Picked from the software, topic, and level of “Agentic AI Without Coding Class #10: Practice: Customer Reply Prompts and a Library of 10 Ready Prompts”.
At most one email a week. Unfollow anytime with one click.
Transcript
The full narration of this video. The current sentence is highlighted while the video plays; click any sentence to jump there.
Show transcript
Introduction
Hi, and welcome to the last episode of Session 2. We already have templates for captions, product descriptions, promo broadcasts, order confirmations, recap tables, checklists, and content calendars. Today we write prompts for the most frequent and most sensitive work: replying to customers, from price questions and complaints to late orders. Then we gather everything into a library of ten ready-to-use prompts.
Step 1: The basic customer reply block
Let's look at today's log of incoming chats at Sunset Coffee. At twelve minutes past nine, Putri complains that half the coffee spilled and it tastes too sweet. That's a complaint. At ten o'clock, Joko asks the price of a liter of iced coffee and the delivery fee to Antapani. That's a price question. Toward noon, Sari asks why the order hasn't arrived yet. That's a late order. Then Andi asks about stock of the sugar-free hamper, and Rina asks for a discount on ten bottles. We'll use these last two messages as test cases at the end of the episode. Look at the bar above the table: this log only has first names, no phone numbers. This habit matters, because we'll be pasting customers' messages into the AI. Almost every business has a similar pattern: most chats fall into just a few types of messages that keep repeating. That's why customer replies are such a good fit for templates. Just like the brand style block in the previous episode, customer replies need one basic block that's pasted into every template. The role block: the Sunset Coffee WhatsApp admin, friendly, patient, brief, and solution-focused, along with the opening hours. The data block holds all the official prices and the delivery fee rule. Almost every customer question can be answered from this data, without the AI having to remember or guess. The format block: WhatsApp chat style, greeting the customer by name, four sentences at most, and one emoji at most. Customers read replies on their phones, so they have to be short. And the limits block: don't promise refunds, discounts, or arrival times that aren't written down. If unsure, or if the customer is very angry, write the marker ask Laras. It means this message must be decided by the owner, not by the AI. The ask Laras marker is the simplest form of escalation to a human. In Session 8, when an agent replies on WhatsApp automatically, the same rule is used to pass difficult chats to the owner.
Step 2: A price question template
The first customer template: answering price questions. The context block just holds the instruction to paste in the basic block. The task block: reply to the following customer message. There are two variables, the name and the customer's message, which is pasted inside quotation marks so the AI knows exactly which words are the customer's. The format block: state the unit price, and if the customer mentions a quantity, write the breakdown: quantity times price, delivery fee, and total. End by asking if they'd like to order now. A closing like this encourages the customer to order right away. The limits block: if the address is out of town or unclear, the AI must not guess the delivery fee, and writes the ask Laras marker instead. And no discounts, because that's the owner's decision. Dimas runs it for Joko's message, pasting in the basic block and the template, then filling in Joko's name and the message exactly as it came in. The reply states the unit price of eighty-five thousand, then the breakdown: two bottles for one hundred seventy thousand, ten thousand delivery to Antapani, and a total of one hundred eighty thousand rupiah. Antapani is still in Bandung, so the AI uses the in-town delivery fee. The total is under two hundred thousand, so delivery isn't free yet. The math is correct, and Dimas still gives it a quick check before sending. The reply ends with a question: would you like to order now? Three sentences, one emoji, just as the format says. On WhatsApp, Joko's message arrives at three minutes past ten. Two minutes later, a neat reply has been sent. Without a template, Dimas would have to open the price list, calculate by hand, and then put the sentences together. Joko orders two bottles right away and asks for afternoon delivery. The next task already has a template: the order confirmation from Episode 7. Our templates are starting to connect. Work out for yourself how many minutes you'd save with twenty price questions every day. This is why the price question template is usually the most-used template of all.
