How to Fix AI-Generated Code When It Breaks (3 Moves)

Your AI just broke something. The screen is white, or nothing loads, or you're staring at red text that means nothing to you. This is not a sign you're doing it wrong — it happens to everyone. Here's what to do right now.

Why AI-Generated Code Breaks (The Short Version)

AI coding tools don't have a full picture of your project. They see what you've shown them in the current conversation, and nothing else. That gap is where most breaks happen.

It doesn't see your whole project

When you ask an AI to change the login form, it edits the login form. It doesn't automatically know that the login form shares a function with the signup page, or that you have a custom auth hook that depends on a specific prop name. It writes code that's correct in isolation but wrong in context.

It doesn't know what changed two prompts ago

The AI treats each response as a fresh answer. If you made three changes earlier in the chat and something only breaks now, the AI has no reliable sense of which change caused it. It can reason about what you've shared, but it can't rewind. That's your job — and git makes it trivial once you're in the habit.

Move 1: Paste the Error Back In (The Right Way)

This is almost always your first move. It sounds obvious but most people skip it or do it wrong.

Copy the exact error message — don't paraphrase it

Don't type "it says something about undefined." Copy the full error text — every line of it — and paste it directly into the chat. Error messages are precise for a reason. "Cannot read properties of undefined (reading 'map')" tells the AI exactly where to look. "It says undefined" tells it almost nothing.

If you're in a browser, open the developer console (F12 → Console tab) and copy from there. If you're in a terminal, copy the full output starting from where it went red.

Add one sentence of context

After the error paste, add one sentence: what you just did. Something like: "This happened after you changed the login form." or "This started after adding the new API call in the last message." That sentence cuts the AI's search space in half.

Keep it to one sentence. Don't explain your whole app — you'll muddy the signal.

What to do if there's no error message — just a broken UI

Sometimes nothing crashes, it just looks wrong. A button doesn't appear. A form doesn't submit. The layout is broken. In this case, describe exactly what you see versus what you expected:

• "The submit button is gone after your last change. It was there before."
• "The user's name no longer shows in the top right corner."
• "The page is blank after I click login."

Specific beats vague every time. "It's broken" is the least useful thing you can say. "The submit button disappears when the form has an error message" is something the AI can actually act on.

Move 2: Roll Back with Git Before It Gets Worse

If Move 1 fails — the AI tries again and makes it worse — stop. Do not keep prompting. Every new edit adds more to undo later. Roll back now.

See the full git survival guide for vibe coders if you haven't set up git yet or need a refresher on committing. What follows assumes you have at least one commit to go back to.

Check what changed: git diff or just ask your AI tool

Before rolling back, see what the AI actually touched. In your terminal:

git diff

This shows every line that changed since your last commit. Green lines were added, red lines were removed. You don't need to understand the code — you're just looking for how much changed. A small diff means a small rollback. A huge diff means the AI went deep and you should definitely roll back.

Alternatively, in Cursor you can click the file in the sidebar and it will highlight changed lines inline. Either way, look before you roll back — it confirms you're undoing the right thing.

Undo the last commit and start the broken section over

To undo your last commit but keep the files in their current state:

git reset HEAD~1

To wipe the changes entirely and go back to exactly how things were before:

git checkout .

If you committed the AI's broken changes, use git reset HEAD~1 first, then git checkout . to clear anything that's still sitting there uncommitted.

Now you're back to your last known-good state. Re-prompt the AI from there with a cleaner, narrower instruction.

Why this is faster than trying to unpick bad AI edits

When the AI makes a mistake and you ask it to fix the mistake, it often compounds it. It doesn't know what "working" looked like — it only knows what it wrote. Rolling back to a working commit and starting the section fresh almost always gets you to a fix in fewer total prompts than trying to patch an already-broken attempt.

The time it takes to learn two git commands pays back immediately the first time you use them.

Move 3: Start a Fresh Chat

If you've pasted the error, if you've tried re-prompting, and the AI is going in circles — the chat itself is the problem. Long conversations make AI tools less reliable. Context builds up, earlier instructions start conflicting with recent ones, and the model can get stuck optimizing around its own previous responses rather than solving your actual problem.

Signs the current chat is the problem (not the code)

• The AI keeps apologizing and trying variations of the same broken fix
• It's now changing things that were working before you started this conversation
• It's contradicting advice it gave you five messages ago
• The responses are getting longer and more hedged with every attempt

Any of these is a signal: start fresh.

How to carry your context into a new session without losing progress

Don't just close the tab. You'll lose the context of what you were trying to build, and you'll waste the first few messages of the new chat re-explaining your project.

Instead, before you open a new chat, copy these three things from your project:

1. A short description of what the feature is supposed to do
2. The relevant file(s) — paste the current code, not a description of it
3. The exact error or broken behavior you're trying to fix

Then open the new chat and lead with those three things. You'll get a cleaner response than if you'd continued the old thread.

The "summary handoff" prompt — ask the AI to summarize before you nuke the thread

If you're in the middle of a complex session and don't want to lose track of what's been discussed, ask the AI to summarize before you close it:

"Before I start a new chat, give me a one-paragraph summary of what we were building, where we got stuck, and what approaches you already tried."

In Claude Code, you can also use the /clear command to reset the conversation context without starting a new terminal session — useful if you want to keep the same working directory but wipe the chat history.

In Cursor, opening a new Composer window starts with a fresh context. The previous Composer tab stays intact if you need to refer back to it.

Paste the AI's summary into the new chat as your opening context. You'll be back up to speed in one message.

Preventing the Next Break

Fixing the current break is urgent. Fixing the pattern is what saves you from doing this every day.

Commit before every significant AI change

Before you ask the AI to do anything substantial — refactor a component, add a new feature, change how auth works — run:

git add .
git commit -m "before AI: [what you're about to ask]"

That's your safety net. If the next AI response breaks something, you have a clean rollback point. This habit alone will cut your debugging time in half.

One task per chat session

Long, multi-feature sessions are where AI tools go wrong. The more context accumulates, the more the AI tries to hold together, and the more likely it is to make an edit in one place that breaks something it touched three messages ago.

Keep each chat to one task. Build the form. Add the API call. Fix the styling. When that task is done and working, commit it and start a new chat for the next thing.

How better prompts reduce break frequency

Most AI-generated code breaks because the prompt was underspecified. "Add a login form" leaves too much to interpretation. "Add a login form with email and password fields that submits to /api/auth/login and shows an inline error message if the response is not 200" gives the AI almost everything it needs.

The guide to writing better prompts for AI coding tools goes deeper on this — it's worth reading once you've got the current fire out. Better prompts mean fewer breaks, which means less time debugging and more time actually building.