Self-Refinement & Critics

Expert

Refine model answers through multi-turn self-criticism loops.

01 The Concept

Self-Refinement & Critics prompts the model to generate a response, review its output against a strict safety or quality rubric, identify mistakes, and rewrite the final completion to eliminate defects before outputting.

By forcing the model to act as its own editor in a critique-and-rewrite loop, you systematically eliminate logical bugs, factual inaccuracies, and safety leaks.

02 Weak vs. Strong

EX 01Self-Refinement Prompt Setup

You are a professional editor. 1. Generate an initial draft email based on the notes. 2. Audit your draft against these rules: No internal project names, and maintain a polite tone. 3. List any violations found. 4. Output the final refined email draft with all violations removed.

→ Why it works

Structuring the task as a sequence of draft, critique, and correction forces the model to verify its assertions and remove safety-policy violations.

03 Key Points

01Critique Rubric: Defining explicit standards for the model to check its draft against.

02Draft Generation: Creating an initial draft response to serve as base text.

03Audit Step: Instructing the model to list all violations or defects in the draft.

04Surgical Rewrite: Recomposing the draft to address and fix each identified defect.

05Stop condition: Ending the refinement loop once all rubric items pass.

04 Model-Specific Notes

Claude is highly analytical, making it an excellent critic. It identifies grammatical and policy violations with high precision.

05 For Your Role

It's like writing a letter, reading it over to check for errors, crossing out mistakes, and rewriting a clean copy before sending.

Next: CoT Decoding Controls →