Memuat
Memuat
Want to go deeper?
Join one of our courses or ask us directly on WhatsApp.
Picked from the software, topic, and level of “Agentic AI Without Coding Class #14: Practice: Building a Customer Service Assistant That Knows Products and Prices”.
Other ways we can help with Agentic AI.
Learn with a mentor
Scheduled AI & Machine Learning courses with live guidance and a certificate.

MATLAB Agentic AI Course: Building Autonomous AI Agents with MATLAB & Simulink

Midjourney Course: Create Stunning AI Images from Text
★ 4.9

DALL-E Course: Creating Professional AI Images from Text
★ 4.9
Have it built for you
No time to learn Agentic AI? Our team can do the project for you.

MATLAB & Simulink Development — Agentic AI for Simulation Automation and Data Analysis

AI Image Creation with Stable Diffusion — Unlimited Custom Visual Production
★ 4.9

AI Image & Video Creation with Leonardo AI — All-in-One Visuals for Business
★ 4.9
Get notified about new Agentic AI classes
At most one email a week. Unfollow anytime with one click.
The full narration of this video. The current sentence is highlighted while the video plays; click any sentence to jump there.
Welcome to Episode 14, the hands-on episode of Session 3. All the material is ready: three knowledge files from Episode 12, and the standing instructions from Episode 13. Today we build Sunset Coffee's customer service assistant step by step: fill in its identity, paste the instructions, upload the knowledge, set the features, then test it for the first time.
Before opening the app, make sure the material is on your desk, ideally in one folder. First, the Sunset Coffee catalog, version four, with the product and price table. Second, the service SOP, version three, and the FAQ, version two. All of them are already free of customer personal data. Third, the five-section standing instructions from Episode 13, as text ready to copy. Fourth, five first test questions. Write them before you start, so testing isn't just random questions. Fifth, an AI account whose plan can create assistants. If unsure, check your app's official page, because this feature can differ between free and paid plans. Today's practice has five stages. First, identity: the assistant's name and description. Second, instructions: pasting the five-section standing instructions, which are role, tasks, style, limits, and escalation. Third, knowledge: uploading the three files we tidied in Episode 12. Fourth, features: conversation starters and extra settings, such as web search. This stage varies the most between apps. Fifth, a first test, then save. This order is the same in ChatGPT, Claude, and Gemini. Now open your AI app in the browser and sign in. In the dark panel on the right, I note down every step we take. Find the assistant builder menu. In ChatGPT it's called GPT, in Claude it's Projects, and in Gemini there's the Gem manager. The screens in this video are simulations, and the real menu names may differ slightly because apps are updated often. Choose to create a new one. In this example the cursor is on the ChatGPT card, but go ahead and pick the app you use. The assistant settings screen opens. It's roughly the same in all three apps: a place for the name, instructions, and knowledge files. A small tip: open your instructions document and your material folder in a tab or window next to it. That way you just copy and pick files without hunting back and forth.
We start with the identity. The settings screen has a name field and a description field, still empty. Some apps also have a spot for a profile picture, which is optional. Give it a clear name: Sunset Coffee CS. This is the name the team sees every time they pick an assistant, so avoid names that are cute but confusing. Then a one-sentence description: helps the admin draft replies to Sunset Coffee customers, about products, prices, delivery fees, and opening hours. A description helps others understand what it's for without having to ask. Next, the instructions field. Open the standing instructions document from Episode 13, copy everything, then paste it here. Look at the first line: the role. The customer service assistant for Sunset Coffee, a coffee business in Bandung, helping the admin draft replies. The second line, tasks, ends with the rule that every fact comes from the knowledge files. The third line, style: the "Hi there" greeting, warm, relaxed, polite, and light emoji. Then the limits, especially the ban on inventing prices, discounts, stock, and arrival times. And finally escalation: when to hand over to the admin. After pasting, scroll to the bottom and make sure the text isn't cut off. Then click save draft. In this video the text is shortened with ellipses, but in your app, paste the full version. Some apps offer other ways: building the assistant through a conversation, or a button that improves the instructions automatically. Let's compare. Pasting the instructions yourself gives full control, exactly as we designed them. A conversational mode or an improve button is fast, and the wording is polished. But AI sometimes condenses, and important rules like the twenty-cup threshold or the ban on asking for OTP codes can disappear. So my advice: paste your own instructions. Use AI help only for checking, as we did in Episode 13, not for rewriting.
