A team shares prompt engineering tips for improving token efficiency in an LLM agent harness, based on lessons learned at Cursor.