01 AI doesn't know who you are

When you open a AI chat for the first time, the screen is just a writing box. On the other side, there’s no one who knows your company, your client, or the way you speak. AI has read a huge amount of text from all over the world, but it hasn’t read anything about you. So it responds to a generic person who doesn’t exist.

This explains the first disappointment almost every beginner has. You type "message for patient" and get a polite, long, generic text that could work for any clinic in the country. It’s not that the AI is bad. You asked for something meant for anyone, and it delivered something meant for anyone. The first part of a good prompt is saying who you are and who the response is for.

Beatriz, the owner of a physical therapy clinic with six professionals, did exactly that on her first try. She asked for "a message to remind someone about an appointment" and got four paragraphs that started with "Dear patient." On her second try, she wrote first: "I own a neighborhood physical therapy clinic, I talk with patients on WhatsApp in a warm tone, and most of them are over 50." The message came back in three lines with the right tone. It was the same AI. What changed was what it knew about Beatriz.

It doesn’t need to be long. One sentence is usually enough: who you are, what you do, and who you’re talking to. It’s the difference between asking for bread at the counter and leaving an order slip with a name, quantity, and time.

02 Say the result, not the steps

The second part is the most important, and the one most people forget. Say what you want to happen after the response is used. "Write a polite email" describes the approach. "I want the supplier to agree to the Friday deadline" describes the result. The AI does the second much better because then it knows where the text needs to lead.

Think about how you delegate to someone you trust. You don’t dictate every sentence. You say what needs to happen and let that person choose the words. It’s the same with AI. When you describe the process, it follows that process and the result is left to chance. When you describe the result, it chooses the path that gets there.

Marta, an HR manager at a packaging factory with 120 employees, needed to announce that vacation requests would be submitted through the system instead of by email. The first time, she asked for "a notice about vacation." She got a generic text, rewrote it, asked again, adjusted the tone, and cut it down. Four rounds, forty minutes. The second time, she described the result: "I want every employee to know what’s changing, why it’s changing, and what they need to do by the 1st, in a text that fits on a phone screen." The notice was ready for approval in five minutes.

03 Say what must not happen

The third part is the shortest and saves the most rework. Say what the response must not include. "No emoji." "Don’t start with Dear Sir or Madam." "Don’t promise a deadline." "No more than eight lines." When the AI isn’t given limits, it tends to do more: make it longer, more formal, and full of more promises. What you rule out is what prevents a second round.

This works because AI wants to please you. Without contrary instructions, it adds a "please let us know if you need anything," invents a deadline that sounds good, and dresses things up with adjectives. None of that is a mistake on its part. It’s what a person would do if you said "make it nice" without saying where to stop. Setting limits is how you tell it where to stop.

Renata, a customer service manager at a company that sells management software to 800 small businesses, learned this the hard way. She asked for a reply to a customer complaining about a system error, and the AI wrote, "we’ll fix it by tomorrow." No one had promised tomorrow. She almost sent it. Now every request of hers ends with the same line: "don’t promise a deadline or refund I haven’t listed." Since then, no reply has included a commitment she couldn’t keep.

All three parts fit in a five-line prompt: who you are, the result you want, and what must not happen. It’s a three-field form, and you fill in all three each time. After a week, you’ll do it from memory.

04 The three-part test

Before you hit send, reread your prompt and ask three questions. Does it say who I am and who this is for? Does it say what needs to happen? Does it say what must not happen? If one of the three is missing, the response will fail at exactly that point. Nowhere else. That’s the test, and it’s more reliable than any opinion about whether "AI is good or bad."

The test also helps later, when the response comes out wrong. Instead of deciding that "it didn’t work," you look at the prompt and find the missing part. It’s almost always there. And, as you’ll see in the next lesson, you can fix it without starting over: just add the missing part in the same conversation.

Here’s an honest observation. The three parts don’t guarantee a perfect answer the first time. They guarantee an answer you can adjust, because you know what you asked for and can point out what’s missing. That’s different from a generic answer, where there’s nothing to work with. You can fix the first with one sentence. You throw the second away and start over, and that’s when most people give up.

Paulo, the marketing coordinator for a chain of four building supply stores, asked for "a post about the paint sale." He got an upbeat text full of exclamation points, written for people who are 25. The stores’ customers are between 40 and 65 and are mostly construction workers and homeowners. When he reread the prompt, he saw it immediately: it didn’t say who the post was for. He added a line, "the readers are customers aged 40 to 65, many of them construction professionals; use a direct and respectful tone," and the next post had the right tone.

01 One chat per topic

Before you click any button, here’s a rule that prevents half the problems: each topic gets its own conversation. A AI chat remembers everything said within a conversation and forgets everything outside it. That’s good and bad at the same time. It’s good because you can say “shorter now,” and it knows what you’re talking about. It’s bad because if you use the same conversation for three topics, it mixes them together.

