“Anthropic's prompting strategy for newer models favors providing more context and fewer negative constraints to avoid conflicts with user instructions.”