Chain of Thought (CoT) Prompting
Before native reasoning models (A7) existed, the best way to make a model smarter was to force it to write out its thinking before giving an answer. This is called Chain of Thought prompting—literally just adding "think step by step" to your prompt.
If you ask a model a complex math problem and demand just the final number, it will often guess wrong because it has to calculate everything in one "token" (prediction). By forcing it to write out its steps first, it uses the generated text as a scratchpad to "show its work," increasing its accuracy exponentially.
Ask Claude a complex word problem and tell it: "Output ONLY the final answer, nothing else." Then ask it again, but tell it: "Explain your reasoning step-by-step, then give the answer." Compare the results.
Why does telling a model to "think step by step" often improve accuracy on math or logic?