You built something. It was going beautifully. Claude was making things, you were directing, and for a while the gap between “I want this” and “here it is” felt impossibly small.

Then something broke. And you have no idea why.

Maybe the page stopped loading. Maybe the thing that worked yesterday does not work today. Maybe you asked for one change and three other things fell apart. You are staring at error messages that mean nothing to you, and the AI keeps suggesting fixes that make it worse.

This is normal. You are not failing. You have hit the wall that every vibe coder hits. Here is what to do about it.

The 80% wall

There is a pattern that plays out across almost every vibe coding project. The first 80% goes brilliantly. The layout appears. The pages look right. The interactions feel smooth. You are amazed at how fast things come together.

Then you reach the last 20%. This is where projects need user accounts, payment processing, connections to other services, handling for all the ways things can go wrong, and the hardening required before real people use it. This is the part that has historically required years of experience to get right.

The AI promised you would not need that experience. It was half right. It can build the 80% without you knowing anything. But the last 20% is where the AI needs you to be a better collaborator — more specific in your descriptions, more patient with the process, and more willing to break problems into smaller pieces.

The 80% wall is not a sign that vibe coding does not work. It is a sign that you have graduated from the easy part. The skills that get you past it are not programming skills. They are communication skills, patience, and knowing when to ask for help.

The debugging mental model for non-developers

Here is the most important thing to understand: you do not need to read code to fix problems.

What you need is a way of thinking about problems that helps the AI help you. Developers call this “debugging.” You can think of it as giving the AI better directions when it is lost.

There are four moves, and they work in almost every situation.

Move 1: Describe what you expected versus what happened.

This is the single most useful thing you can do. Not “it is broken” but “I expected the sign-up button to take me to a confirmation page, but instead the page goes white and nothing happens.” The gap between what you expected and what actually happened is where the problem lives. The more precisely you describe that gap, the faster the AI finds it.

Move 2: Ask for the explanation before the fix.

When something breaks, the instinct is to say “fix it.” Resist that. Instead, say “explain what went wrong in plain language before you try to fix anything.” This does two things. It forces the AI to actually understand the problem rather than guessing. And it gives you a chance to spot if the AI is confused — if the explanation does not match what you saw, it is solving the wrong problem.

Move 3: Ask what changed.

Say “show me the last thing you changed.” Most problems are caused by the most recent change. If you were editing a form and the navigation broke, those two things are probably connected. Knowing what changed narrows the search enormously.

Move 4: Undo.

You can always go back. Say “undo the last change” and the AI will revert your project to before the problem started. This is possible because your project uses something called version control — a system that saves a snapshot every time you make a change. Going back to a snapshot that worked is always an option. It is not giving up. It is the smartest move in many situations.

The five recovery moves

When something breaks, try these in order. Each one escalates a little further. Most problems resolve in the first two steps.

1. Undo the last change

Say: “Undo the last change and get me back to where things were working.”

The AI will use version control to revert your project to the previous state. This is the equivalent of pressing the undo button in a document editor — except it works across your entire project. If things were working five minutes ago, you can be back there in seconds.

This is always safe. You lose only the most recent change. Everything else stays exactly as it was. Start here every time.

2. Start a new session with a clearer description

If undoing the change does not help — or if the problem has been building across several changes — start fresh. Open a new session and describe the problem from scratch.

This matters more than you might think. AI sessions accumulate context — everything you and the AI have discussed. After a long back-and-forth of trying fixes, the AI’s understanding of your project can become muddled. A fresh session with a clear, specific description of the problem often succeeds where the previous session was going in circles.

Write something like: “I have a portfolio website. The gallery page was working this morning. Now when I click on an image, nothing happens. The rest of the site works fine. Can you look at the gallery page and figure out what is wrong?”

Specific. Clear. No history to confuse things.

3. Ask for the explanation before the fix

Say: “Before you change anything, explain what you think is causing this problem. Use plain language.”

