Chapter 2. Asking and Refining Prompts

Learning objectives

• Write questions that clearly state the desired output.
• Provide the situation and audience (who will read it and who it is for) to tailor the response to your circumstances.
• Specify the length and format so that the received response is immediately usable.
• Rewrite vague questions into specific questions.
• Review the initial response, pinpoint what is lacking, and ask for follow-up revisions.
• Recognize when to start a new conversation, restate the necessary conditions, and start fresh.

Introduction

Lee Chun-ja was agonizing over what gift to give for a friend's birthday next month. She wrote "Recommend a gift" to an AI chatbot. A long list came back with items such as wireless earbuds, a smartwatch, perfume, and fine wine. Every single one was far removed from what her friend would like. Just as she was about to close the app, thinking, "As expected, machines don't understand how people feel," her daughter beside her spoke up. "Mom, the AI doesn't know who your friend is. Try speaking to it just as you would ask a person."

Lee Chun-ja wrote again. "I am trying to pick a birthday gift for a longtime friend who is turning seventy. My friend has bad knees, so these days she finds joy in growing houseplants at home. The budget is around KRW 50,000, and I will hand it to her in person. Please suggest only five options briefly, along with the reasons." This time, candidates such as a small plant stand, a watering can, a lap blanket, and a plant label making kit came up along with reasons. Lee Chun-ja picked two of them and stopped by a flower shop.

What changed was not the AI, but the question. This chapter covers how to write good questions and how to ask for revisions when the initial answer is not satisfactory. First, you will learn the basics of using the chat window, and then examine one by one the five elements to include in a question (what, why, to whom, how much, and in what format). Next, it covers exercises in rewriting vague questions, phrasing for refining the initial answer, and when to start a new conversation. This chapter carries the greatest weight on the exam.

Key concepts

Using the chat window for the first time

The screen of an AI chatbot app looks slightly different depending on the company, but the basic structure is similar.

1. At the bottom of the screen is an input box. You type your question here. Many apps also allow you to input by voice by pressing a microphone button instead of using the keyboard.
2. When you finish writing your question, press the Send button (usually shaped like an arrow).
3. The AI's response appears above. If the response is long, scroll up on the screen and read to the end.
4. If you have more to say, type again in the same input box. The AI will then read the earlier conversation in context. This is called asking follow-up questions.
5. If you want to talk about a completely different topic, press the New chat button to start in a new window. The location and appearance of this button vary by app, but it is usually at the top of the screen or inside a side menu, labeled "New Chat" or "New Conversation", or drawn as a "+" or pencil icon.

A question is a request to a stranger

When you get into a taxi and say, "Please take me there," the driver does not know where to go. You must say, "Please drop me off in front of the bank building at the big intersection in front of City Hall." It is not because the driver does not know the streets, but because the driver does not know where you want to go.

The same is true for AI. AI has read many texts in the world, but it does not know the circumstances of this specific request unless you provide them. It does not know how old your friend is, what your budget is, or who will read this text. Depending on settings, some apps may know a little about you through memory features or by referencing previous conversations, but they still do not know the circumstances of this specific request. The AI fills in unknown parts with the most common and generic options. Therefore, if you ask vaguely, you get an answer that applies to anyone—and thus fits no one perfectly.

The Google Workspace Training Center's "Gemini prompting tips" also recommends the same direction. It advises writing naturally as if speaking to a friend or coworker, avoiding ambiguous language, and notes that the more context (situation) you provide, the better the AI understands the request and produces a useful answer.

A good question is like a letter of request to someone you are meeting for the first time. It contains what you are requesting, why it is needed, who it is for, how much, and in what format you want to receive it. Let us examine these five elements one by one.

Stating the desired output specifically

The first ingredient of a question is what you want to receive. "Recommend a gift" is unclear about what you want to receive. It cannot tell whether you want a list, a single selection, or even information on where to buy it.

The method for clearly stating the deliverable is simple. Say "what" + "make this" together.

When you state the deliverable, the AI also knows where to stop. If you say "a text message," it will not produce three pages of letters.

If your request involves text that the AI needs to read, such as "Extract only the tasks I need to do from this notice," you must copy and paste the text of that notice or upload it as a photo along with the question. The AI cannot see the paper in your hand. When uploading a photo, cover names or numbers first (Chapter 4).