Think of a conversation as a blank page. You write the prompt at the top, the response appears below, and the adjustment goes under the response. Everything on the page counts. When the topic changes, you turn the page. In chat, turning the page means starting a new conversation. It takes one click and doesn’t erase anything: the old page stays saved.

Paulo, who coordinates marketing for four building supply stores, learned this the hard way. In his first week, he used one conversation for everything: a post about the cement sale, then the spreadsheet of the month’s campaigns, then an email to the paint supplier. In response to the third prompt, the AI wrote the supplier email mentioning the cement sale and a column from the spreadsheet. Nothing was wrong from its point of view: it was all on the same page. Since then, Paulo has started one conversation per topic and given each one a name.

02 The four buttons: new conversation, attach, copy, history

You’ll see a screen with lots of icons. It looks like an airplane cockpit, but works like a notebook: old pages are listed on the left, the open page takes up the middle, and the box where you write is at the bottom. You’ll only use four things on this screen. The rest can stay as it is.

The first one is new conversation: a button with a plus sign or a pencil, almost always at the top of the list on the left. It's the previous step's "turn the page." The second is attach: a paperclip next to the writing box. It lets you give the AI a file, a photo, or a table instead of typing out the content. The third is copy: appears below each response, usually as two overlapping squares. One click copies the entire response to the clipboard, ready to paste into an email or WhatsApp. The fourth is the history: the list of old conversations. That's where you find yesterday's worksheet.

Beatriz, the owner of a physical therapy clinic with six professionals, wanted a confirmation message that respected everyone’s schedule. Instead of typing out the whole schedule, she clicked the paperclip, chose a photo of the schedule posted at reception, and wrote: "based on this schedule, write a confirmation for patient A on Thursday at 3 p.m., with the physical therapist who works at that time." The AI read the photo, found the right name, and wrote the message. Beatriz clicked Copy and pasted it into WhatsApp. Three clicks, with no typing beyond the prompt.

03 Three versions and one adjustment

The first response in an AI chat is the most common one, not the best one. It chooses the most likely approach, and the most likely approach tends to be the dullest. There’s a simple way around this, and it uses the conversation to your advantage: ask for three different versions at once, then choose one. Choosing among three is easier than fixing one. You can read all three in twenty seconds and point to the one that sounds most like you.

After you choose, comes the second half, and this is where the conversation shows its value. You don’t need to open anything or explain again who you are or what you want. Write below the response: "the second version, shorter and without the word 'dear'". The AI knows which one is the second, knows what it’s about, and returns only the adjustment. This is called refining in the same conversation, and it’s the pair of prompts that pays off most in everyday work: three versions, one adjustment. They’re both in the kit as prompts 19 and 20, and they’re the practice for this lesson.

Marta, an HR manager at a packaging factory with 120 employees, needed a title for a notice about the cafeteria’s new hours. She asked for three: one direct, one warmer, and one shorter. They came back as "New cafeteria hours starting Monday," "New cafeteria hours: see what’s changing for you," and "Cafeteria: new hours." She chose the second and wrote underneath: "that one, but cut it in half." It came back as "Cafeteria: see what’s changing for you." It took a minute, including the time to read it. Before, she would spend ten minutes staring at the screen and choose the first thing that came out.

Notice what Marta didn’t do: she didn’t start a new chat for the adjustment, paste the announcement again, or explain that it was for the cafeteria. The chat already knew. Every time you start over, you throw away everything the AI has already understood about your request.

04 When the response is bad, the problem is almost always the prompt

It happens: you make a request and the answer comes back generic, too long, or in a tone that doesn’t sound like you. The natural reaction is to think the AI "didn’t understand" or that "it’s having a bad day." It doesn’t get worse in the afternoon. In eight out of ten bad answers, one of the three parts from lesson 1 was missing: who you are, the result you want, or what must not happen. The answer fails exactly where the missing part should have helped.

That’s why the fix isn’t to delete everything and type it all again. It’s to look at the response, ask "what was missing?", and write only the missing part below it, in the same conversation. "I forgot to mention: I’m a customer service manager, and this is going to an upset customer." That’s it. The AI revises it with the new detail without losing the rest. And when you’re not even sure what you want, there’s a way out in the kit: prompt 17, which turns the tables and tells the AI to ask questions before answering.

Renata, a customer service manager at a company that sells management software to 800 customers, tried asking for "a message to let customers know about the update." She got a formal, two-page announcement she would never use. On the first day, she deleted everything and tried three times, with a different sentence each time, and each version was worse. On the second day, she used request 17: "I want to let customers know about an update; before you answer, ask me the five questions you need." The AI asked who the customers were, what was changing, whether they needed to do anything, the tone, and the length. Renata answered one by one. The text she got was six lines long and sounded like her.

