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
Welcome to Episode 13. Sunset Coffee's knowledge material is ready: the SOP, catalog, and FAQ. Now we write the second part, which is just as important: the standing instructions. We'll set the assistant's role, a brand voice with the "Hi there" greeting and light emoji, topic limits, and rules for when the assistant must hand a customer over to a human admin.
Step 1: Standing instructions: an SOP for AI
Let's start by separating two things. Knowledge is what the assistant knows: prices, delivery rules, opening hours. Instructions are how the assistant behaves. Picture a new admin at your shop. The catalog tells them product prices. But how to greet customers, what must never be promised, and when to call the owner are taught through a work SOP. Standing instructions are a work SOP for AI. Good standing instructions have five sections. First, the role: who this assistant is and which business it serves. Second, the tasks: what it does, and where it gets its facts. For a CS assistant, the facts always come from the knowledge files. Third, the style: how your brand speaks to customers, from greetings to emoji use. Fourth, the limits: what it must not do, answer, or promise to customers. Fifth, escalation: when the assistant must stop and hand the conversation to a human. We'll write all five, one by one, for Sunset Coffee. Before writing, let's practice sorting. Where does the cold brew price of fifty-five thousand rupiah go? In the knowledge, specifically the catalog. Prices can change, and then we only edit one file. Greet customers with Hi there? That's an instruction, because it's a behavior rule that applies to every answer. Free delivery over two hundred thousand rupiah? That's a business rule that might change someday, so it belongs in the SOP, part of the knowledge. Never invent prices? That's an instruction, because it governs the AI's own behavior. The rule of thumb: if it can change because of a business decision, put it in the knowledge. If it governs how to behave, put it in the instructions. Before writing, get to know the three most common mistakes. First, vague instructions, like answer well and kindly. Well by whose standard? Without examples, every AI interprets it differently. Second, mixing data into the instructions. Prices and promos get written into the instructions, then nobody updates them when the promo ends. As a result, the assistant keeps offering a promo that's already over. Third, too long. Two pages of rambling instructions bury the important rules. Short, firm, and organized by section is better, just like what we'll write now.
Step 2: From Session 2 templates to role and tasks
Here's the first part of Sunset Coffee's standing instructions. The role section: you are the customer service assistant for Sunset Coffee, a coffee business in Bandung, helping the admin draft chat replies. Notice the role is clearly to help the admin, matching the assistant's position in Session 3. The task section names the topics it handles: products, prices, delivery fees, opening hours, how to order, and payment. Its last sentence is key: take every fact from the knowledge files. The data section names all three reference files, and states that for prices, the catalog wins. This rule prevents confusion if two different numbers ever appear. The format section asks for a reply ready to send on WhatsApp: at most three short sentences, no headings, and no tables. Tables look neat on a computer screen, but messy in a customer's WhatsApp chat. Compare it with the Session 2 template. The business name variable, once filled in every time, is now written just once: Sunset Coffee, Bandung. The catalog variable, once pasted every time, is now simply referenced from the knowledge files. The customer message variable is still there, but no longer inside the instructions. Dimas pastes it as a regular chat message whenever a question comes in. The format and limits sections move over whole, then get sharpened. That's what we do in the next steps: tone, limits, and escalation.