Providing the situation

The second ingredient is why, and in what situation it will be used. Even for the same "thank-you text," the content is completely different depending on the situation.

• A text replying to a greeting from a neighbor who shared rice cakes on moving day
• A thank-you text sent to a friend who attended a funeral
• A text sent to an alumnus who sent a gift for a grandchild's first-birthday party

The situation includes the following elements:

There are times when you need to state health circumstances, such as "Has diabetes, so avoid sweet foods." In such cases, there is no need to include names or contact information. The key is to write only the necessary circumstances and leave out information that can identify who the person is. This is covered in detail in Chapter 4.

Specifying the audience

The third ingredient is who it is for and who will read it. Even with the same content, the tone and vocabulary must change depending on the reader.

Phrasing to specify the audience is simple. A single sentence is enough, such as "My second-grade granddaughter will read it," or "I will post this in a group chat for a gathering with many seniors in their 70s." It is also good to specify the tone of voice: "in polite honorifics," "in an affectionate informal tone," or "not too stiff."

Setting the length

The fourth ingredient is how much. If you do not specify the length, the AI usually writes at length. If you need a single text message but receive a one-page letter, you have to shorten it again.

Examples of phrases that set the length include:

• "Within three lines"
• "The length of one text message"
• "Around 100 characters"
• "Only five candidates"
• "In one sentence"
• "One line for each item"

Specifying with numbers is followed much better. "In three lines" is clearer than "briefly." However, since AI sometimes fails to hit the exact character count, count them yourself after receiving the response if the length is a strict requirement.

Setting the format

The fifth ingredient is what form you want to receive it in. Match the format to where you will use the received answer.

For example, if a notice to be posted in a group chat contains a table, the table may look broken on mobile phones. When you tell it where it will be used, the AI selects a shape suited for that destination.

The five-slot question grid

If we combine the five ingredients we have seen so far into a single table, it looks like this. Try filling out this table in your head before writing a question. You do not need to fill in every box. When you do not like the answer, you can fill in the empty boxes one by one.

Turning this table into a question results in the following:

[Prompt sent]
A friend of 50 years is turning seventy, but I cannot attend the party because of knee surgery.
Write a text message expressing both congratulations and my apologies for being unable to go.
In a comfortable casual tone, within five lines, and with only about one emoji since it will be sent via messenger.

[Sample AI response]
Young-sook, happy 70th birthday!
I really wanted to come to the party, and I am so sorry I cannot go because of my knee.
It feels like only yesterday when we first met at sixteen, and now it is already our 70th birthday.
I hope you stay healthy and have only days full of laughter, just like now.
Once my leg is completely healed, I will treat you to something delicious ^^

Turning vague questions into specific questions

Now, let us rewrite vague questions using the five-box question table. When revising, first think, "What does the AI not know?" and fill in those blanks.

Example 1. Wondering about side dishes

• Before revision: "What side dishes should I make?"
• What the AI does not know: How many people will eat, who will eat, ingredients to avoid, timeframe, format
• After revision: "Plan a week of dinner side dishes for a household of two (a couple in their 70s). My husband needs to cut back on salty foods, and I want to shop for groceries once and use them for the whole week. Present it as a table by day of the week, and attach a grocery shopping list at the end."

Example 2. Explaining to a grandchild

• Before revision: "What is photosynthesis?"
• What the AI does not know: Who this will be explained to, how simply, how long
• After revision: "I want to explain photosynthesis to my grandson in the third grade of elementary school. Use a houseplant at home as an example, keep it within five sentences, and write it in an informal tone as if speaking directly to my grandson."

Example 3. Secondhand transaction

• Before revision: "Write a post to sell a bicycle."
• What the AI does not know: Condition of the item, period of use, flaws, price, transaction method
• After revision: "I want to sell a bicycle on a secondhand marketplace app. I rode it for three years, replaced the tires last year, and there is a scratch on the front basket. The price is KRW 80,000, and I only do in-person transactions. Write a sales listing within ten lines, without hiding flaws and being honest."

Example 4. Before going to the hospital

• Before revision: "My knee hurts."
• What the AI does not know: What is wanted (a diagnosis? hospital preparation?), format
• After revision: "I am going to an orthopedic clinic next week because of my knee. Please create a list of only five questions to ask the doctor. Do not make a diagnosis, only provide a list of questions."

