Module 3 · Talking to AI: prompting · Lesson 15 of 32
Let it think step by step
9 minFree GenAI course · BeginnerLesson 3.3By Saket Bhatnagar
Researched and drafted with AI, reviewed and edited by Saket Bhatnagar.
In short
- Asking the AI to work through a problem in steps before answering makes it more accurate on multi-step tasks.
- The written steps become part of the text the model reads, so it can build on them like rough work.
- It costs more tokens and time, and it adds little to simple lookups or creative writing.
Video lesson · coming soon
Chain-of-thought prompting: why “think step by step” works
Lala, your AI mentor, will walk you through this lesson in a short animated video. Until then, everything is here to read and try.
- 01 The ₹110 bat and ball puzzle
- 02 Mrs Iyer’s rough-work rule
- 03 Steps become rough work in my context
- 04 Solving the chai flask problem
- 05 Reasoning models think on their own
- 06 Use steps for multi-step problems
- 07 Try it, then get clean answers
Try this before reading on. A bat and a ball cost ₹110 together. The bat costs ₹100 more than the ball. How much is the ball?
If ₹10 jumped into your head, you are in very good company. Most people say it. But then the bat would be ₹110 and the total ₹120. The ball is ₹5. The quick answer felt right and was wrong, and AI models fall for the same kind of trap.
“Show your working”: why steps help
Mrs Iyer runs a maths tuition class in Chennai. Her one rule: no marks for an answer without working. Her students grumble, until they notice that writing each step is exactly how they catch their own slips.
Here is why it works for a model. It writes one token at a time, and each token gets about the same amount of computing. If the very first token must be the final answer, the whole problem has to be solved in that one go. That is a lot to ask.
When the model writes steps first, each step lands in the context window. Thanks to attention, every later token can look back at those steps. The model is not thinking harder per token. It is giving itself rough work to lean on, spread over many tokens.
Jumping to the answer vs working it out
Drag to exploreThe direct path asks one guess to do all the work. The step path writes each result down, so later steps can use it.
Diagram connections:
- Word problem to Direct guess (rushed)
- Direct guess to Final answer (risky)
- Word problem to Step 1 written (think)
- Step 1 written to Step 2 written
- Step 2 written to Quick check
- Quick check to Final answer (confirmed)
- Word problem: A question that needs a few calculations in a row.
- Direct guess: The answer is written first, so the whole problem must be s…
- Step 1 written: The first result is now in the context for later tokens to…
- Step 2 written: Builds on step 1 instead of holding everything at once.
- Quick check: Put the answer back into the question and see if it fits.
- Final answer: More often right on multi-step problems.
Watch a problem get solved step by step
Neha is ordering chai for her office Diwali party in Nagpur. Eighteen people are coming. Each will drink two cups. One flask holds six cups and costs ₹150. Here is how a step-by-step prompt handles it.
None of these steps is hard alone. What helps is that no step has to hold everything at once. That is the whole idea, and it scales. Planning a trip, finding a bug in code, or comparing three phone plans all break into small steps the same way.
How do you ask for step-by-step thinking?
The shortest version is one line added to your prompt. Researchers found in 2022 that adding a phrase like “Let’s think step by step” noticeably improved accuracy on maths and logic questions. You can do better by saying what the steps should be.
I want to travel from Delhi to Jaipur and back on a Saturday.
Budget: ₹2,000 for travel. I must be home by 10 pm.
Before answering, think it through step by step:
1. List the travel options and rough fares.
2. Check each one against my budget and the 10 pm limit.
3. Pick the best option and say why.
Then give your recommendation in two lines.You can also mix this with the last lesson. Show one worked example with its steps, and the model will copy the step style on the new problem. That is few-shot chain-of-thought, and it is how the technique was first studied.
Reasoning models and when to skip the steps
Since late 2024, many AI makers have trained models to do this thinking on their own. Chat apps often show it as a “thinking” or “reasoning” mode. You ask a question, the model works through it at length, and then answers.
With a reasoning model, you rarely need to say “think step by step”. It already does. Spend your words on a clear goal, the limits and what a good answer looks like. And remember that the thinking still uses tokens, so it is usually slower and costs more.
| Task | Use steps? | Why |
|---|---|---|
| Word problems and bills | Yes | Several calculations in a row, where one slip ruins the answer |
| Planning a trip or a timetable | Yes | Many limits to juggle: time, money, order |
| Finding a bug in code | Yes | Tracing what happens line by line catches the mistake |
| Translating a sentence | Usually not | One step, so steps add tokens without adding accuracy |
| A birthday wish or a poem | Usually not | Creative work has no single right answer to reason towards |
| A quick fact like a capital city | No | Either the model knows it or it does not; steps won’t help |
Check yourself
Check yourself
Which task gains the most from asking the AI to think step by step?
Pick an answer to see why it is right or wrong.
Check yourself
Why does writing steps make a model more accurate?
Pick an answer to see why it is right or wrong.
Remember these three
- For multi-step problems, ask the AI to show its working before the final answer.
- Written steps become rough work in the context that later tokens can lean on.
- Reasoning models do this on their own; skip it for one-step or creative tasks, and still check what matters.
Questions people ask
- What is chain-of-thought prompting?
- It is asking an AI to write out its reasoning in steps before giving the final answer. It improves accuracy on maths, logic and planning tasks, because each written step helps the next one.
- Does “think step by step” still work in 2026?
- It still helps with regular chat models on multi-step problems. Reasoning models already think before answering, so with them a clear goal and clear limits matter more than that phrase.
- What is a reasoning model?
- A reasoning model is trained to work through a problem at length before answering. Chat apps often call this a thinking mode. It is better at hard multi-step problems but slower and more costly per answer.
- Is step-by-step reasoning always correct?
- No. The steps can look neat and still contain a wrong number or a wrong assumption. Check important results yourself, and use a calculator or code for exact maths.