AI tools can suggest outdated packages, or name packages that do not exist, and attackers have begun registering those invented names so an unwary install pulls in malicious code. Check every new dependency, and pin versions so builds are repeatable.
Operational gaps are the other blind spot. A prototype usually has no backups, no logging, no error tracking, no staging environment, and no way to roll back a bad release. None of that matters until the first outage or the first request to restore a customer's deleted data. By then, it matters a great deal.
A checklist before a vibe-coded app meets real data
First, get a person to read it. Have an experienced engineer walk through the architecture, the data model, and the authentication flow, and require them to be able to explain each one. Second, secure it: enforce permissions on the server, move secrets to a proper store, rotate any keys that may have been exposed, and run a dependency and vulnerability scan.
Third, make it safe to change. Add automated tests around the workflows that matter most, set up continuous integration, and put a human review step in front of every merge. Fourth, make it operable: add logging, error tracking, backups you have actually tested restoring, and a clear rollback path. If you handle regulated or personal data, get specialist advice on those obligations before launch.
How to decide whether to harden or rebuild
Sometimes a prototype is worth keeping. If the structure is sensible, the data model reflects the real business, and an engineer can follow the code, hardening it incrementally is usually the fastest route. Keep what works and add the missing safeguards around it, one risk at a time.
Rebuild when the foundations are the problem: a data model that does not fit the business, authorization that would have to be retrofitted everywhere, or code nobody can explain after a serious attempt. A rebuild is much faster than the original guesswork would suggest, because the prototype has already answered the most expensive question, which is what to build.
Keeping AI speed with engineering control
The choice is not between AI-written code and careful engineering. Teams can use agents for much of the typing while senior engineers own the design, the review, and the release. What changes is who is accountable, and that does not move to the tool.
That is the model StackLift AI works to: agents accelerate the work on feature branches, they never push to the default branch or touch production secrets, and senior engineers own every decision. If you have a prototype that is starting to attract real users, an architecture review is a sensible first step before you scale it.
Frequently asked questions
Common questions on this topic.
What is vibe coding?
Vibe coding means building software by describing what you want to an AI tool and accepting the generated code, without closely reading it. Andrej Karpathy popularized the term in early 2025. It differs from using AI as an assistant, where an engineer still reviews and understands every change.
Is vibe-coded software secure?
It can't be assumed to be. Generated code may run correctly while still missing server-side permission checks, exposing keys, or using risky dependencies. These flaws rarely appear in a demo. Treat it as unreviewed code and have an experienced engineer assess its security before it handles real data.
Can a vibe-coded prototype become a real product?
Yes, often. Its most lasting value is usually the proof that people want the product and clarity about what to build. Whether the code itself survives depends on its foundations. An architecture review will show whether to harden it step by step or rebuild it using what you learned.
Do we still need senior engineers if AI writes the code?
Yes. Writing code was never the only hard part. Someone must still choose the architecture, judge trade-offs, review changes, handle incidents, and answer for the system. AI can reduce the typing, but production software still needs a person who understands it and owns the outcome.
Apply this to your project
Turn your idea into an architect-reviewed delivery plan.
Use the StackLift estimate flow to convert your requirements into scope, story points, risks, timeline, and a practical delivery baseline.
Start an estimate