Flashcard: what is the self-refine prompting pattern and what loop does it set up?

**Self-refine** is a three-step loop where the same language model plays three different roles in sequence: author, critic, and reviser. The author step is a normal prompt: produce a draft of whatever you are asking for, like a summary, a piece of code, an email, an argument. The critic step takes that draft and asks the same model, 'review this draft. What is wrong with it? What could be sharper, clearer, more correct?' The reviser step takes the original draft and the critique and asks, 'rewrite the draft to address each point in the critique.' One round of author plus critic plus reviser is one 'refine' step. You can loop the critic plus reviser pair for 1 to 3 rounds before deciding the output is good enough. After each loop, the output is usually a bit better than the previous one. The core insight is that critiquing one's own work is often easier than producing perfect work in one pass. Splitting the labor into separate prompts gives the model a chance to find and fix issues it would have missed in a single attempt.

**Self-refine** is a prompting pattern that decomposes a generation task into three sequential roles, all played by the same LLM with no fine-tuning: author (produce a draft), critic (review the draft and list specific issues), reviser (rewrite to address each issue). One round equals one author + critic + reviser cycle; in practice the loop runs 1-3 rounds before terminating. The motivation is that finding flaws in a piece of work is often easier than producing flawless work directly. When the model is asked to critique an existing draft, it can compare against criteria, notice missing pieces, and flag inconsistencies it would have missed during generation. The reviser then has a structured set of issues to address rather than starting from a blank page. The technique was formalized in Madaan et al. 2023 (Self-Refine: Iterative Refinement with Self-Feedback) and has become a standard component in agent loops. Production examples: AutoGPT-style agents critiquing their own plans, code-generation tools like Cursor doing draft, critique, fix loops, content tools running tone or style checks across rounds. Where it works best: writing and editing, code with stylistic or correctness issues, structured arguments, math solutions with verifiable steps. Where it underdelivers: factual recall (the critic shares the author's blind spots), tasks where the first draft is already correct, time-sensitive tasks where the 3x token and latency cost are unacceptable.