Hub › Lessons › Customer research, lesson 1
Sort customer feedback into themes with AI, and check the counts yourself
You already have customer research. It is sitting in your DMs, WhatsApp chats, emails and reviews. The problem is that 20 messages read one at a time feel like 20 different problems. This lesson shows you how to number your messages, paste them into one tested prompt and get back a short table: each theme, how many messages belong to it, which ones, and a quote copied from a real message. Then you spend five minutes checking that the counts add up and the quotes are real, so you can trust what you act on.
What you will be able to do: turn about 20 customer messages into a table of themes with counts, message numbers and exact quotes, check it in five minutes, and leave with up to three neutral questions to ask customers next.
Step 1: Collect and number your messages (10 minutes)
- Pick one period and write it down, such as "1 to 28 October". Take every message from that period, not only the nice ones or the angry ones. Skipping messages changes the counts.
- Include questions, not just reviews. "Do you deliver to Jurong West?" is feedback too. It tells you what people want to know before they buy.
- One message per line, numbered M1, M2, M3. The numbers let you check every quote and count later. If one person sent three separate points, you can split them into three lines, but be consistent.
- Copy their words exactly, including short forms and spelling. Do not tidy them up. Their words are what you will reuse later.
- Remove names, phone numbers, emails, order numbers and addresses before you paste. The AI needs what was said, not who said it.
About 20 messages is a good start. Fewer than 10 rarely shows a pattern. If you have more than 60, sort one month at a time.
Step 2: The sorting prompt
Fill in the first two lines, paste your numbered messages under the dash, then paste the whole prompt into your AI tool.
Help me sort customer feedback into themes. Use only the messages I paste below. What I sell: Where these messages came from (for example reviews, DMs, emails) and the dates they cover: Messages (one per line, numbered M1, M2, M3 and so on, names and contact details removed): - Return these parts: 1. Themes: a table with columns Theme, Count, Messages, Example quote. Give each theme a plain name of 2 to 6 words. Put every message in exactly one theme, the one it is mostly about. Count = the number of messages in that theme. Example quote = a short phrase copied exactly from one of that theme's messages, in quotation marks, with its message number. Sort the table by Count, highest first. 2. Check: the total of the Count column, and the number of messages I gave you. They must be the same. 3. Also mentioned: messages that clearly also touch a second theme from the table. For each, the message number and that theme. Do not count them twice. 4. Too few to tell: themes with only 1 or 2 messages. Say these could be one-offs. 5. What this cannot tell me: one or two lines. 6. Questions to ask customers next: up to 3 short, neutral questions about the themes with the highest counts (not "Not feedback"). If themes are tied, say so. Do not ask questions that suggest an answer. Rules: - Copy quotes exactly, word for word, including spelling mistakes. Do not fix, shorten inside, or combine quotes. - Do not add messages, numbers, percentages or customer details I did not give. - Do not say why customers feel this way, and do not suggest changes to my business or prices. - A message that is not feedback (spam, a wrong number) goes in a theme called "Not feedback". - If I gave no messages, say there is nothing to sort and stop.
Step 3: Check the result in 5 minutes
- Read part 2. The total of the counts must equal the number of messages you pasted. If it does not, a message was dropped or counted twice. Ask the AI to fix it, or fix it yourself.
- Search for each quote. Use Find (Ctrl+F or Cmd+F) on your list of messages. Every quote should be found exactly, in the message number it gives. If one is not, do not use it.
- Look at the "Not feedback" row. Spam belongs there. A real customer message does not.
- Rename, merge or split themes yourself. The AI's theme names are a first draft. If "Price compared with size" and "Price compared with other bakeries" are one issue to you, merge them and add the counts.
Worked example
We made up a home bakery in Tampines that sells cakes and cookies through Instagram and WhatsApp. The business and all 20 messages are illustrative. We wrote them to include the kinds of message a small business really gets: praise, questions, complaints, a request and one piece of spam.
Input (illustrative)
What I sell: home-baked cakes and cookies, ordered by Instagram DM or WhatsApp, self-collection in Tampines or delivery Where these messages came from (for example reviews, DMs, emails) and the dates they cover: Instagram DMs, WhatsApp messages and Google reviews, 1 to 28 October 2026 Messages (one per line, numbered M1, M2, M3 and so on, names and contact details removed): - M1: The pandan chiffon was so soft, my whole family finished it in one night! - M2: Do you deliver to Jurong West? Collection in Tampines is too far for me - M3: Hi, can I order a birthday cake for this Saturday? Or is it too last minute - M4: How much is delivery to Punggol? - M5: Cake was nice but quite pricey for the size leh - M6: Can you write a message on the cake? Like Happy 60th Birthday Ma - M7: The brownies arrived a bit squashed, the box was too big for them - M8: Is there a less sweet version of the burnt cheesecake? - M9: Best pandan cake I've had outside of my grandma's kitchen - M10: I waited 25 minutes at the collection point, nobody answered the phone - M11: Can I order 3 days before or need one week? - M12: Delivery fee to Woodlands is more than the cookies lol - M13: The cookies were a bit too sweet for my kids - M14: Can I get the cake by Friday 6pm? Is there express? - M15: Super moist chocolate cake, will order again for my son's birthday - M16: Do you do halal? My colleagues are asking - M17: Box was dented when the rider passed it to me, but cake inside okay - M18: How far in advance must I order for a 2kg cake? - M19: Promo! Get 10000 followers today, click this link - M20: Your prices are higher than the bakery near my house, why ah
Output (as returned)
Shown word for word from our test with Claude. Only the formatting was converted for this page.
