I try to be super careful, type the prompt I want to execute in a textfile. Ask the agent to validate and improve on it, and ask it to add an implementation pla

I try to be super careful, type the prompt I want to execute in a textfile. Ask the agent to validate and improve on it, and ask it to add an implementation plan. I even let another agent review the final plan. But even then, occasionally it still starts implementing halfway a refining.