Constitutional prompting refines LLM outputs with self-critique

A new prompting technique called "constitutional prompting" has been developed to improve the safety and adherence of LLM outputs. This method involves a three-pass process: first, the model generates a draft response; second, it critiques its own draft against a predefined set of rules or principles; and third, it revises the draft based on the critique. This approach ensures that feedback is specific and auditable, focusing on named violations rather than general quality improvements, and can be iterated for more robust results. AI

IMPACT Enhances LLM reliability by enabling models to self-correct against defined principles.

RANK_REASON Describes a new method for improving LLM output quality and safety.

Read on dev.to — LLM tag →

• Constitutional prompting
• JSON

• product
• other

AI-generated summary · Google Gemini · from 1 sources. How we write summaries →