If you skipped this step earlier, do it now. The AI’s explanation will tell you whether it actually understands the problem. If the explanation matches what you are seeing, let it try a fix. If the explanation sounds wrong — if it is talking about a page you were not on, or a feature you never asked for — correct it before it starts changing things.

This is your quality check. It costs nothing and prevents the AI from confidently making the wrong thing worse.

4. Break the problem into smaller pieces

If the AI is struggling with a complex fix, make it smaller. Instead of “fix the checkout flow,” try:

  • “First, just make the button clickable again. Do not worry about what happens after.”
  • “Now make the button open the payment form.”
  • “Now make the payment form send the information.”

Small, concrete steps. Each one can be tested and verified before moving to the next. This is how professional developers work, too — they just do not talk about it in those terms.

The smaller the step, the less that can go wrong, and the easier it is to undo if it does.

5. Know when to stop and ask a human

This is not a failure. This is good judgment.

If you have tried the steps above and the problem persists, you need a person who can read the code. This might be a developer friend, a freelancer on a platform like Upwork, or a community like the Claude Discord or Stack Overflow. What you bring to them is valuable: a clear description of the problem, what you have tried, and a project that is mostly working with one specific thing broken. That is a much easier problem for a developer to solve than building from scratch.

Knowing when to hand off is a skill, not a weakness.

Red flags that mean stop

Not every problem is worth another attempt. Here are the signs that you should stop trying and escalate — either to a fresh approach or to a human.

The same fix keeps breaking something else. You fix the form and the navigation breaks. You fix the navigation and the images disappear. This is a sign the project has accumulated what developers call technical debt — layers of quick fixes that conflict with each other. The AI cannot untangle it because each fix creates a new problem. You need either a fresh start on that section or a human to look at the structure.

The AI starts contradicting itself. In one message it says the problem is in the layout. In the next it says the layout is fine and the problem is in the data. This usually means the AI’s context has become too cluttered to reason clearly. Start a new session. If it contradicts itself again with a fresh start, the problem is genuinely complex and you need human help.

The project was working and now nothing works. If a single change cascaded into a total failure, undo immediately. Do not try to fix forward. Get back to the last state that worked and approach the change differently — or ask a developer why that change broke everything.

You are on your third session trying to fix the same thing. Two honest attempts is a reasonable investment. Three means the problem is beyond what you and the AI can solve together in this way. That is fine. It does not mean vibe coding failed. It means this particular problem needs a different kind of expertise.

The security blind spot

This section is short and important.

If you have built something that handles passwords, payments, or personal data — and it is going to be used by real people — you need a human developer to review the security. This is not optional.

Recent research paints a stark picture. A 2026 study testing over 500 code samples from major AI models found that roughly one in four contained a confirmed security flaw. Separate studies found AI-generated code introduces vulnerabilities at significantly higher rates than human-written code — particularly around password handling, data exposure, and common web attacks.

The AI builds things that work. “Works” and “secure” are not the same thing.

You would not wire your own house just because you successfully installed a light switch. The same principle applies here. If real people will enter passwords, credit card numbers, or personal information into something you vibe-coded, pay a developer to review it. An hour of their time can prevent a serious breach.

This is the one area where “it seems fine” is not good enough.

You are still a vibe coder

Hitting the wall does not mean the approach is broken. It means you have built something complex enough to have real problems — and that is a sign of progress, not failure.

The developers who build software for a living hit the same walls. They just have more practice recovering from them. Every strategy in this guide — undoing changes, starting fresh, breaking problems into smaller pieces, knowing when to ask for help — is something working developers do every day. You are not cheating by using these techniques. You are doing exactly what they do, just with an AI doing the typing.

Keep building. Keep describing. Keep refining your ability to say precisely what you mean. The 80% wall gets further away the better you get at communicating with the AI — and communication is a skill you already have.

Art & Algorithms publishes guides, tutorials, and prompt packs at the intersection of art and code. Subscribe for the full archive.