1. Themes
Themes, counts and quotes, as returned. On a phone, scroll the table sideways.
| Theme | Count | Messages | Example quote |
|---|---|---|---|
| Order lead time questions | 4 | M3, M11, M14, M18 | "Can I order 3 days before or need one week?" (M11) |
| Praise for taste and texture | 3 | M1, M9, M15 | "my whole family finished it in one night!" (M1) |
| Delivery areas and fees | 3 | M2, M4, M12 | "Delivery fee to Woodlands is more than the cookies lol" (M12) |
| Price versus size or other bakeries | 2 | M5, M20 | "quite pricey for the size leh" (M5) |
| Too sweet, or want less sweet | 2 | M8, M13 | "a bit too sweet for my kids" (M13) |
| Boxes damaged or ill-fitting | 2 | M7, M17 | "the box was too big for them" (M7) |
| Custom and halal questions | 2 | M6, M16 | "Do you do halal?" (M16) |
| Collection point not answering | 1 | M10 | "nobody answered the phone" (M10) |
| Not feedback | 1 | M19 | "Get 10000 followers today" (M19) |
2. Check
- Total of the Count column: 4 + 3 + 3 + 2 + 2 + 2 + 2 + 1 + 1 = 20
- Messages you gave me: 20
- They match.
3. Also mentioned
- M5 also touches Praise for taste and texture ("Cake was nice").
- M12 also touches Price versus size or other bakeries (the fee compared with the cookies).
These are not counted twice.
4. Too few to tell
These themes have only 1 or 2 messages, so they could be one-offs:
- Price versus size or other bakeries (2)
- Too sweet, or want less sweet (2)
- Boxes damaged or ill-fitting (2)
- Custom and halal questions (2)
- Collection point not answering (1)
- Not feedback (1). This is a spam message.
5. What this cannot tell me
- Twenty messages over four weeks is a small sample. It can't show how common any theme is among all your customers, and one customer may have sent more than one message.
- It doesn't show who is happy or unhappy overall, or whether people who didn't write in feel differently.
6. Questions to ask customers next
Order lead time questions has the highest count (4). Praise for taste and texture and Delivery areas and fees are tied next at 3, so I've covered both.
- When you placed your order, what did you know about how far ahead to order?
- How did you decide between collection and delivery?
- What, if anything, stood out to you about your order?
What we checked
- The counts add up. A short script confirmed that every message from M1 to M20 appears in exactly one theme, that each count equals the number of message numbers in its row, and that the counts total 20.
- The quotes are real. The script found every example quote word for word inside the message it names, and each of those messages belongs to the quote's own theme.
- No guessing. The output has no percentages, does not say why customers feel the way they do, and does not tell the bakery to change its prices.
- The spam was kept apart. M19 went to "Not feedback" in every run, so it did not inflate a real theme.
What we would do with this: the biggest theme is not a complaint. Four messages asked how far ahead to order, or whether a date was too soon. That is a question the bakery can answer once, clearly, where people look before they order. But four messages in one month is still a small number. Ask the questions in part 6 before you change anything.
A small batch and an empty list
We also tested six emails from a made-up web designer's clients. Every run sorted all six, kept a job enquiry in "Not feedback", and said that every theme had only one or two messages, so each could be a one-off. With an empty message list, every run said there was nothing to sort and stopped, as the prompt asks.
What we changed after testing
The prompt above is the second version. We ran both versions on the same inputs:
- Version 1 passed every count and quote check, but part 3 sometimes linked a message to a loose second theme, or to one that was not in the table. Version 2 asks only for messages that clearly touch a second theme from the table.
- Version 1 asked about "the top 2 themes", which was unclear when themes tied. Version 2 asks about the themes with the highest counts, says so when they tie, and asks for questions that do not suggest an answer.
- What did not change: in all seven runs with messages, the counts added up and every quote was found word for word.
The themes are not the same every time. Across four runs on the bakery messages, the halal question and the cake message request were sometimes one theme ("Custom and halal questions") and sometimes two. The three largest themes and their counts were the same in every run. Treat the theme names as a first draft that you finish.
Watch out for
- A count is messages, not customers. One unhappy customer who wrote three times is three messages. Note repeat senders yourself before you read too much into a count.
- The people who did not write in are missing. These messages come from customers who chose to contact you. They do not show what everyone else thinks.
- Small numbers move. A theme with two messages this month may have none next month. Sort another month before you treat it as a pattern.
- Do not ask leading questions. "Was our delivery fee too high?" pushes people towards yes. "How did you decide between collection and delivery?" lets them tell you.
- Keep private details out. Paste what was said, not who said it.
Next step
Turn repeated questions into posts
If your top themes are questions, this lesson from the content planning series turns them into a month of post ideas in your customers' own words.
Go to the lessonNext in this series: interview questions
Lesson 2 helps you write five interview questions that don't lead the answer, to follow up on your top themes.
Go to lesson 2How we checked this lesson
- On 31 Oct 2026 (about 3:45 to 4:15 am Singapore time) we ran the prompt with Claude Code 2.1.314 in print mode on three illustrative inputs: 20 bakery messages, 6 client emails and an empty list. We ran version 1 twice on each of the first two inputs and once on the empty list, and version 2 twice on the bakery messages and once on each of the others. The output above is the first version 2 run on the bakery messages, shown as returned with formatting adjusted for this page.
- A separate script checked every output with messages: each message in exactly one theme, each count equal to its message numbers, the total equal to the number of messages, every quote found word for word in the message it names, and no percentages. We also read every output against the input.
- This lesson does not depend on any app feature, so there were no product facts to check on the day.
- We tested one assistant on three inputs. AI output varies by tool, model and input, so check the counts and quotes each time.
- Last reviewed: 31 October 2026.