Single learning path · 6 lessons · ~2h30
Everyone is using Astra, the new GPT-6 model, to make games. You’ll use it to find your business’s bottleneck, bring your numbers together on one screen, test your site as a customer, create a visual identity, and find out where your money is leaking. Each lesson ends with a result you can check yourself.
Lessons
Leave with a real problem from your work chosen to hand off to AI, and a sentence that says how you’ll check the result.
Come away with your bottleneck for the next three months stated in one sentence and a plan that’s been through a round of criticism.
Leave with a dashboard that tells you in five seconds whether your quarter is on track, built using only your own numbers.
Leave with a numbered list of what breaks, confuses, or puts off people visiting your site for the first time.
Come away with three social media cards that look like they were made by the same person, based on a single reference you admire.
Leave with a table of your money leaks, with proof on every row, and no cancellations made without your decision.
Lesson 1 of 6 · Foundation
By the end of this lesson, you’ll choose a real problem from your work to hand off to AI and write, in one sentence, how you’ll check whether the result is correct.
Astra arrived, and the internet filled up with little games made in five minutes. Fun, but they don’t pay the bills. The difference between playing around and getting results isn’t the tool: it’s the problem you point it at and your ability to tell whether the answer is right. This lesson addresses that before you make any request.
↓ scroll to study
Astra is the name of the new model from OpenAI, GPT-6, released in September 2026. You talk to it in the same place as always: the ChatGPT chat box. What changed is under the hood. It can work for hours on a long task, find a detail in a huge pile of documents, and see what's on your computer screen much more accurately than the previous model.
This matters to you because, until yesterday, AI was good at short answers. Now it’s good at work. For Regina, an accountant who serves 40 businesses on her own, the difference is between asking “write a deadline reminder” and asking “read the 40 contracts and tell me which clients have a price adjustment clause coming due in October.” The second request was impossible last year. Today it’s an afternoon of work for AI and an hour of review for Regina.
Point AI at the problem you’ve been putting off. The rest is just details.
What you write to the AI has a name: prompt, or simply a request. The five requests in this course are long, and they all follow the same structure. Recognizing the four parts lets you adapt each one to your situation without breaking what makes it work.
The four parts are: context (who you are and what you have on hand), task (what should be done, in order), limits (what the AI cannot do without asking) and evidence (what “done” means, in a way you can open and inspect). When a request fails, it’s almost always because one of these four is missing.
Marcos, the dentist who owns a clinic with three chairs, asked, “improve my schedule.” He got generic productivity advice. He tried again with the four parts: “I own a dental clinic with 3 dentists. Here’s my schedule from the last 2 months. Find the time slots that are empty most often. Don’t suggest changing prices. Give me a table with the day of the week, time, and how many times it was empty.” This time he got the table, and saw that Tuesday at 2 p.m. was empty in 7 out of 8 weeks.
Test yourself
A real estate agent wrote: “Here are the 30 listings I published this year. Rewrite the 5 worst ones to make them more attractive. Don’t make up square footage or prices.” What’s missing from this request?
There’s just one rule for choosing what to hand over to Astra, and it applies to all five prompts: choose the hardest problem you can still verify. If it’s too easy, you won’t get anything you couldn’t do in ten minutes. If it’s too hard to verify, you’ll be left with a polished answer you can’t trust. The sweet spot is where it hurts and where you can still look at the result and say, “that’s right.”
Paulo, the real estate agent, had three candidates. “Write the listing for the house on Bento Street” is too easy: he can do it in his sleep. “Tell me the price per square foot in the Petrópolis neighborhood” is hard to verify: AI answers confidently, but where did the number come from? But “compare my last 12 closings and tell me at which stage I lost the most days” is both hard and verifiable: the closings are his, the dates are in the contracts, and he can check every line.
Before
"Give me some ideas for selling more homes." You get a list of ten tips that work for any real estate agent in the country.
After
"Here are my last 12 closings with dates. At which stage did I lose the most days? Show me the math." You get a table with the days spent at each stage, and it checks them against the contracts.
