Habits and rules

Vibe coding best practices: the 10 mistakes that sink AI-built apps

Updated November 17, 20268 min read

Most vibe coding best practices come from the same handful of mistakes that people make again and again. Below are 10 common ones, what goes wrong when you make them, and the simple habit that fixes each. Read them as vibe coding tips in reverse: each mistake shows you the habit to build. At the end there is a one-page list of vibe coding rules you can paste into your project notes.

Mistake 1: asking for the whole app in one message

It is tempting to paste in a long wish list: logins, payments, a dashboard, emails, dark mode. The AI will try. You get a huge pile of changes at once, some parts work, some don't, and you have no idea which change broke what.

The fix: one feature per message. Build it, test it, save it, then ask for the next. If you want the AI to see the bigger picture, give it the full plan once and say "we'll build step 1 only for now".

You can tell you are making this mistake when the AI's reply lists more than a handful of changed files, or when you can't say in one sentence what you are about to test.

Tip

A good size for one request is something you can test in under five minutes.

Mistake 2: skipping the plan and letting the AI guess

When you start with "build me a booking site", the AI fills every gap with its own guesses: which pages, what data, which services. Some guesses are fine. Others lock you into choices that are painful to undo later, like a database you didn't want.

The fix: ask for a plan before any code, and read it. If something in the plan surprises you, ask about it now. Changing a sentence in a plan takes seconds. Changing a built app takes hours. Write down what is not in version one, too.

A useful plan answers five things: who the app is for, the one main thing they do, which screens exist, what data the app must remember, and what you are leaving out. It fits on one page. If the AI's plan is much longer than that, ask it to cut the plan down to the smallest version that still does the main thing.

Note

Coding agents like Claude Code have a plan mode made for exactly this step.

Mistake 3: building for hours with no save points

Everything works at 3pm. At 5pm the AI makes a change that breaks the login and, while fixing it, breaks two other things. With no saved version from 3pm, you are stuck trying to repair forward.

The fix: save a checkpoint every time something works, and always before anything risky like a redesign or a new database. Coding agents use git for this (a tool that keeps a history of every saved version). Just ask the AI to "save a checkpoint with a short note". Going back is almost always faster than fixing.

Tip

If you only take one habit from this page, take this one.

Mistake 4: keeping one chat going for days

AI tools can only hold so much in their working memory, called the context. A chat that has run all week is full of old ideas, old errors and decisions you later changed. The AI starts mixing them up, forgets rules you gave it on Monday, or brings back a bug you already fixed.

The fix: start a fresh chat for each new task. Keep a short notes file the AI reads at the start of every chat (CLAUDE.md for Claude Code, AGENTS.md for Codex) with what the app is, how to run it and your rules. When you start a new chat, add one short paragraph on where you are.

A handy trick: at the end of a long session, ask the AI to "write a short summary of what we did today and what is left to do". Paste that summary into the next chat, or add it to the notes file.

Tip

VibeCodingDeck shows a context meter on each agent's box, so you can see a chat filling up and clear it before the AI starts forgetting.

Mistake 5: telling the AI "it's broken"

The AI can't see your screen. "It's broken" gives it nothing, so it guesses, often changing things that were fine.

The fix: report bugs in three parts. What you did, what you expected, what happened. Paste the whole error message if there is one, even if it looks like nonsense. Then ask the AI to find the cause and explain it before it changes anything. That one extra line stops a lot of random edits.

Screenshots help too, if your tool accepts them. A picture of a squashed layout on a phone says more than a paragraph about it.

Tip

Ready-made bug report wording is in our vibe coding prompts library.

Mistake 6: letting the AI fix its own fix, again and again

This is the fix loop. The AI tries a fix, it fails, it tries another on top, that fails, and after ten rounds the code is a tangle of half-fixes. Each try feels close, so you keep going.

