Finally set a hard 3-prompt limit on AI debugging and it saved my sanity

I kept falling into the trap of letting Cursor/Claude try to "fix" the same issue 4 or 5 times in a row, only to realize an hour later that it had completely wrecked adjacent files and hallucinated imports. Now I have a strict rule: if it doesn't fix it in 3 attempts, I immediately git-checkout rollback and read the stack trace myself. Jotted down a quick checklist of rules I use to stop AI context degradation. Happy to share the doc if anyone wants it, but curious: what are your personal guardrails to stop agentic tools from looping?