Several prompts in this course mention subagents. Think of it as a team of interns Astra hires on its own, assigns work to, and then brings back together. You don’t manage any of them. You talk to just one coordinator. That’s what lets you ask for big things: read 40 contracts, test an entire website, review 90 days of statements.
What doesn’t change, with or without helpers, is your requirement: evidence you can open. A table, a file, a screenshot, the exact line from the statement. If the AI says "I checked everything and it’s correct," that’s not evidence; it’s a promise. Accountant Regina has a standard line at the end of every request: "Never mark a task as complete without showing me something I can open." Since she started using it, she’s stopped discovering errors after sending work to a client.
Remember that sentence. It will show up, with slight variations, in all five prompts in the upcoming lessons. It’s what separates an AI that works for you from one that just sounds good.
Practice now 0/4 done
Leave with a problem chosen and a sentence that says how you’ll know it worked. Paper and pen, or your phone’s notes app. About 10 minutes.
Nothing here involves AI yet, and none of your data is touched. Just think and take notes. If the choice seems wrong tomorrow, cross it out and pick another problem: the list will still be useful.
You’ve just done what most people skip: chosen a problem worth spending AI time on, and figured out how you’ll check the answer. The five prompts in the next lessons start here.
Summary
Lesson 2 of 6 · Request 01
By the end of this lesson, you’ll have your bottleneck for the next three months written in one sentence, and a plan to break through it that has already survived a round of criticism.
You have ten problems and don’t know which one to start with. So you start with none, or with the loudest one, which is rarely the most important. This prompt flips the order: AI doesn’t answer anything until it interviews you and finds the one thing that, once unblocked, unblocks the others. Only then does it build the plan. And before showing it to you, it gets someone to attack the plan.
↓ scroll to study
The first phase of the request forbids AI from proposing anything. It can only ask questions, one at a time, and wait for your answer. At least eight questions, digging in four directions: what you’ve been avoiding, what costs you the most time or money, what everything else depends on, and what you would fix if you could fix only one thing. When it’s confident, it names the bottleneck in one sentence and asks you to confirm. If you say no, it keeps going.
The one-question-at-a-time format isn’t a gimmick. When AI dumps ten questions on you at once, you answer the easy ones and skip the painful ones. One at a time, there’s nowhere to hide. Marcos, the dentist, came in thinking his problem was attracting new patients. On the sixth question (“what do you do when someone misses an appointment?”), he realized he was losing 20% of his schedule to no-shows and doing nothing about it. His bottleneck wasn’t attracting patients. It was confirming appointments.
Once the bottleneck is confirmed, the AI builds the 90-day plan. First, it lists all the ways it can help: tasks it handles itself, research it runs, drafts it writes, things it checks every week. Then it groups them into workstreams. For each workstream, it gives you four things: the goal, the first task, what “done” means, and how to measure it. Finally, it tells you what it needs from you: files, access, decisions.
Accountant Regina identified her bottleneck: “Every month I repeat the same request for documents for 40 clients, and half of them are late.” The plan came in three parts: a payment reminder template where only the name and deadline change, a reminder calendar for each client, and a list of the 8 clients who are always late, so she could have a different kind of conversation with them. Each part had a first task for the following week and something she needed to provide: the client list and the deadlines for each tax regime.
What every workstream in the plan needs
Here’s the part nobody does alone. The prompt tells the AI to stop and, before showing you the final plan, submit the plan to a critic. In the original, Astra calls Claude, another AI, as a helper and asks it to argue against the plan: what’s wrong, what’s missing, what’s too ambitious, and what you won’t actually do. The three strongest objections lead to a rewrite. Then the debate runs again on the rewritten version.
If your account can’t trigger this on its own, do it manually in two minutes: open a second conversation (in Claude, or another ChatGPT window), paste in the bottleneck and the plan, and ask for the three strongest objections. Then paste the objections back into the first conversation. The result you get is a short table: objection, what changed. Two rounds. That’s the difference between a plan that sounds good and one that has taken a hit and held up.
Paulo, the real estate agent, got a plan with five workstreams. The critic pointed out: “five workstreams for an agent working alone is a company plan, not a person’s plan; he’ll drop three by the second week.” The plan came back with two workstreams. Paulo says it was the first plan of his that lasted more than a month.
