Skip to main content
Vibe Code Forum

Question: The AI fixed one bug and broke two other things. What do I do?

Almost everyone hits this. You ask the AI to fix something, it says it's fixed, and now two other features don't work. You ask it to fix those, and the first bug comes back. How do you get out of the loop?

Solution

By Vibe Code Forum team in post #2

Stop asking for fixes on top of fixes. Each round adds more changes to untangle. Instead:

  1. Go back to the last version that worked (version history in app builders, or Git: see Git and GitHub explained). Losing an hour of broken changes is cheaper than another three hours of loop.
  2. Start a new conversation. The old one is full of failed attempts the AI keeps referring to. Begin with a short summary of the project and the one problem.
  3. Ask for a diagnosis, not a fix: "Before changing anything, explain what causes this bug and which files are involved."
  4. Ask for the smallest possible change: "Fix only this. Don't refactor or change anything else."
  5. Test the old features too after the fix, then save a version straight away.

To avoid it next time:

  • Save a version after every working change.
  • Ask for one feature at a time. See Prompting that works.
  • Once the project has real users, ask the AI to add automated tests for the important features, so it can tell when it breaks them. See Test before you trust.

The full routine is in When it breaks: a calm debugging routine.

Read it in context

Stop asking for fixes on top of fixes. Each round adds more changes to untangle. Instead:

  1. Go back to the last version that worked (version history in app builders, or Git: see Git and GitHub explained). Losing an hour of broken changes is cheaper than another three hours of loop.
  2. Start a new conversation. The old one is full of failed attempts the AI keeps referring to. Begin with a short summary of the project and the one problem.
  3. Ask for a diagnosis, not a fix: "Before changing anything, explain what causes this bug and which files are involved."
  4. Ask for the smallest possible change: "Fix only this. Don't refactor or change anything else."
  5. Test the old features too after the fix, then save a version straight away.

To avoid it next time:

  • Save a version after every working change.
  • Ask for one feature at a time. See Prompting that works.
  • Once the project has real users, ask the AI to add automated tests for the important features, so it can tell when it breaks them. See Test before you trust.

The full routine is in When it breaks: a calm debugging routine.

Nobody has confirmed this yet. If you try it, say how it went.

More in Ask for help