01 AI makes mistakes with the same confidence it gets things right

When a person isn’t sure, you can tell. They hesitate, say "I think so," or ask to confirm later. The AI chat It does none of that. It writes an invented number with the same polished phrasing it uses for a correct number. There’s no sign of uncertainty. The wrong answer and the right answer arrive with the same look, at the same pace, with the same confidence.

This isn’t a flaw in any particular product. It’s how this technology works: it puts together the most likely sentence, not a verified one. Most of the time, the likely sentence is also the right one, which is why AI is so useful for writing. But "most of the time" isn’t good enough for a price going out to 800 customers.

Think of a new coworker who’s very articulate, answers everything right away, and never says "I don’t know." You’d like them. You’d also learn within a week not to pass along a number they gave you without checking the spreadsheet. That’s the right relationship with AI: it writes, you check. This doesn’t mean distrusting everything; it means being wary of certain types of passages, which you’ll learn to recognize in the next step.

Renata, a customer service manager, asked the AI for a notice about a system change. The text came ready to send, well written, and said that "the update reduces processing time by 40%." She had never mentioned 40%. The AI completed the sentence with a number that sounded good. Renata almost forwarded it to 800 customers. She stopped because the number was too round, and this lesson will turn that suspicion into a habit.

02 The three things you always check: number, deadline, promise

You don’t need to distrust everything. That’s exhausting and makes people give up using it. You need to watch out for three types of passages, always the same ones. Number: any amount, percentage, quantity, or calculation. Deadline: any date, “by Friday,” “within 48 hours,” “starting in October.” Promise: any commitment on behalf of your company, such as “we’ll refund you,” “at no cost,” “we guarantee.”

These three have something in common: if they’re wrong, someone will hold you accountable. The wrong adjective looks bad. The wrong deadline leads to a complaint. The rule is simple: before passing it along, highlight the three types and check each one against the source. If you don't know the source, the passage doesn't belong.

Marta, in HR, asked for a text explaining the vacation policy and got "employees are entitled to 30 consecutive days, which can be split into up to two periods." At her company, the correct number is up to three periods. Paulo, in marketing, asked for a post and got a price the AI had copied from an example he provided the week before, which was already out of date. In both cases, the AI didn’t lie: it filled in what seemed right.

03 Never paste anyone’s real data

This is the only rule in this lesson with no exceptions. A customer’s full name, ID number, phone number, address, or account balance: none of that goes into the chat. The chat looks like a private conversation on your screen, but it’s a service run by an outside company. What you paste leaves your building. You don’t need to know where it goes to decide it won’t go.

The good news is that the prompt works exactly the same with made-up information. The AI doesn’t need to know the patient’s name is Ana Lúcia to write a good confirmation message. It needs to know that the person is a patient, the appointment is Thursday at 3 p.m., and your tone is warm. Replace the name with "patient A," the amount with any round number, and you’ll get the same result. Then, when it’s time to send the message, add the real name.

What counts as real personal data: full name, ID number, phone number, email address, home address, date of birth, account or student ID number, and any value tied to a specific person. What you can include safely: the type of person ("a patient," "a store customer"), the situation ("they were late twice"), your tone, and the result you want. If you’re unsure whether something is real personal data, treat it as if it is. Replacing it with a made-up name takes two seconds and never goes wrong.

Beatriz, the clinic owner, tested the rescheduling message with "patient A, Thursday appointment." She approved the text, pasted it into WhatsApp, and only then replaced "patient A" with the real name. It took ten extra seconds. The receptionist does the same: she has the text saved with "patient A" and has never typed a real name into the chat. If anyone ever asks what the clinic puts into AI, the answer is "nothing that identifies a person."

04 Ask for the method, not the verdict

When you hear a claim and want to check it, the temptation is to ask the AI itself: "is this right?" Don’t. It answers "yes" and "no" with equal confidence, for the same reason explained in the first step of this lesson. What works is asking for a way to check: "how can I verify this in five minutes, with a source?" Then it gives you something you can do yourself, and you make the call.

Adding one sentence at the end of the request helps a lot: "if you’re not sure, say so." The AI starts flagging where it’s guessing, and that warning is worth more than a polished answer. Request 13 in the kit does exactly that: it asks what would need to be true, how to check in five minutes, and what changes if it’s wrong.

A supplier gave Paulo the claim "70% of renovations start in January," and he wanted to use it in an ad. Instead of asking whether it was true, he pasted in prompt 13. The response said it didn’t know where the figure came from and suggested checking with the industry association or reviewing the four stores’ own sales from previous Januaries. He looked at the sales spreadsheet. The peak was in March. The ad went out with his data, not the supplier’s.