The fourth phase is where the helpers come in. The request says: one helper per workstream, each responsible for their own, with a weekly report in five fixed lines: done, blocked, next, evidence, risk. Everything each workstream produces goes into a folder with its name. The coordinator keeps the plan up to date, flags anything that’s slipping, and never marks a task as done without something you can open.
If you only use AI through chat, without a program that runs tasks on its own computer, this phase changes hands: you do the first task for each workstream and return to the conversation every week to ask for a five-line report on what you delivered. It works the same way. The five lines and the evidence are essential. Dentist Marcos, for the “appointment confirmation” workstream, gets this every Monday: done (messages sent), blocked (3 patients without WhatsApp), next (call the 3), evidence (spreadsheet with responses), risk (the receptionist isn’t following the script). Five lines, and he knows where things stand.
Common mistake
Have the debate once and never again. The plan withstands criticism in week 1 and becomes law. But the request says to repeat the debate at the end of each workstream, because what you learn by executing changes the plan. Skip this, and you’ll spend 90 days flawlessly executing a plan that was outdated by day 20.
Practice now 0/5 done
Leave with the bottleneck in one sentence and a plan after a round of objections. About 15 minutes; the interview may take a little longer if you answer thoughtfully.
It’s just a conversation. Nothing is sent to anyone, nothing is spent, and none of your files are changed. Avoid writing clients’ or patients’ names; use "client A" and "client B." If at any point it seems like the AI is going in the wrong direction, say "go back to the interview" and it will pick up again.
You're my operator for the next three months. We'll work in four phases. Don't skip any. PHASE 1 — INTERVIEW (one question at a time) Your task is to find the single factor that most limits <my business / my professional life> over the next 3 months. Interview me. Ask one question, wait for my answer, then ask the next. Ask at least 8 questions. Dig into this: what I've been avoiding, what costs me the most time or money, what everything else depends on, what I'd fix if I could fix only one thing. When you're confident, name the bottleneck in one sentence and ask me to confirm it. Don't move on until I say yes. PHASE 2 — PLAN Build the plan to break through this bottleneck in 90 days. - List every way you can help me. Be specific: tasks you do yourself, research you conduct, drafts you write, things you check every week. - Group them into workstreams. For each workstream: the objective, the first task, what "done" means, and how we measure it. - Tell me what you need from me: files, access, decisions. Then STOP and run the debate below before I see the final plan. PHASE 3 — THE DEBATE (required) Submit the bottleneck and the full plan to a critic. If you can bring in another AI as a helper, do that. If you can't, give me the text ready to paste into a second conversation and bring back the objections. Ask the critic: what's wrong, what's missing, what's too ambitious, what I won't actually do. Take the three strongest objections seriously and rewrite the plan. Debate the rewritten version again. Show me both rounds in a short table: objection → what changed. Repeat this debate at the end of each workstream, not just once. PHASE 4 — EXECUTION If you can bring in helpers to do tasks: assign one helper per workstream, each responsible for theirs, with a five-line weekly report (done / blocked / next / evidence / risk) and a folder named after the workstream. If you can't, I'll do the first task in each workstream and come back every week; give me the five-line report based on what I bring. You're the coordinator: keep the plan updated, flag what's slipping, and never mark a task done without evidence I can open. Rules: ask before spending money or messaging anyone. If you're guessing, say so. Prefer the hardest problem I can still verify.
You’ve just done what consultants charge a lot to do: answered questions until you found the bottleneck, and received a plan that’s already been challenged before it reached you. Save the objections table. It records why the plan is the way it is.
Summary
Lesson 3 of 6 · Request 02
By the end of this lesson, you’ll have a dashboard built only from your own numbers, open in your browser, that tells you in five seconds whether the quarter is on track.
Your numbers live in ten places: your bank app, a spreadsheet, the card reader, a notebook, your memory. Every month, you “feel” whether things went well. Feelings don’t decide anything. This prompt asks three questions, builds the model, and delivers a screen that answers one thing. If a chart doesn’t help answer it, it’s removed.