Step 3: The Sunset Coffee brand voice
Now the part that makes the assistant sound like Sunset Coffee: its voice. Rule one, greet customers with Hi there. It's neutral for every customer, warm, and still polite. Second, light emoji: at most one or two per message. And no emoji at all when a customer is complaining, because a cheerful emoji in the middle of a complaint feels insensitive. Third, short, like a chat. Two to three sentences, one idea per sentence. Customers read on a phone screen, often while doing something else. Fourth, relaxed but not sloppy. Avoid slang shorthand, and don't write in all capital letters. Fifth, no overpromising. Avoid words like definitely, guaranteed, or claims like the best in Bandung. A promise you can't keep damages trust faster than an ordinary answer. See the difference for the same question: how much is the one liter? The first style is too stiff. Dear customer, the price of palm sugar iced coffee, one liter, is eighty-five thousand rupiah. Correct, but it reads like an official letter. The second style is too hyped: all capitals, a barrage of exclamation marks, too many emoji, and a claim of being the best. It grabs attention, but it isn't convincing. The third style is on brand: the "Hi there" greeting, a clear price, one emoji, and a follow-up question that invites the customer to order. This is what we write into the instructions. Here's the second part of the standing instructions. The style block sums up the greeting, the tone, and the ban on sloppy slang in two sentences. The emoji block sets a limit of one or two emoji per message, gives examples that fit the brand, and forbids emoji during complaints. The example block holds a reply structure, not a full example with prices. Hi there, then the answer from the catalog or SOP, one emoji, and a follow-up question. It deliberately has no numbers, so the instructions don't go stale when prices change. Here's the result on WhatsApp. A customer asks briefly, even in shorthand: got the sugar-free hamper? Dimas's reply, drafted by the assistant, stays neat: the "Hi there" greeting, the price from the catalog, one emoji, the one-day-ahead reminder, and then a follow-up question. The customer asks whether the thirtieth works. That date is well beyond the one-day-ahead limit, so it's fine. The next reply is short, has one emoji, and goes straight to asking for the address and the recipient's name. Notice the assistant doesn't copy the customer's shorthand. The brand voice stays intact.
Step 4: Topic limits: what to answer and what not
The next section is topic limits. The first thing it may answer: products, sizes, and prices, always from the catalog. Second, delivery fees, the delivery area, opening hours, and how to order and pay, from the SOP. Third, product recommendations. For example, if a customer asks for an office gift idea, the assistant may suggest a hamper, as long as the product is in the catalog. Now what it must not do. First, discounts or prices outside the catalog. The assistant has no authority to give price cuts. Second, health claims, for example a question about whether sugar-free coffee is safe for people with diabetes. That's a doctor's territory, not a coffee shop admin's. Third, topics unrelated to the business, like politics, gossip, or helping with school homework. A CS assistant isn't an all-purpose chatbot. Fourth, promises that aren't certain: an exact arrival time, or stock for tomorrow. Things like that are answered by an admin who knows the real situation. Let's test one limit. A customer asks: the sugar-free coffee is safe for diabetics, right? A reasonable question, but risky to answer carelessly. The assistant doesn't say yes or no. It politely suggests consulting a doctor. Then it still helps with facts from the catalog: cold brew with no sugar and no milk, and a sugar-free hamper. Declining politely while offering relevant help. That's exactly the pattern we want. Here's the third part of the standing instructions. The first limit: only discuss Sunset Coffee, and for other topics decline politely while offering help with products and orders. The second limit is the most important: never invent prices, discounts, stock, or arrival times. If it's not in the files, say you don't know yet, then refer to the admin. This sentence keeps the assistant honest. The third limit: never give health advice, and never ask for card details, PINs, or OTP codes. Scams often use a shop's name, so customers need to know Sunset Coffee never asks for data like that.
