We Do Not Teach Thinking to AI

Short summary

Reasoning models like Claude's extended thinking don't benefit from step-by-step prompting guidance; they excel when given clear goals and success criteria. Shift from directive prompts (specify the thinking) to outcome-based prompts (define what good looks like). This evolution unlocks reasoning models' potential to find superior paths independently.

• •Reasoning models are built to solve problems through internal deliberation, not external step-by-step direction
• •Over-specifying thinking processes or examples constrains reasoning models and prevents them from finding better approaches
• •Best practice: provide clear goals, success criteria, and output format—then let the model handle the reasoning

Generated with AI, which can make mistakes.