Usage & Prompting
Chain-of-Thought
If you ask the model to think step by step instead of answering immediately, the error rate drops considerably for anything multistep. The intermediate steps also make errors findable: you can see where the chain breaks. This prompt trick has grown into a model category: reasoning models build the chain of thought in by default.
An example
“A customer has a three-year term at 49 euros, cancels after 20 months, and is two months behind on payments. What is outstanding?” Asked directly, the answer is often a wrong number. With the instruction to write out the calculation step by step, it is usually right. The reason is mechanical: every intermediate result produced then sits in the context and can be used for the next step. The model creates a notepad for itself.
A closer look
The model is prompted to spell out intermediate steps before stating the answer. This measurably helps with multistep tasks, because each step is in the context and the next one can build on it.
Why it matters
The dangerous confusion: the chain of thought reads like a justification but is not one. It is generated text like any other and can dress up a wrong answer with a plausible derivation, or attach to a correct answer a derivation that does not match the actual computation. That makes it unsuitable for accountability toward customers, regulators, or courts. As a tool for making multistep tasks more reliable, it is excellent. You just must not confuse the two.
Commonly misunderstood
On simple tasks it tends to worsen the result and costs tokens. And the written-out chain is no guarantee: it can look coherent and be quietly wrong in one step.