Vibe coding debugging: five checks before another prompt
code-fix.dev ·
A practical sequence for debugging AI-generated apps: reproduce the failure, collect evidence, narrow changes, and verify the fix.
1. Separate expected and actual behavior
Replace “login is broken” with “email login should open the dashboard, but returns to the login screen.” Record the page, account type, action and timing so another person can reproduce it.
2. Reproduce under the same conditions
Compare behavior before and after refresh, signed in and signed out, and in another browser. Record whether it always fails or only sometimes. Write the exact sequence of clicks rather than a general description.
3. Separate interface and server failures
Inspect Console and Network in browser developer tools. If no request is sent, inspect the event handler first. If a request fails, examine its response and server logs. A status code alone is not a diagnosis. Remove cookies, tokens and personal information before sharing logs.
4. Find the last working change
Compare Git changes around the first failure. Test one suspected change at a time. For deployment-only failures, compare environment variable names, start commands and dependency versions. Do not test by deleting production data or exposing credentials.
5. Verify the original flow too
Check valid input, invalid input, refresh and signing in again. Give AI the reproduction steps, relevant code and redacted logs, and ask for the reason behind each change and how to verify it. If failures repeat or affect stored data, a repository review can establish a safer repair scope.