Don’t Just Think, Focus—Unlock the Magic of Reasoning Models with Chain-of-Refinement Prompting

Source: Nate’s Newsletter Date: 2025-04-30 URL: https://natesnewsletter.substack.com/p/dont-just-think-focusunlock-the-magic

Summary

Nate introduces “Chain-of-Refinement” (CoR) prompting as a successor to Chain-of-Thought for modern reasoning models like o3 and Gemini 2.5 Pro. Where CoT tries to teach models to simulate deliberate thinking, CoR treats the model’s reasoning capacity as a given and instead provides structured checkpoints that direct and refine outputs iteratively — essentially meta-prompting that guides the model through latent-space concept processing in multiple passes. The author’s key counterintuitive point is that effective CoR demands more precision from the user, not less.

Implications

• Reasoning models (o3, Gemini 2.5 Pro) have shifted the effective prompt contract: the limiting factor is now human specification quality, not model reasoning depth.
• CoR is practically significant for agent loop design — multi-step refinement checkpoints map naturally onto orchestration patterns where intermediate outputs need human or tool-mediated evaluation before proceeding.
• Feeds thread: prompting methodology evolution / reasoning-model operator patterns.