Exam-style
A subscription business uses a foundation model to work out refund amounts. The policy is tiered: the rate depends on how long the subscriber has been on the plan, an unused part-month is refunded pro rata, and the whole refund is capped at one month's fee. The prompt returns a single figure, and about one refund in five is wrong, usually because a step was skipped or the cap was applied before the pro-rata calculation instead of after. The team has already put eight worked examples in the prompt, each showing a subscriber, a policy tier and the finished figure, and the error rate has not moved. What should they change?
Reveal the answer
B. Ask the model to work through the policy steps in order and show the intermediate figures before it gives the final amount
This is a reasoning failure, and examples do not fix reasoning. A few-shot prompt shows the model finished input-output pairs. That pins down a format or teaches an unusual category boundary. Single-shot does the same with one pair, and zero-shot supplies instructions and no examples at all. Eight examples of subscriber-in, figure-out have already shown the model what a right answer looks like. It still cannot get from the policy to the figure, because none of the working behind those figures is visible. Chain-of-thought prompting asks for the steps first: identify the tier, compute the rate, work out the unused days and the pro-rata amount, add them, then apply the cap and state the result. The order of operations that keeps going wrong becomes something the model writes down rather than something it has to hold. Of the rest, zero-shot removes context rather than adding it, so it makes the same failure more likely and cheaper. Temperature controls how varied the wording is, and a refund calculation wants the opposite of variety. Prompt templates are a reuse mechanism: a fixed skeleton with slots for the subscriber and the plan. They help the team ship one prompt everywhere. They do not help the model reason. Chain-of-thought is not free: the working is generated text, so it costs output tokens on every call and adds latency before the figure appears. Keep the steps in a labelled section the application strips before the number reaches a customer. The extra tokens are the price of the fifth refund being right.
Q. A prompt turns a tiered refund policy into a figure and gets one in five wrong, usually by skipping a step or applying the cap in the wrong place. Eight worked examples did not help. What next?
A. Chain-of-thought. Ask for the tier, the rate, the pro-rata amount and the cap check as numbered steps, and the final figure last.
Why? Examples and reasoning fix different faults. Few-shot examples, and single-shot with one of them, show the model what a finished answer looks like. They settle format and awkward edge cases. Zero-shot gives instructions and nothing to copy. None of them expose the working. Eight pairs teach the shape of a refund figure and nothing about how to reach it. Chain-of-thought puts the intermediate steps in the output where the order of operations can hold. It buys that with output tokens and latency on every call, which is the trade. Raising temperature adds variation to a calculation that wants none, and prompt templates standardise how the prompt is assembled and reused rather than how the model thinks.