Take note of Example 4. It defined the deliverable as: "Do not make a diagnosis, only provide a list of questions." As seen in Chapter 1, conclusions about your body are made by a doctor, not an AI. Instead, AI is very helpful for organizing what to ask at the hospital. When you define the deliverable clearly, you end up using the AI for what it does well.

Refining prompts based on the initial response

It is rare for the initial answer to be completely satisfactory. Rather than struggling to write a perfect question from the start, it is faster to ask for follow-up revisions after receiving the initial response. It is like trying on clothes during a fitting and saying, "Please shorten only the sleeves a little."

There are three tips when asking for revisions:

1. Pinpoint exactly what is lacking. Saying "It is too long, and the second line is stiff" is better than saying "It is no good."
2. Ask to keep the good parts. Saying "The first line is good, so leave it as is" prevents the AI from changing the parts you liked.
3. Do not change too much at once. If there are many things to fix, ask for one or two things at a time, starting with the most important.

Try this: Refining a congratulatory text message

[Initial prompt]
Write a congratulatory text message for my granddaughter entering elementary school.

[Sample initial response]
To my beloved granddaughter,
I sincerely congratulate you on finally entering elementary school.
I hope that you make many friends in this new environment, discover the joy of learning,
and grow healthy in body and mind. Grandmother will always pray for you.
...

The initial response is not bad, but the vocabulary is difficult and long for an eight-year-old child to read. Follow up by writing like this:

[Follow-up prompt to refine]
It is good, but my granddaughter is eight years old, so she will read it herself.
Leave out difficult words ("environment", "grow"), and shorten it to three lines.
Keep the sentiment of "Grandmother will always pray for you."

[Sample revised response]
My sweet child, congratulations on going to school!
Have fun playing with your new friends and enjoy learning.
Grandmother will always pray for you. I love you!

The second request contains what was lacking (difficult words, length), the reason (an eight-year-old will read it directly), and what to keep (the sentiment of praying). As a result, it came much closer to the desired answer in one go.

Useful phrases for refining prompts

Here is a collection of frequently used phrasing for refining prompts. You can use them directly as they are.

Look at "Contains incorrect information" above the last row. If a fact you already know is incorrect, you can provide the correct fact and have the AI fix it. Conversely, when you wonder whether a fact you do not know is correct, asking the AI for revisions will not solve it. At that point, verify it using the methods in Chapter 3.

Instructing the AI to ask back and admit when it does not know

There are times when it is difficult to write all the information essential to a question. There are even times when you do not know what to write. In such cases, you can ask the AI to ask you first.

[Prompt sent]
I have to give a congratulatory speech for my nephew who is getting married next month.
Ask me five questions first about what I need to tell you so you can write a good speech.
Once I answer, write the speech after that.

The AI will ask questions like, "What memories do you have with your nephew?" and "About how many minutes is the speech?" As you answer them one by one, the five-box question table fills up naturally.

There are also phrases to prevent the AI from making things up when it does not know. These are the two sentences foreshadowed in Chapter 1:

• "If you are not certain, please say you do not know."
• "Do not make up parts you are not certain about; mark them as 'Needs verification'."

As seen in Chapter 1, AI tends to fill in unknown blanks with plausible words. Adding these phrases increases the likelihood of receiving an answer where parts that need verification stand out clearly. However, these phrases do not completely prevent incorrect answers. Even for parts without a "Needs verification" label, verify them if they fall under the six checklist questions from Chapter 1.

Starting a new conversation

Within a single chat window, the AI reads what was said earlier in context. That is why asking follow-up revisions is possible. However, because of this characteristic, there are times when it is better to start a new conversation.

How much a new conversation knows about earlier conversations depends on the app and settings. If the memory feature or the feature referencing previous chats is turned on, parts of earlier conversations may carry over into the new conversation. For example, OpenAI Help's "Memory in ChatGPT" explains that when the memory feature is turned on, ChatGPT can remember relevant preferences and details from conversations, and when the "Reference chat history" setting is turned on, it can use relevant information from past chats in future responses. When these features are turned off, a new conversation knows nothing about previous conversations. Either way, it is most reliable to restate the essential conditions in the new conversation. At this point, it is convenient to copy the answer you liked or the organized conditions from the earlier conversation and say, "Rewrite this with these conditions."

