Anthropic's Claude uses a Constitution; adding reasons to prompts improves compliance

Constitutional AI replaces the human‑by‑human grading of RLHF with two AI‑driven stages. In the first stage the model critiques its own answers against the Constitution and rewrites them; in the second stage a separate model compares two answers and learns to rank them, a process Anthropic calls RLAIF. According to Anthropic’s prompt‑design guide, adding a one‑sentence reason to a constraint (e.g., “Please avoid abbreviations because the text will be read aloud”) makes Claude’s responses more consistent and less likely to refuse, because the model has been trained to consider the reasoning behind rules.