Chain-of-Thought Prompting: Reasoning That Actually Improves Answers
Chain-of-thought prompting improves multi-step reasoning, but only when applied carefully. Here is how I use it without wasting tokens. Chain-of-thought prompting is the practice of asking the model to show its reasoning before giving its answer. For tasks that require multi-step logic, math, or deduction, a prompt that asks for the answer directly produces worse results than one that asks the model to reason step by step. The improvement is large enough that chain-of-thought has become a standard technique, but it is often applied carelessly, which produces long outputs that reason in circles without reaching better answers. This guide covers how I use chain-of-thought prompting to get reasoning that actually improves results. The Basic Trigger Phrase The simplest form of chain-of-thought is a single phrase appended to a question. Asking the model to think step by step before answering causes it to generate intermediate reasoning that improves the final answer. The phrase costs almost nothing in effort and reliably improves accuracy on multi-step tasks. // Basic chain-of-thought trigger A store sells pencils at 2 for 75 cents. How much do 12 pencils cost? Think step by step, then give the final answer. Without the trigger, the model often jumps to a plausible-sounding but wrong answer. With the trigger, it works through the unit price, multiplies by twelve, and arrives at the correct total.