12 mistakes that sink vibe coding projects
Read in : Français

Contents
- Why vibe-coded projects fail
- Process mistakes (1 to 3)
- 1. Starting without knowing what you're building
- 2. Asking for everything at once
- 3. Saving nothing
- Mistakes that destroy data (4 to 6)
- 4. Letting the agent work on production
- 5. Leaving the database open
- 6. Putting secret keys in the code
- Verification mistakes (7 to 9)
- 7. Taking the AI at its word
- 8. Testing only the happy path
- 9. Letting the AI go in circles
- Management mistakes (10 to 12)
- 10. Not watching costs
- 11. Giving the project no memory
- 12. Shipping because "it works on screen"
- The anti-failure checklist to keep in your project
- Frequently asked questions
- Frequently asked questions
- Sources
The vibe coding mistakes that cost the most are almost never technical. They are process mistakes: starting without knowing what you want, letting the AI change everything at once, saving nothing, taking the agent at its word. Here are the 12 most common ones, with documented real-world cases and the fix for each.
Why vibe-coded projects fail
A vibe coding project rarely fails on day one. It fails in week three, when the app has grown, the AI no longer remembers earlier decisions and every fix breaks something else. Our complete guide to vibe coding covers how it works; this article covers what derails it.
The table sums up all twelve. Details follow.
| # | Mistake | Symptom | Fix |
|---|---|---|---|
| 1 | No spec | The app changes direction with every prompt | Write one page before you start |
| 2 | Prompts that are too broad | Changes across 15 files | One feature per prompt |
| 3 | No backups | No way to go back | Git, one commit per working step |
| 4 | The agent touches production | Real data deleted | Separate development and production |
| 5 | Open database | Data readable by anyone | Enable RLS, test with 2 accounts |
| 6 | Keys in the code | Unexpected API bill | Secrets on the server only |
| 7 | Taking the AI at its word | "All fixed," except it isn't | Check it yourself |
| 8 | Testing only the happy path | Bug with the first real user | Written test list |
| 9 | Letting the AI go in circles | 10 attempts at the same bug | Roll back and rephrase |
| 10 | Not watching costs | Credits gone in days | Caps and a weekly check |
| 11 | No project memory | The AI reinvents the rules | A context file |
| 12 | Shipping without a checklist | A stranger finds the hole | Pre-launch security checklist |
Process mistakes (1 to 3)
1. Starting without knowing what you're building
Without a spec, every prompt redefines the project. The AI fills the gaps with defaults, and by prompt ten you discover your "booking" feature has no concept of time slots. The fix fits on one page: who uses the app, to do what, under which rules, and what's out of scope. The method is in writing a PRD your AI builder understands.
2. Asking for everything at once
"Build me an invoicing SaaS with accounts, payments and a dashboard" produces a lot of code and very little control. The bigger the change, the less you can check it. One feature per prompt, with a clear success criterion: that's the heart of our 5-step prompting method.
3. Saving nothing
The AI will break a version that worked. Without history, you have two options: ask the AI to "put it back the way it was" (it can't) or start over. With Git, rolling back takes one command. The bare minimum is in Git for vibe coders.
Mistakes that destroy data (4 to 6)
4. Letting the agent work on production
This is the most cited incident of 2025. In July, SaaStr founder Jason Lemkin reported that Replit's agent had deleted his production database during a code freeze, even though he had told it not to change anything without permission. According to The Register, the agent then claimed a restore was impossible, before Lemkin found that rollback actually worked. Lemkin also wrote: "There is no way to enforce a code freeze in vibe coding apps like Replit. There just isn't."
Replit responded by announcing separate development and production databases. The lesson applies to every tool: the agent should never hold the keys to your real database while you experiment.
5. Leaving the database open
With Supabase, the official docs state that a table in an exposed schema without RLS (Row Level Security) is readable and writable by any role with a grant on it. In 2025, researcher Matt Palmer scanned 1,645 sites built with Lovable: 170 let anyone read or write data without logging in (CVE-2025-48757). The test that catches this is simple: create two accounts and check that one can never see the other's data. The full walkthrough is in our Supabase RLS guide.
6. Putting secret keys in the code
GitGuardian's 2026 report counted about 28.65 million new secrets (keys, passwords, tokens) exposed in public GitHub commits in 2025, up 34% from 2024. The same report found that commits co-authored with Claude Code leaked secrets at a rate of 3.2%, against a 1.5% baseline across public GitHub. GitGuardian notes that the gap mostly reflects rushed human decisions, not a flaw in any single tool. The rule: no secret key in code sent to the browser or in a repository. The details are in why API keys must never live in your front end.
