Vibe Coding App: Lessons and Technical Debt Risks
A vibe coding app can ship fast, but speed can hide risks. Learn the lessons from a CEO's son's project and how non-coders can test AI-built apps.
On this page
A vibe coding app is one you describe in plain language while an AI writes the code for you. The approach is fast, and it is getting more common among founders, freelancers and small teams who want a working tool without hiring a developer. A recent Fortune story about Zendesk's CEO and his son gives a useful, concrete picture of both the appeal and the pitfalls. This article walks through that case and turns it into practical steps for anyone building software without writing code.
What happened: a CEO and his son vibe code an app
According to Fortune, Zendesk CEO Tom Eggemeier had been telling his children since 2023 that AI would matter enormously and that they would fall behind without it. His son Will, a Boston College senior and basketball player, took him up on it. Earlier this year they started building a predictive analytics app called Underdog AI. The goal was to help college athletics programs spot players in the NCAA transfer portal before better-funded schools got to them. Tom helped design the interface.
The early stage was easy. Tom described the process as vibe coding, and he said it was exciting because new features kept appearing. Then Will noticed that the AI was making recruitment suggestions that did not sound right. Fast development had also created technical debt in areas like security and API protocols, and the pair spent hours debugging.
The outcome was humbling. Two college basketball general managers told them the app matched about 70% of what was already for sale. Will also cannot sell to rival schools while he is on the Boston College roster, so he is reworking the project until he graduates. The story does not present this as a failure. It presents it as a clear lesson about what the first version of an AI-built app can and cannot do.
Vibe coding technical debt: why speed without structure costs you later
Technical debt is the cost you pay later for shortcuts taken now. In traditional software, a developer might skip writing tests or hard-code a value to meet a deadline. The shortcut works today, but every change becomes harder and riskier over time. Vibe coding can build up that kind of debt quickly, because the AI produces working-looking code faster than most people can review it.
The Fortune piece makes a point that matters for non-coders: there are no real shortcuts to software that is seamless and secure. An app can look finished in a live preview and still have weak access rules, fragile connections between parts, or logic that breaks on unusual input. You may not see those problems until a real user finds them.
This does not mean vibe coding is a bad idea. It means the speed is most useful when you pair it with checks. The rest of this article covers the checks that matter most for someone without a technical background.
Warning signs a non-coder can spot
You do not need to read code to notice when an AI-built app is drifting. Watch for these signs:
- Outputs that feel wrong. Will noticed the recruitment suggestions did not sound right before anyone formally tested them. If a recommendation, score or summary surprises you, stop and ask why.
- Inconsistent data. A record that shows up in one view but not another, or a total that changes when you refresh, usually points to a problem in how information is stored or passed along.
- Fragile logins. If you can reach a page you should not see when you are logged out, or if sessions expire at random, access control needs attention before anything else.
- Features that work only on the demo path. If the app behaves well with your own example and fails with a slightly different name, date or file, the logic is not yet robust.
- Changes that break older parts. When a new feature causes an unrelated screen to stop working, you are seeing debt come due.
When you see any of these, describe the problem to your builder in specific terms, save a copy of the current version, and test again before moving on.
The vibe coding lessons that apply to builders
Tom Eggemeier said the experience made him "a better leader" for his product and R&D teams. Several of his takeaways transfer directly to people building their own tools.
Use the technology yourself. Tom used an AI agent to renegotiate his insurance, internet, mobile plan and subscriptions, and he said it saved him $3,150 a year. He also trained an AI tool to write in his voice and noticed that it kept ignoring his instruction to avoid contractions. That repeated failure led him to think harder about how memory and skills affect AI products. For a builder, the lesson is simple: the more hands-on time you spend with a tool, the faster you learn where it breaks.
Put AI to work on real problems. Tom's advice was to create projects that solve concrete issues, not to chase features for their own sake. Underdog AI had a clear, narrow purpose. That focus made it easier to judge whether the output was useful.
Watch expenses, time and complexity. Tom framed good AI use around cutting costs, saving time and reducing complexity. Before you add a feature, ask whether it removes a task, a cost or a source of confusion. If it only adds another screen, it may be adding debt.