↓ scroll to study
The prompt starts by limiting AI to exactly three questions, asked one at a time. First: what was your revenue, month by month, over the last 12 months? Second: what’s the one goal for this quarter that you really want to hit? Third: what are the inputs that you control to get there. It’s not "your metrics." It’s what you do with your hands: calls, proposals, visits, hours delivered.
The third question is the one almost nobody can answer right away, and it’s the most important. Revenue is an outcome. Inputs are causes. The dashboard that comes next is useful only if it connects the two.
In your profession, the inputs are
With the three answers, AI builds the model. There’s nothing mystical about it: it’s a table. The month-by-month revenue trend and the current pace. The quarterly goal as a target, with the gap between where you are and where you need to be. And each input as a lever, with a weekly target and a link to the bigger goal.
After showing you the table, it stops and asks what’s wrong. That pause is the part you can’t skip. Dentist Marcos saw the table say he needed 22 confirmed appointments a day. The clinic has only 3 chairs: its physical maximum is 18. The target was wrong, not the input. He adjusted the quarter’s target to what the setup could handle, and the dashboard was built on honest numbers.
Once you’ve corrected the model, AI delivers the dashboard. It comes as a file you download and open with two clicks. Don’t be put off by the format: it’s like a web page, except it’s stored on your computer. No one else can see it. It’s not on the internet. Close it and it’s gone; open it again and it’s there.
The request describes the screen precisely: dark, calm, a single screen. The big numbers first. Then the revenue chart. Then the goal ring, which shows how much of the quarter is covered. Then the supply gauges. Every number has to come from one of your responses. If data is missing, the display stays blank with a notice about what’s missing, instead of making something up. Hover over any number to see where it came from. And there’s an "update" area: paste in the week’s numbers and everything recalculates.
Accountant Regina pasted in all 12 months and saw in the big number that her current pace put her at 82% of her goal. She hovered over it: the 82% came from the average of the last 3 months. That made sense. For the first time, she knew where she stood without opening five different things.
Common mistake
Accept a number you didn’t provide. When a month is missing, the AI tends to fill in a plausible value to make the chart look good. The request forbids this, but the prohibition only works if you check: hover over each big number and look at its source. If the source says "estimated" or "average," and you didn’t ask for that, remove that month and leave it blank. A dashboard with a gap is honest; one with a guess is dangerous.
The last phase of the request asks what real data sources you can provide: the bank statement, billing system or card reader, issued invoices, calendar, customer records. For each source you connect, AI replaces the field you used to enter by hand with a direct read and tells you in plain words what that number now reads from.
You don’t need to connect anything on day one. Pasting the numbers manually works, and for many people it’s enough for months. Connect things when you get tired of pasting. Realtor Paulo connects just one thing: his bank statement, which he downloads from the app once a month and pastes into the update area. He keeps entering visits by hand because they’re in his head and in no system. A dashboard that’s partly automatic and partly manual, and entirely his.
The request ends with a rule that matters more than anything else: the dashboard answers just one question, "Is the quarter on track?" If a panel doesn’t help answer that, it gets removed. That’s why the dashboard fits on one screen. It’s not a technical limitation. It’s discipline.
The weekly routine is short. Dentist Marcos does it every Monday at 8 a.m., before his first patient: he pastes in the week’s numbers, checks the big number, sees which supply gauge is furthest behind, and picks just one thing to do that week. Five minutes. He doesn’t look at the dashboard again until the next Monday.
The five-minute ritual, every week
Practice now 0/5 done
Leave with the dashboard open on your screen, showing your current pace against the quarterly goal. About 15 minutes, plus the time it takes to find your monthly revenue.
Your numbers stay in your conversation and in a file on your computer. No one is charged, and no accounts are connected for this exercise (phase 4 can come later, if you want). If you prefer, round to the nearest hundred and don’t write down any client names. If you don’t have 12 months handy, use 6: the dashboard starts smaller, and you can fill in the rest later.