Try this: Organizing a tangled conversation into a new chat

1. Briefly summarize the conditions established in the previous conversation. Example: "Gathering name: Neulbom Hiking Club, second Saturday of November, 9:00 AM, in front of the community center, membership fee KRW 20,000, pack your own lunch."
2. Open a new conversation.
3. Write all at once following the five-box question table: "Write a notice to post in the hiking club members' group chat. Include all the conditions below, break lines short with simple words since most members are in their 70s, and do not use symbols or tables."
4. Paste the summarized conditions below that and send it.

Going deeper

How many times to refine a prompt

You can refine a prompt multiple times. However, depending on the service and subscription plan, there may be limits on usage volume or the length of a single conversation. And if you are refining the same task for the tenth time, it is time to change your approach. It is good to keep the following criteria in mind:

1. If the same area remains unsatisfactory even after refining three times, open a new conversation and write again from the beginning using the five-box prompt worksheet. In most cases, a box was missing from the initial prompt.
2. If it still does not work in a new conversation, provide an example of the format you want (see the next section).
3. If it still does not work, writing it directly yourself may be faster. You can also take just a few good sentences generated by the AI and blend them into your own writing.

The purpose of using AI is to save time and effort. If refining prompts takes more time than writing the text directly, that task may not be well suited for AI. That, too, is a useful finding.

Showing an example is fast

There are nuances that are difficult to explain in words. Examples include "the way my family talks" or "the feel of text messages I usually send." In such cases, showing an example is fastest.

[Outgoing prompt]
This is the tone I usually use in my family group chat:
"Did everyone eat?~ Let's have lunch at Mom's place this Sunday. I'll make the side dishes!"
In this tone, write an announcement for my father's birthday gathering next month.
Leave [Date] and [Location] blank because I will fill them in later.

When you provide an example, the AI mimics habits down to sentence length, endings, and even punctuation like the tilde (~). One or two examples are sufficient. Also note the last line of this prompt. It asked to leave the date and location blank. This ensures that the AI does not make up the date and location, while you fill in the accurate facts yourself. In effect, it puts the "three-box division" from Chapter 1 into practice within the prompt.

At this time, verify that the example does not contain other people's names, phone numbers, or private stories. If you copy and paste an entire family group chat, family members' names and personal circumstances will be included. Select only one or two sentences that demonstrate the tone, remove the names, and use those.

Do not disclose too much

This chapter has repeatedly emphasized "be specific." However, while being specific, people naturally tend to share a lot of personal stories. These include details such as a grandchild's name, apartment complex, medical conditions, and bank balance.

What is needed to receive a good answer is usually circumstances, not identity.

You can receive a good answer with only the column on the left. Including the items on the right does not improve the answer and only increases risk. This principle will be practiced again in Chapter 4 as "masking before asking."

At work

Group announcement by Seo Gyeong-ja, manager of Neulbom Hiking Club

Situation. Seo Gyeong-ja was the manager of the 30-member Neulbom Hiking Club. The November hike was postponed by one week due to rain in the forecast. Because most members were in their 70s, dozens of phone calls would come in if a group chat announcement was long or confusing. Seo Gyeong-ja wrote to the AI, "Write an announcement postponing the hiking club's hike."

First response. The response was polite, but had three problems. First, there were many difficult expressions, such as "Please be advised that the schedule has unavoidably been deferred due to inclement weather." Second, the AI made up a new date, inserting "the third Sunday of November." In reality, it was Saturday. Third, it contained a table that caused lines to break awkwardly on mobile phones.

Discarded options.

• Post the first response as-is and manually correct only the date. She discarded this because difficult words and broken tables would remain, causing a flood of calls.
• Ask again, saying "Make it better." She discarded this because if she did not explain what was lacking, the AI might change the wrong parts.
• Give up on AI and write it directly. She discarded this because she lacked time, and refining sentences was something AI did well.

Decision. Seo Gyeong-ja refined her prompt in succession using the five-box prompt worksheet: "Most members are in their 70s, so change it to simple words (Audience). The date is the third Saturday of November; use only the facts I stated, and Don't make things up for the date (Fact correction). Format in short lines without tables because it will be pasted into a group chat (Format). Within five lines (Length)." The new announcement was short and clear. Before posting it, Seo Gyeong-ja checked the date and meeting time against the calendar once more. Only two phone calls came in that day.

