WorkflowAI tools

The AI keeps 'fixing' the same bug and making it worse

Same error after 5 AI attempts; the code keeps getting longer

In short: Each blind retry adds more code and more guesses. Stop patching and get a diagnosis first.

Why it happens

AI agents tend to change code when they're unsure, rather than gather evidence. After a few rounds, the original error is buried under workarounds. The fix is to reset to a known-good state and force a diagnose-before-edit step.

How to fix it

1. Revert to the last working commit or checkpoint (git, or the tool's version history).
2. Copy the exact error text, the file, and what you did right before it broke.
3. Ask for a diagnosis only, with no code changes: likely causes, ranked, and how to confirm each.
4. Confirm the cause with a log or a small test, then ask for the smallest fix.
5. Commit after every working step so you always have a checkpoint.

Copy-paste fix prompt

I'm getting this error in my app:

Same error after 5 AI attempts; the code keeps getting longer

[Paste the full error, stack trace, and the file it points to here]

Likely cause: Each blind retry adds more code and more guesses. Stop patching and get a diagnosis first.

Task: Do not edit any files yet. Read the error and the relevant files, list the 3 most likely root causes ranked by probability, and for each tell me one quick check (a log line or command) that confirms or rules it out.

Rules:
- First explain the root cause in 2 sentences and list the files you'll change.
- Make the smallest change that fixes it. Don't refactor unrelated code.
- Don't disable security or checks (RLS, auth, signature verification, type checking, CORS allow-all) to make the error disappear.
- Never print or hard-code secret keys.
- Finish by telling me exactly how to verify the fix.

Common with: Cursor, Claude Code, Codex, Lovable, Bolt, Replit, v0. Last reviewed 2026-11-13.