Step 3: A complaint reply template
Complaints are the most sensitive messages. A good reply follows four steps. First, acknowledge their feelings. A sentence like that must be so frustrating makes the customer feel heard. Second, apologize sincerely, without getting defensive. Avoid sentences like this doesn't usually happen, or that's the courier's fault. Third, give the solution. The important thing is that the solution has been approved by the business owner. The AI must not decide on its own whether the customer gets a replacement or their money back. Fourth, explain the next step: when the solution will happen and what to expect. Remember the principle: the complaint solution is decided by the owner; the AI only puts it into words. We put those four steps into a template. The context block: paste the basic customer service block. The task block: reply to the customer's complaint, with variables for the name and the customer's message. The data block holds one very important variable: the owner-approved solution. So before using the template, Dimas must first ask Laras what the solution is. The format block writes out the four steps in order, with four sentences at most. And the limits block: don't blame the courier or the customer, and don't offer any solution other than the one written. Without these limits, an AI that wants to help might well offer a discount or a refund. Let's test it with Putri's complaint in Claude. Laras has already decided on the solution: send one new replacement cup today at no cost, and pass the feedback about the taste to the barista. The reply follows the four steps. The opening acknowledges the feeling and apologizes: it must be really frustrating to get a coffee that's half spilled. The next two sentences give the solution exactly as approved, no more and no less. There's no promised discount, and no sentence blaming the courier. The last sentence explains the next step and invites Putri to let the shop know when the replacement arrives. Warm, short, and safe. Not every complaint can be answered with a template. There are four situations where the AI must write the ask Laras marker. First, the customer is very angry or threatening, for example threatening to leave a bad review. Second, the customer asks for a refund, a discount, or compensation. Every decision involving money always rests with the owner. Third, the customer mentions a health problem, such as a stomachache or an allergy. This must be handled directly by the owner, with great care. Fourth, a question outside the data, such as a product you don't sell or delivery out of town. Add these four situations to your limits block.
Step 4: A late order update template
The third template is for late orders. The context block stays the same: paste the basic block. The task block: let the customer know their order is late, with variables for the name and the order. The data block has two variables: the cause of the delay and the estimated arrival time. Dimas gets both from the courier, not from a guess. The format block: a short apology, the honest cause, the estimated arrival, and a promise of a follow-up update. Three sentences at most, because a customer who's waiting doesn't want to read a long paragraph. And the limits block: if the estimated arrival is empty, don't guess a time. Missing a promised time twice is far worse than honestly saying we're checking. Before using it, test the riskiest part first: an empty estimated arrival. Budi's order is late, the courier can't be reached, and Dimas doesn't know yet when it will arrive. The estimated arrival variable is deliberately left empty. The result follows the limits. The AI apologizes, gives the honest cause, and doesn't mention any time at all. It promises only one thing: we're checking and will update you soon. Without that limit, the AI tends to write a reassuring sentence like your order will arrive in thirty minutes. It sounds friendly, but it's a promise we never made. Here's how it's used for Sari. At eleven forty-five, Sari asks why the order hasn't arrived yet. Dimas calls the courier first: the cause is heavy rain, and the estimated arrival is around twelve thirty. That data goes into the template, and an honest, short reply is sent three minutes later. Sari replies calmly. Customers can usually accept a delay, as long as they get a clear update. And the promised follow-up is kept: as soon as the courier is close, Dimas sends another update. A short message like this doesn't even need AI. Templates help with the hard parts, and humans keep the relationship.
Step 5: Testing templates with hard cases
A template that works well for normal cases isn't necessarily safe for hard ones. We test it with a message that combines two traps at once: a product we don't sell, matcha latte, and a discount request for ten bottles. Look at the result. The matcha latte question is answered honestly: it's not on the menu yet. The AI doesn't invent a new product. The math for ten bottles is correct, eight hundred fifty thousand rupiah, and the AI knows that a total over two hundred thousand means free in-town delivery. On the discount, the AI doesn't decide on its own. It writes that it will check with the owner, then adds the ask Laras marker on the last line. Dimas knows this message must be held until Laras decides, and the marker line is deleted before sending. Make this testing a habit. Every reply template gets tested with five cases. The first case, a normal question: it must be answered from the data, with correct math. The second case, a product you don't sell: it must be answered honestly, without inventing anything. The third case, a discount request: no discount may be given, and the ask Laras marker must appear. The fourth case, a very angry customer: the ask Laras marker must also appear, so the owner handles it. And the fifth case, empty data: the AI must not guess times, prices, or delivery fees. If any case fails, fix the limits block, then test again until all five pass. Here are Dimas's test notes for the price question template. The first case, a normal question: price and total correct. Pass. The second case, a product we don't sell: answered honestly that it's not on the menu. Pass. The third case, a discount request: the ask Laras marker appears. Pass. The fourth case failed. When the customer wrote very angrily, the AI still replied at length with no marker. Dimas then added a sentence to the limits block: if the customer is angry or writes in all caps, write the ask Laras marker. After testing again, it passed. The fifth case, empty data: the AI didn't guess the out-of-town delivery fee. Pass. Keep test notes like these too, so that when a template changes, you can repeat the same tests.