Build me a dashboard to track my business. Step 1 — ask me exactly these three questions, one at a time, and wait for each answer: 1. What was my revenue each month for the last 12 months? (I'll paste the numbers or a file.) 2. What is the ONE goal this quarter that I truly want to hit? 3. What inputs do I need to deliver to get there? (Things I control: <proposals sent, visits made, appointments confirmed, hours delivered>.) Step 2 — turn my answers into a model: - The month-by-month revenue trend, at the current pace. - The quarterly goal as a target, with the distance between where I am and where I need to be. - Each input as a lever, with a weekly target and how it connects to the bigger goal. Show me the model in a table and ask me to correct anything that's wrong. Step 3 — build a polished dashboard as a single file I can open in my browser. - Dark, calm, one screen. Big numbers first. Then the revenue chart. Then the goal ring. Then the input gauges. - Every number must come from my answers. No made-up data. If something is missing, leave a blank space saying what's missing. - Include a small "update" area where I can paste the new numbers for the week and the dashboard recalculates. - When I hover over any number, show where it came from. Explain in two lines how to save the file and open it with a double-click. Step 4 — connect what I already have. Ask which of these I can provide: bank statement, billing system or card reader, issued invoices, calendar, customer list. For each one I connect, replace the manual field with the actual data and tell me exactly what it started reading. (Skip this step today; I'll let you know when I want it.) Rule: the dashboard answers just one question: "Is the quarter on track?" If a panel doesn't help answer that, remove it.
You’ve just brought ten places together on one screen that answers one question. Next Monday, five minutes is all it takes to know where the quarter stands.
Summary
Lesson 4 of 6 · Request 03
By the end of this lesson, you’ll have a numbered list, from worst to best, of everything that breaks, confuses, or drives away someone visiting your site for the first time.
You know your site by heart, which makes you the worst person to test it. You know where to click, forgive the button that takes too long, and don’t even notice the form that doesn’t submit. Your new customer doesn’t know any of that. They try, get stuck, and leave without telling you. This prompt puts AI in their place, on a computer and a phone, and it tells you everything it saw.
↓ scroll to study
The original request has two halves. The first is for someone who’s going to build the site from scratch with AI, and it uses an idea worth knowing even if you never build anything: the critics' circuit. The golden rule is that students don’t grade their own work. Whoever builds it submits it to three critics with clean memory, each judging one thing: does it meet the goal? Does it follow the visual standard? Does it measure up to a real reference, side by side?
The second half is the one anyone with a website can use today, and it’s the focus of this lesson: AI pretends to be a first-time customer and tries to do what a customer would do. Dentist Marcos used both approaches. He commissioned the clinic’s website using the critic loop, with a São Paulo clinic’s website that he admires as a reference. Then he asked AI to try to schedule an evaluation as if it were a patient. It was in the second half that he discovered the “schedule” button didn’t do anything on a phone.
If you’re commissioning a website with critics, require
The request starts with a character instruction: "you are a first-time customer, you’ve never seen this site, act like a real person, not a test technician." That changes everything. A technician tests buttons. A customer tries to get something done and gives up when they can’t. For this to work, the AI needs to be able to open real websites: in ChatGPT, that’s agent mode; in Claude, it’s the extension it uses in your browser. It’s one more button in the conversation, not a new program.
The journey has four steps. Land on the home page and say what the business does in five seconds; if you can’t, that’s discovery number one. Try to accomplish the goal by clicking where a normal person would, filling out forms with realistic information, and deliberately making one mistake to see what happens. Follow every link at the top and bottom of the page. Then finish the task and check that the result was actually saved: does the order exist? Did the confirmation arrive? Do all of this twice: on a computer screen and on a phone screen.
Paulo, the real estate agent, set the goal as “schedule a showing for the apartment on the cover.” It worked on the computer. On the phone, the showing form was hidden behind the WhatsApp button. Three months of paid ads were sending people to a form they couldn’t tap.
AI returns a numbered list, from worst to best. Each item has five parts: what it did, what it expected, what actually happened, an image of the screen at that moment, and how to reproduce the problem in one line. The list is also split into three drawers: broken (doesn’t work), confusing (works, but the person doesn’t understand it), and ugly (works and is understandable, but puts people off). This keeps a crooked button from appearing above a form that doesn’t submit.
