One of the most influential discoveries in prompting is almost embarrassingly simple: ask a model to show its work. Chain-of-thought (CoT) prompting — telling a model to reason step by step before answering — reliably improves performance on hard problems.
Why it works
A language model generates one token at a time, and each token it produces becomes context for the next. If you demand an immediate answer to a hard problem, the model has no room to work — it has to leap straight to a conclusion. Chain-of-thought gives it that room: by writing out intermediate steps, the model builds up the reasoning it needs, and each step scaffolds the next. The computation happens in the output.
Asking for the answer forces a guess. Asking for the reasoning gives the model space to actually think — on the page.
How to use it
The classic trigger is adding "Let's think step by step" or showing a few worked examples. But the deeper point generalized: modern reasoning models internalize CoT — they're trained to produce long reasoning chains automatically, no prompt trick needed. What started as a prompting hack became the foundation of test-time compute and the entire reasoning-model wave.
The limits
CoT isn't magic. Longer reasoning can also wander or confidently justify a wrong answer — length isn't correctness. The best results pair reasoning with verification (checking the chain) or self-consistency (sampling several chains and voting). Still, the core insight endures: give a model room to reason, and it reasons better. It's the small idea that unlocked much of what came next.