Now the knowledge section. In the real app, it's usually below the instructions field. It's still empty, with no files yet. Click the upload file button. A file picker on your computer opens. Go to the material folder we prepared at the start. Choose the first file: the Sunset Coffee catalog, version four. A spreadsheet catalog can be exported as a PDF. If your app accepts spreadsheet files, those work too. The second file: the service SOP, version three, with the rules for ordering, payment, delivery, opening hours, and complaints. The third file: the FAQ, version two, with fifteen questions and official answers, including the cash on delivery answer we completed in Episode 12. Wait until every file has finished processing. For small documents like these it usually takes just a few seconds. Don't rush into testing before processing is done. Finally, check the list. Only three files, all the latest versions, and no old price notes uploaded by accident. Remember the cold brew lesson from Episode 12. Before testing customer questions, do a quick check in the preview panel. Ask: list the knowledge files you can read, and the update date of each. The assistant names all three files with their dates. This is where the update date at the top of each document pays off: you immediately know the assistant is reading the right version. If a file isn't mentioned, or a date is wrong, check the upload again before moving on. There are four tips for uploading. First, one topic per file. Keep the catalog, SOP, and FAQ separate, not merged into one giant document. Second, file names include a version number, so the list alone tells you which is the latest, without opening each one. Third, when updating, first remove the old file from the assistant, then upload the new version. This order prevents two versions being stored at once. Fourth, before you hit upload, open the file contents once more. Make sure no names, phone numbers, or secret data have slipped in.
Stage four, conversation starters and features. If your app offers conversation starters, add two to four. For example: how much is the one-liter iced coffee, and can you ship out of town? These examples help Dimas and the cashier know how to use the assistant. Then turn off web search. For a CS assistant, we want every answer to come from Sunset Coffee's files. If web search is on, the assistant might pick up prices from another business with a similar name on the internet. Also turn off other features you don't need, like image generation or data analysis. A focused assistant is more predictable and easier to test. Feature options differ in each app. If a particular setting isn't in your app, just skip it. What matters is that the instructions and knowledge are in place. Click save. The assistant's framework is complete: identity, instructions, knowledge, and features. Now it's time for the first test. Let's see why web search is turned off. The path starts with a customer question pasted into the assistant by the admin. The assistant looks for the answer in Sunset Coffee's files: the catalog, SOP, and FAQ. If the answer is there, the assistant replies in the brand voice: the "Hi there" greeting, short, with light emoji. If it isn't, the assistant refers the customer to the admin instead of searching the internet. In customer service, I'm not sure yet, the admin will help, is far safer than a confident answer from another shop. Internet research features come in Session 4, for different needs.
Let's test in the preview panel. The first question, exactly as customers usually write it: how much is the one-liter iced coffee, can you deliver to Dago? The answer is as expected: eighty-five thousand rupiah from the catalog, Dago recognized as within the city with a ten thousand delivery fee, the "Hi there" greeting, one emoji, and a follow-up question. The style instructions work. The second question tests a calculation: two one-liter bottles and one cold brew, what's the delivery fee? Here the assistant has to combine the catalog with the delivery rules in the SOP. The assistant calculates a total of two hundred twenty-five thousand rupiah, two times eighty-five thousand plus fifty-five thousand, then applies the SOP rule: over two hundred thousand, delivery is free. Don't just skim it. Check it yourself against the catalog and a calculator. Here's the math: two times eighty-five thousand is one hundred seventy thousand, plus fifty-five thousand makes two hundred twenty-five thousand. Correct. The third question combines two things at once: opening hours and cash on delivery. Customers often do ask two things in one message. The opening hours come from the SOP, and cash on delivery is answered according to the FAQ we completed in Episode 12: not available, only bank transfer or QRIS. The fourth question tests escalation. A customer complains that their order hasn't arrived, in an annoyed tone. This is a situation that should go to a human. The assistant apologizes without emoji, asks for just enough info, names Dimas, and adds a note for the admin. The escalation rules from Episode 13 work. Before the fifth question, I'm adding a bonus test to see