What is 'Constitutional AI' as a prompting pattern and how does it differ from a standard system prompt with refusal rules?

Constitutional AI started as a way Anthropic trained Claude using a set of written principles instead of only human labels. The same idea can be applied as a prompting pattern at inference time, without any training. The pattern works in three steps. First, the model generates a draft answer. Second, the model critiques the draft against a named list of principles, saying for each one whether the draft violated it. Third, the model revises the draft based on the critique. The reason this matters more than a long system prompt with rules is that critique becomes a discrete model step. You can inspect it. You can log it. You can grade it. With a plain system prompt, you have to hope the model followed every rule. With the critique step, you can see whether it noticed when it did not.

Constitutional AI started as a training technique from the 2022 Anthropic paper, where Claude learned from a written set of principles plus model-generated critiques rather than only human labels. The same idea works as an inference-time pattern, and that is what this question is really about. The pattern is a structured loop. Generate a draft. Critique the draft against each named principle in turn, with the model explicitly stating whether the draft violated it and how. Revise the draft to address the critique. The whole loop runs in one or two LLM calls per turn. The difference from a standard system prompt is structural. A system prompt with rules is a hope: the model may or may not follow each rule, and you cannot tell from the output which rules it considered. The critique step is evaluable. You can read it, grade it, and use it as a signal for routing or rejection. That is why constitutional prompting shows up in production guardrail stacks even when the underlying model was not trained with CAI.