Stay curious about failure. The most useful moment in the story came when something seemed off. Will did not dismiss the odd suggestions. He investigated them. Treat strange output as information.
How to test an AI-built app before people rely on it
Before you share an app with customers, partners or your team, run it through a simple test process.
- Test with real data. Use the actual spreadsheet, list or documents you will use day to day. Demo data is usually too clean.
- Try edge cases. Enter blank fields, very long names, duplicate entries, dates in the past and unusual characters. Note what breaks.
- Check outputs against a human expert. The two basketball general managers in the Fortune story exposed a gap that the builders had missed. Ask someone who knows the field to review the results and tell you whether they are accurate and useful.
- Check access. Log out and try to open each page directly. Ask a colleague to test with their own login.
- Keep versions. Save a working version before each significant change so you can go back if an update breaks something. FromScratch saves every change, which makes rollback practical.
- Limit scope for the first release. Launch the core job well, then add features. A small app that works is more valuable than a large one with hidden problems.
Once the app has passed these checks with real users, you can publish it with confidence. For a first project, many non-coders start with an internal tool, which lowers the risk because only your team uses it. For more on that approach, see How to Build an Internal Tool with AI (Step by Step).
Build it with FromScratch
FromScratch is an AI app builder that works in a chat. You describe what you want, and an AI agent builds the app with a React front-end, a Python (FastAPI) back-end and a database. You can see a live preview within minutes, then refine it by continuing the conversation. It is a non-coder AI app builder, so you do not need to write or read code to use it. If you are new to the terminology, What Is Vibe Coding? A Plain-Language Guide explains the basics.
To build a simple team dashboard for tracking prospects, follow these steps:
- Create a new project. Sign up on the free plan and open a new project in the chat.
- Describe the app. Paste the prompt below. It names the fields, the search and filter features, the login requirement and the spreadsheet you want to use as a starting point.
- Attach your spreadsheet. Upload your lead list (a spreadsheet file). The app can use your real content rather than placeholder data.
- Review the live preview. Check that the table shows the right columns and that your records appear correctly. Look for the warning signs covered earlier.
- Refine by chatting. Ask for changes in plain language, such as "make the follow-up date column sortable" or "show overdue follow-ups in red." Test each change.
- Restore a version if needed. If a change breaks something, go back to an earlier saved version.
- Test with your team. Have two or three people use the app with real leads and report anything that seems off.
- Publish when it is ready. Publish to a public URL, or to your own domain if you prefer. For domain setup, see Custom Domain for an AI-Generated Website: Step-by-Step.
Here is the exact prompt to type into the chat:
Build a simple internal dashboard for my small team that tracks prospect leads in a table, with fields for name, company, status and next follow-up date. Add a search bar and a filter by status, and require login so only my team can see it. Use the spreadsheet I attach as the starting data.
If you want a more detailed walkthrough of this type of project, Build a Dashboard with AI covers the common patterns. If your needs grow into customer tracking, Build a CRM with AI is the next step.
FAQ
What is vibe coding?
Vibe coding means building software by describing what you want to an AI in plain language, then guiding the result through conversation rather than writing the code yourself. It is fast and accessible, which is why it appeals to non-coders. The main risk is that speed can hide problems like weak security or inconsistent logic, so testing matters.
Is an app built with AI safe to use with customer data?
It depends on how the app is built, tested and configured. Before you put real customer information into any app, check access controls, test logins from a logged-out state, and confirm who can see each page. For sensitive data, consider getting a technical review from a qualified professional before launch.
How long does it take to get a working prototype?
With an AI app builder, you can often see a live preview within minutes of describing the app. Getting a prototype that works well with your real data takes longer, because testing and refinement are where most of the value is. Plan for multiple rounds of changes before you share it widely.
Do I need to know how to code to maintain an AI-built app?
Not necessarily, but you do need to test changes and keep versions so you can roll back. Being able to describe problems clearly and spot unusual outputs is the most important skill. If the app grows in complexity, a developer may eventually be needed for specialized work.