Step 5: When to hand over to a human admin
The last section, and an often forgotten one, is escalation: when the assistant must hand the conversation to a human. The first situation: complaints. A damaged product, a wrong item, a late delivery, or an unhappy customer. Second, requests for a refund or to cancel an order that's already been paid. This involves money, so the decision belongs to a human. Third, large or special orders. Laras sets the threshold: more than twenty cups, office events, or delivery out of town. Fourth, questions whose answers aren't found in any file. Better to hand it over than to guess. Fifth, the customer asks to talk to a person. Respect that request without trying to keep the customer with the assistant. Customers who feel respected tend to come back. Even handing over has a method. First, acknowledge the customer's feelings with brief empathy, without getting defensive or blaming anyone. Second, collect just enough information so the admin doesn't have to ask again. For a complaint, a photo of the product and the order time are enough. Third, hand over clearly. Say the admin will follow up directly, by name if needed, so the customer knows a real person is responsible. Fourth, give an honest time estimate based on opening hours. Don't promise a reply within five minutes if the shop is already closed. Let's see the result. A customer writes that the one-liter coffee they received yesterday tastes sour, and they're disappointed. The assistant opens with a sincere apology and no emoji, following the style rule for complaints. Then it asks for a photo of the product and the order time, names Dimas as the person who will follow up, and gives a deadline within opening hours. Notice the last line: a note for the admin. This line isn't sent to the customer. Its purpose is to flag to Dimas that he must handle this case himself. In Session 8, a flag like this can trigger an automatic notification. Here's the fourth part of the standing instructions. The escalation block sums up all five situations in one clear sentence, including the threshold: more than twenty cups. The how block holds the four handover steps, plus an order to add a line for the admin at the end. Clear rules like this keep the assistant consistent, whoever uses it. Here's the handover on WhatsApp. The complaint comes in at five past nine in the morning. Dimas sends the assistant's draft: an apology, then a request for a photo and the order time. The admin note line, of course, isn't sent. The customer sends a photo and the order time. That's enough for Dimas to act without asking again. Then Dimas takes over himself. He introduces himself and decides the fix: sending a replacement today at no charge. Decisions like this are deliberately not left to the assistant, because they involve money and customer trust.
Step 6: Assembling, checking, and saving the instructions
Once the four parts are assembled into one document, check it with this checklist. First, complete: it has a role, tasks, style, limits, and escalation. Second, no prices or data that can change. If you find a price or a promo name in the instructions, move it to the catalog or SOP. Third, the instructions name the knowledge files to refer to, so the assistant knows where to look for answers. Fourth, short command sentences, one rule per line. Long paragraphs bury important rules, and AI can easily miss them. Fifth, include an example answer structure. Examples are easier for AI to imitate than long explanations, exactly what we learned about the example section in Session 2. Sixth, no rules that contradict each other. For example, one line asks for short answers, but another asks it to explain every product variant in detail. AI can also help check your instructions. Laras pastes hers into Claude and asks for three things: find rules that conflict, rules that are ambiguous, and common situations not yet covered. Notice her request: don't rewrite yet, just list your findings. Finding one: the at-most-three-sentences rule conflicts with the request to explain every variant. Laras deletes the second one. Finding two: there's no rule yet for chats outside opening hours. Laras adds one: say the order will be processed tomorrow from eight in the morning. Finding three: the term large order isn't consistent. Laras aligns it: more than twenty cups. By asking for a list of findings, the decisions stay in Laras's hands instead of simply being handed to the AI. One important habit: instructions need maintenance too. Don't change many rules at once. Save the old version first, change one thing, then test again. That way you know which change actually improved the answers. How to test systematically is covered in Episode 15. Here's the output of this episode. Now it's your turn to write for your own business. First, standing instructions with five sections, in one text document. Second, your brand style guide, with an example answer structure. Adapt the greeting: your business might suit Sir and Madam, first names, or another form of address. Third, a list of topic limits and escalation rules, including which admin handles them in your business. Fourth, make sure everything has been checked against the checklist. Save it as text, because in Episode 14 we paste it into the assistant builder.
Done!
Let's sum up. Standing instructions are a work SOP for AI, with five sections: role, tasks, style, limits, and escalation. Prices stay in the catalog, the brand voice is locked in with the "Hi there" greeting and light emoji, and every sensitive case goes to a human admin. In the next episode we get hands-on: building Sunset Coffee's customer service assistant step by step, from filling in the name and instructions and uploading the knowledge, all the way to the first test. See you in Episode 14.
Practice exercises
Try these on your own after watching to lock in what you learned.
Write standing instructions for your business assistant with five sections: role, tasks, style, limits, and escalation.
Create a short brand style guide: greeting, emoji limit, reply length, and three words or claims to avoid.
Define at least five situations in your business that must go to a human, including who handles them.
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
53 video classes1,291 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.
