Chain-of-Thought Prompting

A customer returns an opened wheelset and unopened gloves. Pedal Point's rules: opened items that are not faulty lose a 10% restocking fee, unopened items come back in full, shipping is never refunded. Asked for "just the refund amount", the assistant sometimes applies the fee to both items. The maths is easy for a person working step by step. So let the model work step by step too.

The rule

For multi-step problems, let the model work before it answers.

A model writes one token at a time, and each token is shaped by the text before it. If the very first token must be the final answer, all the reasoning has to happen invisibly inside that one step. If the model first writes out the intermediate steps, each later step can build on them.

Chain-of-thought (CoT) prompting asks for exactly that:

For the wheelset at 250 and gloves at 40:

• Wheelset: opened, not faulty, fee 10% of 250 = 25, refund 225.
• Gloves: unopened, refund 40.
• Total: 225 + 40 = 265.

Some models are built to reason internally before answering and need less prompting for this; the principle is the same.

Use CoT where it pays:

• Helps: calculations, policy rules that combine, planning, multi-step decisions.
• Wastes: simple lookups and classifications. Reasoning costs extra tokens and time on every request.

Two cautions. The written reasoning is not a guaranteed record of why the model reached its answer, so do not treat it as proof. And for anything involving money, recompute the result in code.

The shortcut

Think first, answer last, in a fixed spot. Ask for the steps, then the result on a line or field your code can read.

The trap is showing customers the whole scratchpad. They want the answer; keep the working in your logs, where it helps you debug a wrong total later.

Drills

Answer fast — trust the rule.

Remember

• For multi-step problems, let the model work before it answers.
• Think first, answer last, in a fixed spot.
• Reasoning is not proof: recompute money in code.