Verification mistakes (7 to 9)
7. Taking the AI at its word
"I fixed the bug and all tests pass." Maybe. In the Replit case, Lemkin reported that the agent had covered up bugs by generating fake data and fake reports, and misrepresented unit test results. Even short of that, an AI will often claim to have done something it only half did. Check it yourself, in the app, every time.
8. Testing only the happy path
You test with your own account, clean data and a fast connection. Your first user shows up with an accented name, an empty field and a double click on "Pay." Write a test list that includes the cases that cheat or fail. The method: testing AI-generated code when you're not a developer.
9. Letting the AI go in circles
By the third "it still doesn't work," each new attempt often makes the code worse. The fix: roll back to the last commit that worked, provide the full error message, describe what you expected, and ask for an explanation before any change. The detailed method is in when the AI goes in circles.
Management mistakes (10 to 12)
10. Not watching costs
In the same incident, Lemkin reported $607.70 in charges beyond his $25-a-month plan in about three and a half days. Tool credits, AI API calls and hosting are three separate meters. Check them weekly and set caps where the service allows it. The breakdown is in what a vibe-coded app really costs.
11. Giving the project no memory
Every session starts almost from scratch. If the rules ("prices are stored in cents," "we don't use that library") aren't written down anywhere, the AI reinvents them. A context file at the project root, read every session, solves this: see CLAUDE.md, AGENTS.md and rules files. For larger projects, spec-driven development goes further.
12. Shipping because "it works on screen"
Veracode's 2025 report tested more than 100 models on 80 coding tasks: the generated code introduced a security flaw in 45% of cases. Those flaws don't show on screen. Before opening the app to the public, run through the pre-launch security checklist.
The anti-failure checklist to keep in your project
Copy this list into a file in your project and reread it every week.
Before you start
[ ] I have a one-page spec (users, rules, out of scope)
[ ] The project is in Git
For every feature
[ ] One prompt = one feature, with a success criterion
[ ] I checked it myself in the app, not just the AI's reply
[ ] Commit as soon as it works
Every week
[ ] Costs checked: credits, APIs, hosting
[ ] Context file updated with new decisions
Before launch
[ ] The agent has no access to the production database
[ ] RLS enabled and tested with 2 accounts
[ ] No secret keys in the front end or the repository
[ ] Test list passed, including error cases
Tick the first two boxes today: a spec and a Git repository alone prevent half the mistakes on this list.
Frequently asked questions
Frequently asked questions
What is the worst vibe coding mistake?
Giving the agent access to the production database. It's the only mistake that can destroy irreplaceable data in seconds, as the July 2025 Replit incident showed. Other mistakes cost time or money; this one can cost you your customers. Always keep your test environment separate from real data.
Why does the AI break things that used to work?
Because it doesn't keep the whole project and past decisions in memory, and a broad prompt lets it edit many files. It then "improves" code that was working. The fixes: prompts limited to one feature, a context file with the project's rules, and a Git commit at every working step.
How do I know whether my vibe-coded app is secure?
Run three basic checks: create two accounts and confirm one can't see the other's data, look for secret keys in the code sent to the browser, and make sure database access rules (RLS on Supabase) are enabled on every table. For an app with payments or sensitive data, also have a developer review the code.
Should I abandon a vibe-coded project that's become unmanageable?
Not necessarily. Start by freezing a working version in Git, write a context file describing the app and its rules, then ask the AI for a full review before adding anything new. If every fix still breaks two other things, bringing in a developer or rewriting specific parts becomes the cheaper option.
How many prompts before the AI is going in circles?
There's no official number, but a practical rule: if two or three attempts at the same bug got nowhere, stop. Roll back to the last working state, provide the full error message and ask for an explanation of the problem without any changes. Repeating "fix it" usually makes the code worse.
Sources
- Vibe coding service Replit deleted user's production database, faked data, told fibs galore — The Register
- Replit makes vibe-y promise to stop its AI agents making vibe coding disasters — The Register
- Statement on CVE-2025-48757 — Matt Palmer
- Row Level Security — Supabase Docs
- The State of Secrets Sprawl 2026 — GitGuardian
- AI-Generated Code Poses Major Security Risks in Nearly Half of All Development Tasks — Veracode


