Step-by-step reasoning prompts

Some questions can't be answered well in one jump: a word problem, a bug that depends on three things, a choice between two job offers. People solve these by working through them on paper. Models do better when they do the same.

In this tutorial you will try step-by-step prompts on a maths problem, a logic puzzle and a real decision, then learn how to check the reasoning instead of trusting it. You need a free chat tool (ChatGPT, Gemini, Claude or Copilot). You should know the basics from *Anatomy of a good prompt*.

A language model writes its answer one piece at a time, and each piece is based on everything written so far. If it states the final answer first, it has done no working. If it writes the working first, every step gives it something correct to build on.

This is called chain-of-thought prompting: you ask the model to show intermediate steps before the answer. It helps most on tasks with several steps (arithmetic, logic, planning, debugging). It helps little on simple facts or rewriting a sentence.

Brew & Bean sold 14 cappuccinos at ₹180, 9 cold brews at ₹220 and 23 filter coffees at ₹80 on Saturday. A customer coupon took ₹150 off the total. What was the day's revenue? Reply with just the number.

What to expect: The correct answer is ₹6,190: 14 x 180 = 2,520; 9 x 220 = 1,980; 23 x 80 = 1,840; total 6,340; minus the ₹150 coupon = 6,190. Models told to reply with just a number sometimes get this wrong, which is the point.

Brew & Bean sold 14 cappuccinos at ₹180, 9 cold brews at ₹220 and 23 filter coffees at ₹80 on Saturday. A customer coupon took ₹150 off the total. What was the day's revenue?

Work through it step by step: first the revenue for each item, then the total before the coupon, then subtract the coupon. Show each calculation, then give the final answer on its own line starting with "Answer:".

What to expect: Expected working: 2,520 + 1,980 + 1,840 = 6,340, minus 150 gives Answer: ₹6,190. Naming the steps yourself is stronger than a bare "think step by step".

Three friends, Asha, Bilal and Charu, each ordered a different drink from Brew & Bean: an espresso (₹120), a masala chai (₹90) and a cold brew (₹220). Clues:
1. Asha did not order the masala chai.
2. Bilal ordered the cheapest of the three drinks.
3. Charu did not order the cold brew.

Who ordered what? Reason step by step, listing what each clue rules out. Then check your answer against all three clues one by one and say whether each clue holds. Finally, say whether the answer is the only possible one.

What to expect: Answer: Bilal has the masala chai (clue 2, the cheapest at ₹90). Charu is not cold brew (clue 3) and chai is taken, so Charu has the espresso. Asha has the cold brew. Clue 1 holds (Asha did not have the chai) and the answer is unique. If the model skips the uniqueness check, ask again.

Notice the last request: "say whether the answer is the only possible one". Models often state one answer with confidence without checking whether the clues allow another. Asking for a uniqueness check is a cheap way to catch this.

When you write your own puzzles for a model, test them first: a question with two valid answers makes any reply look right or wrong by luck.

Many real questions have no single correct answer. For these, ask for reasoning you can argue with: the options, the criteria, how each option scores on each criterion and what could change the result.

I need to decide: [the decision, e.g. accept a 4-month unpaid internship at a startup or do a paid certification course].

My situation: [your context: year, finances, goals, deadlines].

Do this in order:
1. List the options, including one I may not have thought of.
2. List 4-5 criteria that matter given my situation, and say which matter most.
3. In a table, rate each option against each criterion (high / medium / low) with a short reason.
4. State your recommendation and the 2 facts that, if different, would change it.
5. List what I should check myself before deciding.

What to expect: A useful answer is specific to your context, shows the table, names what would change the recommendation and tells you what to verify. Treat it as a thinking aid, not a ruling.

Here is a problem and your answer.

Problem: [paste problem]
Your answer: [paste answer and steps]

Act as a strict reviewer. Check each step. List any step that is wrong, unsupported or assumes something not stated. If the answer is correct, say so and explain how you verified it.

What to expect: A good reply goes through the steps one by one. Be wary if it simply says "looks correct" without checking anything; ask it to recompute the numbers.

Interviewers for AI-related roles ask "How do you improve an LLM's accuracy on multi-step tasks?" A good answer: ask for intermediate steps (chain-of-thought) or use a reasoning model, give worked examples in the prompt, verify with code or a second pass for calculations, and measure on a test set rather than relying on a few good runs (see *Testing and evaluating prompts*). Also mention the cost: more reasoning means more tokens, so slower and more expensive calls.