Step-by-Step Reasoning: Breaking Big Tasks Down

Introduction

Ask a person a tricky word problem and demand an instant answer, and they will often blurt out the intuitive — and wrong — response. Give them a moment to work it through on paper and they usually get it right. Language models behave in a similar way, and this lesson shows how to give them that “paper”.

Why Reasoning Steps Help

Remember that a model produces its answer one token at a time. If the very first words must be the final answer, the model has no room to work anything out first. When it writes intermediate steps, each step becomes part of the context and helps produce the next, much like your notes on paper. This technique is often called chain-of-thought prompting.

Many modern models — often called reasoning models — already think through hard problems internally before answering. Even with them, asking for steps is useful when you want to check the logic yourself, and breaking big tasks into smaller pieces still improves results.

Asking for Steps

The simplest version is to add “Work through this step by step before giving your answer.” Better still, tell the model which steps to take when you know them — that combines its reasoning with your expertise.

Decide whether this customer qualifies for a refund.
Work through these steps in order:
1. Find the purchase date and today's date in the message.
2. Check whether it is within our 30-day refund window.
3. Check whether the item is in the "non-refundable" list.
4. Give the decision: Refund or No refund, with one reason.

Guided steps make the model consistent from one request to the next, and when it gets something wrong, you can see exactly which step failed — useful for the iteration process in the next lesson.

Not every request needs reasoning steps. For simple tasks — translating a sentence, fixing a typo, writing a short caption — asking for step-by-step working just makes the answer longer and slower. Save it for problems with several dependent steps: calculations, comparing options against criteria, applying rules to a case, planning, or debugging. A quick test: if you would need scrap paper to solve it yourself, the model will probably benefit from writing its steps too.

Separating Working from the Answer

Step-by-step working is useful to check but annoying if a program, or a busy reader, only needs the result. Ask for the two parts separately, for example with the tags from Lesson 5: “Put your working in <thinking> tags and only the final decision in <answer> tags.” Your code can then read just the <answer> section, while you can still inspect the reasoning when something looks wrong.

Do not treat written reasoning as proof. A model can write convincing steps that contain a mistake, or reach the right answer through flawed logic. Steps make errors easier to spot; they do not remove the need to check results that matter.

Breaking Big Tasks into a Chain

Some tasks are too big for one prompt: researching and writing a report, translating and then reviewing a document, or turning raw notes into a polished article. Asking for everything at once produces shallow work. Instead, split the job into stages and give each stage its own prompt, passing the output of one into the next. This is called prompt chaining.

Chaining has three benefits: each prompt is simpler, so quality goes up; you can review and correct the work between stages; and when something goes wrong, you know which stage to fix.

Key Takeaways

Summary

Reasoning steps and prompt chains help the model handle problems that are too big or too tricky for a single quick answer. Next you will learn a systematic way to improve any prompt that is not working: diagnose the problem, change one thing, and test again.