Prompting current Claude models

Your firm's searches-report summariser was written two years ago. It reads the local authority, drainage and water, and environmental search results for a purchase, and drafts the report a fee earner checks before it goes to the buyer. The prompt has grown to 2,400 tokens. Half its lines are in capitals, it tells the model to "think step by step", it forces JSON with an assistant prefill, and it sets temperature=0. On the new model it now returns a 400 error, and once that's fixed, the reports read as if the model is nervous. This lesson is about writing prompts for the models you have now, and about knowing which lines in an old prompt are doing harm.