Anthropic's prompting guide for Claude Opus 5.5 says the model's effort setting should be the main control for how much it thinks. The company reports that, in one chat-product test, removing a system instruction telling the model to think carefully made replies begin sooner with no clear decline in answer quality. That is narrower than a blanket ban on phrases such as “take a deep breath” or “think step by step.” Anthropic also advises users migrating from thinking-disabled setups to remove prompts that ask the model to reproduce private reasoning and to test lower effort levels when latency matters.