The fix: set yourself a limit, like three tries. After that, go back to the last checkpoint, start a fresh chat, and ask for a different approach. Another option is a second opinion: hand the problem to a different AI. Some people build with Claude and ask Codex (OpenAI's coding agent) to review, or the other way round. Our Claude Code review guide covers how to ask for a good review.

Watch out

If three fresh attempts all fail, that is the moment to ask an experienced developer for an hour of help.

Mistake 7: testing with one account and only the easy path

You test as yourself, you type sensible things, and you click in the right order. Everything works. Then a real user types an emoji into the price field, or logs in on a second account and sees your data.

The fix: test like an annoying user. Use a phone-sized screen. Leave forms empty, paste in huge text, try a negative number. Make two accounts and check that one can't see or change the other's things, including by changing numbers in the address bar. Then ask the AI to write automatic tests (small checks the computer runs for you after every change) for the important parts (sign-up, saving data, payments), so a future change can't quietly break them.

Tip

Ask a friend to use the app with no help while you watch. You'll learn more in ten minutes than in an hour of your own testing.

Mistake 8: pasting secret keys into the chat or the code

API keys are passwords that let your app use paid services like AI, email or payments. Paste one into a prompt and it sits in the chat history. Let the AI put one into code that runs in the browser and any visitor can copy it and spend your money.

The fix: keep keys in a private settings file, usually called .env, that is never uploaded. Tell the AI the name of the key, not the key itself: "the payments key is in .env as PAYMENTS_KEY". Before launch, ask the AI to search the project for any keys outside that file. If a key ever leaks, make a new one at the provider and delete the old one.

Watch out

More on this and the other big risks is in the vibe coding security guide.

Mistake 9: giving the AI the keys to your live app

Once your app has real users, its real database matters. In July 2025, Jason Lemkin was building an app with Replit's AI agent in public. The agent deleted his project's live database, even though he had asked it not to change anything. Replit's chief executive called it unacceptable and added new safeguards. The same thing could happen with any AI tool that has the power to do it.

The fix: build and test against a copy of your data, never the live one. Keep backups. Don't hand the AI passwords that can delete real customer data. If you need to change the live app, test the change on the copy first, then apply it yourself or with a careful, reviewed step.

Note

An AI agent will follow instructions most of the time. Plan for the times it doesn't.

Mistake 10: forgetting about the bills

Vibe coding costs money in two places. Your AI plan while you build, and the services your app uses once it is live: hosting, database, email and any AI features. A bot hammering your AI chatbot or signup form can run up a large bill overnight.

The fix: set a monthly spending limit on every paid service before you launch. Limit how often each user can use costly features, which is called rate limiting. While building, keep tasks small and chats fresh, since most AI plans have daily or weekly usage limits and big messy requests use them up faster.

Check your bills in the first week after launch, even if they look small. A slow rise you spot early is easy to fix. A surprise at the end of the month is not.

Tip

Ask the AI: "List every paid service this app uses and how I set a spending cap on each."

Vibe coding rules on one page

These are the habits from above, written as rules. Paste them into your project notes file so the AI follows them too. This is the short answer to how to vibe code well.

  • Ask for a plan before any code, and read it.
  • Write down what is not in version one.
  • One feature per message. Test it before the next.
  • Save a checkpoint every time something works.
  • Start a fresh chat for each new task.
  • Keep a notes file the AI reads first.
  • Report bugs as: what I did, what I expected, what happened.
  • Three failed fixes means go back and try another way.
  • Test on a phone and with two accounts.
  • Keep secret keys in .env, never in prompts or browser code.
  • Never let the AI work on live customer data.
  • Set spending limits on every paid service.
  • Get a review (another AI or a person) before taking payments or storing sensitive data.

If you are just starting out, our how to vibe code tutorial puts these rules into practice on a real weekend project.

Sources

Run up to 12 AI coding agents in one window.

Claude Code, Codex, Gemini, Grok and more, side by side. Windows and Mac. From $29 a month.

Get started

Keep reading