A team builds a customer service chatbot using GPT-4. They write prompts inline in the application code as Python string literals. Three months later, a prompt change to improve response quality breaks customer satisfaction metrics. No record exists of what the original prompt was. What practice would have prevented this?