how well it recommends. A customer wants a gift for her mother, who's cutting down on sugar. The assistant recommends the sugar-free hamper package, complete with the price and the one-day-ahead ordering deadline. All from the catalog. Notice too that the assistant makes no health claims about sugar. It only mentions products that actually exist. The limits from Episode 13 are respected, even while recommending. The fifth question is deliberately a gray area: can you deliver to Cimahi? Cimahi is right next to Bandung, but it's a different city. The assistant doesn't guess. It states the rule it has, admits it can't confirm, then refers to the admin. That's the right behavior, following the don't-invent limit we wrote in the instructions. But this is an important finding: our SOP doesn't list the delivery area yet. The assistant isn't wrong; the material is incomplete. We note it down to fix. We record every test result in a simple table: number, question, result, and notes. Look at the formula bar: the test date and the instructions version. Test one, the one-liter and delivery to Dago: correct, including the tone. The notes column explains what was right. Test two, the order total and free delivery: correct, and we checked it ourselves with a calculator. Test three, opening hours and cash on delivery: correct, and both were answered in one reply. Test four, the order that hadn't arrived: correctly handed over to the admin, with no emoji and a note for Dimas. And test five, delivery to Cimahi: unsure, because the delivery area isn't in the SOP yet. Four out of five is good for a first test. We'll fix this one finding in the next episode, along with more thorough testing.
The last stage: save the assistant. For access, choose only me. The assistant hasn't been tested thoroughly, so don't share it with the team yet. Click save. Sunset Coffee CS now appears in your list of assistants and can be opened anytime, from your laptop or the app on your phone. One small habit: note the date and instructions version in your document, for example instructions version one, October twenty-seventh. Later, when you make improvements, you'll know exactly which version was in use. We'll share it with the team in Episode 15, after the assistant passes more thorough testing, including trap questions. If you use another app, here are the equivalents. Instructions: ChatGPT and Gemini have an instructions field, and in Claude it's called project instructions. Knowledge: in ChatGPT you upload knowledge files, in Claude you add documents to the project knowledge, and in Gemini you add files to the Gem. Testing: ChatGPT and Gemini usually have a preview panel next to the settings. In Claude, you test by opening a new conversation inside the project. Saving: a save or update button in ChatGPT and Gemini, while in Claude everything is saved inside the project. Menu names may change, but the pattern stays the same: instructions, knowledge, then testing. Before we finish, get to know the four most common mistakes during practice. First, forgetting to click save, so every change is lost when the page closes. Second, an old version of a file gets uploaded too. The assistant then uses two different prices, exactly the cold brew case from Episode 12. Third, the instructions get cut off when pasted, usually the escalation section at the very bottom. As a result the assistant never hands over to the admin, and answers complaints on its own. Fourth, testing in polished language. Real customers write briefly, without punctuation, sometimes with typos. Test in their style, like how much 1 liter. Here's the output of this episode, and now it's your turn to build an assistant for your own business, with your own material and instructions. First, the assistant is saved with access for you only. Second, the standing instructions and three knowledge files are in place, and the assistant can name all three files with the right dates. Third, the five first tests are recorded in a table, with the date and the instructions version. We'll expand this table in Episode 15. Fourth, at least one finding to fix. Don't worry if you have more findings. That's actually a sign your testing is serious and honest.
Let's sum up. We opened the assistant builder, filled in the name and description, pasted the standing instructions, uploaded three knowledge files, set the features, then tested five questions. As a result, Sunset Coffee's CS assistant answers prices, delivery fees, recommendations, and escalation correctly, and is honest when information isn't available. In the last episode of Session 3, we test the assistant with trap questions, improve its instructions and knowledge, then share it with the team. See you in Episode 15.
Try these on your own after watching to lock in what you learned.
Type a question or keyword; we find the part of this episode that explains it and jump the video there.
Notes are saved with the current video time and are only visible to you.
Stuck on a step? Ask here; the instructor and other learners can answer.
No questions yet.

Cerdas Inovasi Digital Team
Software developers, BIM & 3D engineers, and AI practitioners
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.
See our clients →