chain of thought · think before answering

For multi-step problems, telling the model to "think step by step" (or showing worked reasoning) markedly improves accuracy — it allocates intermediate computation instead of jumping to an answer. Chain-of-thought exposes the reasoning, which also aids debugging. Modern "reasoning" models internalize this. Caveat: the stated reasoning isn't always the true cause of the answer, so don't treat it as a guarantee.