“Anthropic has advised that modern language models require fewer instructions, not more, for optimal performance.”