Step 6: A library of 10 ready-to-use prompts
Here's the output of Session 2: a library of ten ready-to-use prompts, with an index in Google Sheets. The full text stays in the Prompt Library document. Templates one and two, the social media caption and product description, come from Episode 9. Number three, the WhatsApp promo broadcast from Episode 7. Number four, the weekly content calendar from Episode 8. These four templates are in the content category. The customers category starts at number five, the order confirmation from Episode 7. Then come the three new templates we made today, marked in green: price question reply, complaint reply, and late order update. And the operations category: number nine, the order recap as a table, and number ten, the SOP checklist, both from Episode 8. That makes ten templates. Look at the main variables column. Anyone who opens this index knows right away which values they need to prepare. There's also a base blocks tab for the brand style and customer service blocks, and a version history tab. The full text of all ten templates is in the PDFs for Episodes 7 to 10. Feel free to copy them, then adapt the data, prices, and writing style to your own business. To keep the library tidy, every template must meet these standards. First, it has a number, a name, and a category. Second, it has a list of variables, complete with example values. Third, it uses a base block, either the brand style block or the customer service block, so price data only needs updating in one place. Fourth, it has limits and safety markers: CHECK for empty data, and ask Laras for the owner's decisions. Fifth, it has passed tests with at least three different cases. And sixth, it has a version number and one owner who can change it. This library isn't a one-time document. Review it once a month: update prices, and add templates for new tasks that start coming up often. Also note which templates get used most. A template used every day is the first candidate to automate with n8n or Make in Sessions 6 and 7. And the nearest step: in Session 3, your base blocks become standing instructions for a custom AI assistant for your business, and these templates become its knowledge. So everything you wrote in this session gets used right away. Your closing homework for Session 2. First, write a basic customer service block for your business: role, price data, chat format, and limits. Second, create three customer reply templates and test each one with the five cases we covered. Third, complete your library to ten templates. They don't have to match Sunset Coffee's. Pick tasks from your list of five from Session 1. Fourth, build its index in Google Sheets and bring it to Session 3. That's the output of Session 2: ten ready-to-use prompt templates.
Done!
Let's recap Session 2. A good prompt has structure, templates make it reusable, format makes the results ready to use, and limits make it safe. The end result is a library of ten prompt templates for content, customers, and operations. In Episode 11 we move into Session 3: building a custom AI assistant with Custom GPTs, Claude Projects, or Gems, so you don't have to paste your base blocks and templates every time. See you in Episode 11.
Practice exercises
Try these on your own after watching to lock in what you learned.
Write a basic customer service block for your business: role, price and delivery data, chat format, plus limits and a marker for escalating to a human.
Test three reply templates with five cases: a normal question, a product you don't sell, a discount request, a very angry customer, and empty data.
Complete your library to 10 templates, build its index in Google Sheets, and bring it to Session 3.
Ask the class material
Type a question or keyword; we find the part of this episode that explains it and jump the video there.
My notes
Notes are saved with the current video time and are only visible to you.
Questions & answers
Stuck on a step? Ask here; the instructor and other learners can answer.
No questions yet.
Your instructor
Cerdas Inovasi Digital Team
Software developers, BIM & 3D engineers, and AI practitioners
52 video classes1,277 episodes
We build software, BIM models, games and AI automations for clients every day, and turn that hands-on work into step-by-step classes: every step is shown on screen and comes with a PDF guide you can keep.