Second-hand sales post by Lee Chun-ja

Situation. Lee Chun-ja decided to sell a stroller on a second-hand marketplace app because her grandchildren had outgrown it. When she asked the AI, "Write a sales post for a stroller," an exaggerated post appeared with phrases such as "Top-tier stroller, almost like new! Mint condition!" The actual stroller had been used for five years, and the handle leather was slightly peeling.

Thought process. Lee Chun-ja thought that selling it with an exaggerated post could cause a dispute with the buyer later. The AI merely imitated the tone common in sales posts; it did not know the actual condition of this stroller. The only person who knew its condition was herself.

Discarded options.

• Use the exaggerated post as-is. She discarded this because a post that departed from facts would deceive the buyer.
• Ask the AI to "write it so it sells better." She discarded this because an even more exaggerated post could result.

Decision. Lee Chun-ja refined her prompt as follows: "Do not exaggerate; write truthfully. Used for 5 years, handle leather slightly peeling, wheels and folding mechanism work well, price KRW 30,000, in-person transaction only. Do not hide flaws; disclose them upfront. Within ten lines." The new post was honest and clean. Lee Chun-ja also took photos of the actual item and uploaded them without enhancing them with AI. Because she disclosed even the flaws, the buyer completed the transaction without hesitation.

Common mistakes

1. "Because AI is smart, it will figure things out on its own even if I ask briefly." The AI does not know the circumstances of this request unless you tell it. The same applies to apps with memory features. Because it fills in unknown parts with what is most common, vague questions lead to vague answers.
2. "Being told to ask specifically means writing all my information." What is needed are the circumstances (age group, conditions, budget), not personal identity (name, address, phone number).
3. "I must write a perfect question from the start." It is faster and easier to look at the first response and refine your question. A question is completed through conversation.
4. "If I don't like it, all I have to do is say 'Rewrite this'." If you do not state what is lacking, the AI will change the wrong parts. State both what needs improvement and what should be kept.
5. "Refining the question can also correct facts." While you can have it make corrections by providing facts you know, whether facts you do not know are correct cannot be verified by refining the question.
6. "Even if I open a new conversation, the AI will carry over the conditions stated earlier on its own." Depending on the app and settings, some may carry over or nothing may carry over at all. Do not rely on either possibility; write down the necessary conditions again.
7. "It is convenient to do everything in a single chat window." If topics get mixed, irrelevant answers will appear, and if the conversation becomes long, previous conditions may be missed. Open a new conversation when the topic changes.
8. "Now that I wrote 'Don't make things up', there will be no incorrect answers." Such requests help, but they cannot completely prevent incorrect answers. Verification is still necessary.

Summary at a glance

Check the official docs

• Google Workspace Learning Center — Tips to write prompts for Gemini: Speaking naturally, avoiding vague language, providing context, and breaking multiple tasks into separate questions.
• OpenAI — Memory in ChatGPT: How memory features and reference-past-conversations settings affect new conversations, and how to turn them off.
• OpenAI — Data controls in ChatGPT: Explanation that temporary chats do not appear in history and do not create memories.
• Anthropic — Is my data used for model training?: Explanation that Incognito chats are not used for Model Improvement.

Self-check

2-1. Write down the names of the five boxes in the "five-box question table" from this chapter, and give an example from your daily life for each box.

2-2. Rewrite the question "Tell me how to cook delicious kimchi stew" more specifically using the five-box question table. Do not include your name or address in the revised question.

2-3. Write down three question-refinement tips that are better than only saying "Rewrite this" when you do not like the first answer.

2-4. When the places where the response will be used are as follows, write one line for each explaining how you should ask for the format to be specified: (a) Family group chat (b) Grocery shopping list to look at in the supermarket (c) Announcement to read out loud at the senior center

2-5. Write two cases where it is better to start a new conversation, and write one thing you must do when opening a new conversation.

2-6. Explain why it is beneficial to ask the AI to "leave blanks" for dates or store names, connecting your explanation to the content of Chapter 1.

2-7. Choose one task where you do not yet know what to ask (e.g., preparing for your first overseas trip), and write a request asking the AI to ask you questions first.

Hands-on tasks