01 Beatriz, from her first message to her first kit

It’s worth bringing the three lessons together in one story, because that’s how they make sense. Beatriz owns a physical therapy clinic with six professionals and about three hundred active patients. She had never used a AI chat for work.

In lesson 1, her first prompt had four words, "message to confirm appointment," and came back sounding like a bank. She added the three parts—who she is, the result she wants, and what must not happen—and the message came back in three lines with a question at the end. In lesson 2, she discovered the four buttons and asked for three versions before choosing one. In lesson 3, she replaced the patient’s name with "patient A" and checked the appointment time on the calendar before sending. It worked, and then a new problem came up: every Monday she opened the chat and typed the whole prompt again from scratch.

This is the point where most people stop using it. Not because the AI failed, but because asking well takes work, and work that keeps repeating without getting easier eventually gets abandoned. The answer isn’t to memorize. It’s to save it. What Beatriz needs, and what you’ll put together in this lesson, is a kit: five requests with her context already included, in a file she can open and copy.

Notice how big the leap is. Beatriz doesn’t need to learn anything more about AI to take this step. She only needs to choose, fill in, and save. Everything she learned in the three lessons still applies; the difference is that now it applies without the cost of repeating yourself.

02 Five, not twenty

The published kit has twenty prompts, and that’s by design: twenty different situations to help you recognize your own. But a file with twenty saved prompts is a file no one opens. If you save everything, you use nothing. The number that works in practice is five, and there’s only one way to choose: does the situation come up every week?

Think about what you do on Monday morning and Friday afternoon. The messages that always need to go out, the text that always gets stuck, the check that’s always delayed. That’s where a saved prompt earns back the effort you put into filling it out. A situation that comes up once a year doesn’t need a prompt card; you can open the published kit when it comes up.

Paulo coordinates marketing for four building supply stores. When he looked at the twenty prompts, he marked eleven as "useful." When he asked which ones he’d use every week, four remained: a post about the week’s new item, three versions of a headline, a comparison between two campaign ideas, and a price check before publishing. He left the fifth open, and a month later added a prompt to summarize Monday’s meeting with the store managers.

The list above is a starting point, not a test. If your fifth prompt is different from the one suggested, that means you know your work better than a list does. Write down your five numbers before moving on to the next step.

03 Fill it out once, use it every week

Every request in the kit has parts between the less-than and greater-than signs, such as "who you are" or "your tone with customers." They’re left blank in the kit because the page is for everyone. In your file, they’re filled in because the file is just for you. That’s the whole difference between the published kit and your kit.

Separate the parts into two types. The fixed parts don’t change from one week to the next: who you are, the size of your team, your tone, who your customers are. The variable parts are the details that change each time: this week’s product name, the deadline this time, what happened today. Fill in the fixed parts once and never touch them again. Leave the variable parts marked so you can swap them out in ten seconds.

Marta, in HR at a factory with 120 employees, filled in the prompt for a notice like this: "I’m an HR manager at a packaging factory with 120 employees, most of them on the production floor, and I write short, direct notices without 'Dear employees.'" That part stays the same, so she never has to type it again. What changes—the rule change and the deadline—stays between the brackets. The following week, the notice about the cafeteria’s new hours took four minutes, three of them spent reading.

A reminder from lesson 3 matters twice as much here: the fixed details can include your name and your company’s name, but never a customer’s or patient’s name, ID, or phone number. The kit is yours; other people’s information isn’t. If a request needs a customer example, write "Customer A" in the file and leave it that way forever.

04 Where to save it, and your Monday routine

The right place for the kit is the simplest place you already use: your phone’s notes app, a text file on your computer, or a note in your note-taking app. No new program. There’s only one criterion: you can open it in two taps without searching. Give it a name you’ll recognize, like "my prompts," and add the date at the end of the file so you know when you last reviewed it.

This may seem technical, but it works like the folder of letter templates every office used to have: each template with the fields that don’t change already filled in, and a blank space for what does. You open the folder, pick the right template, fill in the blank, and send it. The only difference is that here the template goes into the chat, not the printer. You only change the marked parts.

Renata, a customer service manager at a company with eight hundred customers, saved the five requests in a phone note called "pedidos-renata". Her Monday morning routine, before the first support ticket, takes less than ten minutes: she opens the note, copies the request for translating a technical explanation, pastes it into the chat with the technician’s text, swaps in the system name, checks that no new deadline has appeared, and sends it. When a request works well in a new way, she pastes the new version into the note, replacing the old one.

Step six is what makes the kit improve over time instead of getting stale. In two months, your five prompts will be different from the five you have today—and better—because each week added a correction. That’s what separates people who use AI from people who tried it once.