Recursive Reflection: A Three-Stage AI Prompting Method to Improve Output Quality

A prompting technique called Recursive Reflection proposes that AI outputs can be significantly improved by running them through a structured Draft-Critique-Rewrite loop rather than accepting a single generated response. The approach is rooted in research by Madaan et al. (2023), which found that large language models perform measurably better at self-improvement when given explicit feedback criteria alongside their own prior output. The core insight is that AI models critique more effectively than they generate from scratch, switching from a 'fill the blank' mode to a 'find the gap' mode when evaluating existing content. By feeding critique tokens back as context for a rewrite, the output is constrained to a higher-quality probability space without requiring a more powerful model. The framework is positioned as a practical workflow fix for users who find themselves heavily editing AI-generated content that should have required minimal revision.