Examples, Step-by-Step Reasoning and Iteration
Once you master role, context, task, format and constraints, three simple techniques can noticeably improve your results: showing examples, asking the model to work step by step, and above all iterating. Together, they turn prompting from a one-shot guess into a short, productive conversation.
In the previous lesson you learned to build a structured prompt. But even the best first prompt rarely produces the perfect result. Professionals do not expect perfection on the first try; they expect a good starting point and a method to improve it. That method is what we cover now.
1. Give examples (few-shot prompting)
Sometimes it is easier to show than to explain. With the "few-shot" approach, you provide a few examples of the expected result, and the model reproduces their style, length and logic. If you provide none, it is called "zero-shot"; with a single example, "one-shot."
Turn each customer review into one neutral sentence.
Review: "Awful delivery, I waited three weeks!!!"
Neutral: The customer reports a three-week delivery delay.
Review: "Great product but the packaging was damaged."
Neutral: The customer likes the product and reports damaged packaging.
Review: "Support never answers, this is outrageous."
Neutral:
The two examples set the format, the length and the tone far more precisely than a paragraph of explanation could. The model completes the third line in the same pattern.
How to choose good examples
- Keep them correct. The model will imitate their flaws as faithfully as their qualities.
- Make them varied. If all examples look alike, the model may over-copy their wording. Show different cases, including a tricky one.
- Keep the format consistent. Same labels, same punctuation, same structure in every example.
- Use a small number. Two to five is usually plenty for everyday tasks.
- Include a counter-example when needed. You can write "Bad: ... Good: ..." to show what to avoid.
Tip
Examples are perfect for tone of voice. Paste two emails you wrote and say: "Write the new message in the same style as these two examples." The model picks up your habits (sentence length, greetings, level of formality) without you having to describe them.
2. Reason step by step
For a logic problem, a calculation or a decision with several criteria, asking the model to lay out its approach often improves reliability. It builds its answer progressively instead of jumping to a conclusion, and each step provides context for the next one.
Before:
Which plan is cheapest over 12 months?
After:
Compare these two plans over 12 months. Work step by step:
1) calculate the total cost of each plan,
2) list the main advantages and drawbacks of each,
3) give your recommendation in two sentences.
Plan A: 9 per month, 20 setup fee.
Plan B: 12 per month, no setup fee.
Here the structure also tells you what a good answer looks like. And there is a bonus: visible reasoning lets you spot a mistake at a specific step, such as a wrong multiplication, instead of facing a bare conclusion you cannot judge.
Ways to ask for step-by-step work
- "Think through this step by step before giving your final answer."
- "First list the criteria, then evaluate each option against them, then recommend."
- "Show your calculations, then give the result on a separate line."
- "List your assumptions first."
Warning
Reasoning that looks logical can still contain an error, and a neat explanation can be wrong. Check the calculations and the important facts yourself. Asking for steps raises the chances of a good answer; it does not guarantee one. For anything involving exact arithmetic, redo the key calculation with a calculator or spreadsheet.
3. Iterate: the conversation as a tool
The first result is rarely the best. Treat the exchange as a working dialogue: read the answer critically, then give precise feedback. Compare these two reactions to a mediocre draft:
- Weak feedback: "Not good. Try again."
- Strong feedback: "The structure is good. Keep the first two paragraphs. Shorten the third to two sentences and replace the formal closing with something more casual."
Strong feedback names what to keep, what to change and in which direction. Here are the most useful kinds of follow-up instructions:
- Adjust length: "Shorten it to 100 words." "Expand the second point."
- Adjust tone: "More direct." "Warmer, but still professional."
- Target part of the text: "Keep part 2, rewrite part 3 in a more casual tone."
- Add material: "Add a concrete example from the healthcare sector."
- Ask for alternatives: "Give me three different openings."
- Challenge: "What is the weakest point in this argument?"
Let the model interview you
One of the most powerful techniques is to reverse the roles: "Ask me three questions to better understand my need before you answer." The model then asks about the audience, the purpose and the constraints you forgot to mention. Answer them, and you have just built a much richer context, which also teaches you what to include next time.
Change one thing at a time
When you iterate, adjust one variable per round if you can: tone, then length, then audience. If you change everything at once, you will not know which change helped. This habit becomes essential in chapter 3, when you start testing and improving prompts systematically.
Know when to stop
Iteration has diminishing returns. If the third round brings only small cosmetic changes, you are probably better off taking the draft and finishing it yourself. The goal is a result that serves your purpose, not a perfect one.
Tip
If a conversation gets long and confusing, summarize the decisions made so far and start again with a clean new prompt. A clear context beats a cluttered history. You will see why in the next lesson.
Putting the three techniques together
These techniques combine naturally. A realistic workflow might look like this:
- Write a structured prompt (role, context, task, format, constraints).
- Add two short examples of the style you want.
- For a decision or a calculation, ask for step-by-step work.
- Read the answer, then iterate with precise feedback, one change at a time.
- Check the facts and numbers yourself.
Your turn: examples and iteration
Exercise 1. Write a few-shot prompt that turns rough meeting notes into action items in the format "Who - What - By when." Include two examples, then one new set of notes.
Exercise 2 (project). Take a short text you wrote (a message, a paragraph from a report). Ask the AI to rephrase it, then iterate three times, changing one thing each time: the tone, the length, then the target audience. Note which instruction had the biggest effect and add the best final prompt to your notebook.
See a solution
A possible few-shot prompt for Exercise 1:
Turn meeting notes into action items. Format: Who - What - By when.
If a deadline or owner is missing, write "to be defined".
Notes: "Sam will send the budget next week, and we need someone
to book the room."
Action items:
- Sam - Send the budget - next week
- To be defined - Book the room - to be defined
Notes: "Lina is checking the contract by Friday."
Action items:
- Lina - Check the contract - Friday
Notes: [paste your new notes here]
Action items:
Notice the rule for missing information: it prevents the model from inventing owners or dates. For Exercise 2, people often find that changing the audience produces the most dramatic difference, because it affects vocabulary, structure and level of detail all at once.
Key takeaways
- A few good examples (few-shot) guide style and format better than a long explanation; keep them correct, varied and consistent.
- Asking for step-by-step reasoning helps on multi-criteria problems and makes errors easier to spot, but does not guarantee accuracy.
- Iterate with precise feedback (what to keep, what to change), ideally one change at a time, instead of starting over from scratch.
- Let the model ask you questions to clarify your need, and know when to stop and finish the work yourself.