The most important instruction comes at the end: “don’t fix anything yet; I want the raw findings first.” Only after you say “go” does AI fix the broken items, one at a time, and repeat the exact step that failed to prove it’s resolved. Accountant Regina received 11 items: 2 broken (the contact form didn’t submit and the “areas of practice” link led to a blank page), 6 confusing, and 3 ugly. She had the 2 broken items fixed that afternoon and left the other 9 for a conversation with the person who built the website.
Common mistake
Let the AI fix things as it finds them. It seems efficient, but you lose the raw list and have no way to tell whether the fix created a new problem. The request deliberately separates the two steps: first the complete list, then "go," then one item at a time with proof that the step now works. If the AI starts fixing things on its own, say "stop, just the list."
This test finds what’s broken, confusing, and ugly. It doesn’t tell you whether your offer convinces anyone to buy. It isn’t a security test for break-ins, or a test of how many people the site can handle at once. When you tell someone what you did, say “customer test,” never “load test” or “security audit,” unless you actually did that.
After fixing the phone button, Marcos, the dentist, wanted to know if the website was “good for selling.” The honest answer is: now it doesn’t stop anyone from booking. Whether people will want to book depends on the price, location, a photo of the clinic, and things no navigation test can measure. Knowing what the tool can’t do is what lets you trust what it can do.
Test yourself
After running the customer test and fixing the broken items, which of these statements can you honestly make?
Practice now 0/5 done
Leave with the numbered list, from worst to best, grouped into broken, confusing, and ugly. About 15 minutes; AI works on its own most of the time.
AI only browses, clicks, and fills out forms with test data. It doesn’t buy, pay, or fix anything until you say “go.” If your website creates an account when someone fills out the form, a “Maria Test” will appear on your list: delete it afterward. If you don’t have a website, use the page your customers use most to reach you: the booking link, your Instagram page, or your Google page.
Você é um cliente de primeira viagem. Nunca viu este site. Aja como uma pessoa real, não como um técnico de testes. Site: <endereço do seu site> Seu objetivo: <o que um cliente deveria conseguir fazer, por exemplo: agendar uma avaliação / pedir um orçamento / marcar uma visita> Percorra a jornada inteira no computador, e depois de novo numa tela do tamanho de um celular: 1. Chegue na página inicial. Diga em voz alta o que você acha que este negócio faz, em até 5 segundos. Se não conseguir dizer, essa é a descoberta número 1. 2. Tente cumprir o objetivo. Clique onde uma pessoa normal clicaria. Preencha os formulários com dados realistas de teste (nome "Maria Teste", e-mail inventado). Erre um campo de propósito (e-mail errado, campo vazio) e anote o que acontece. 3. Siga todos os links do topo e do rodapé. Relate os que estão quebrados, lentos ou levam para um lugar surpreendente. 4. Termine o objetivo. Confirme que o resultado ficou salvo de verdade: o pedido existe? a confirmação chegou? o cadastro foi criado? Relate como uma lista numerada, do pior para o melhor. Para cada item: o que você fez, o que esperava, o que aconteceu de fato, uma imagem da tela e como repetir o problema em uma linha. Separe "quebrado" de "confuso" de "feio". Não conserte nada ainda. Quero as descobertas cruas primeiro. Depois, só quando eu disser "vai", conserte os itens quebrados um de cada vez e repita exatamente o passo que falhou para provar que está resolvido.
You’ve just seen your site through the eyes of someone who’s never visited it. The list you have now is what separates a site that exists from a site that works.
Summary
Lesson 5 of 6 · Request 04
By the end of this lesson, you’ll have three social media cards that look like they were made by the same person, created from a single reference you admire and ready to post.
Your posts don’t look like yours. Each one comes out differently: different font, different color, different spacing. People scrolling past don’t recognize you. This prompt takes a reference you like, figures out why it works in terms you can check, and turns it into a style guide that makes the second and third cards look like the first.
↓ scroll to study
The request starts with you showing just one thing: an image, a post, a screen you think looks good. The AI has to say, in five lines, why it works. And here’s the requirement: mechanisms, not adjectives. “Looks elegant” isn’t enough. “The title is four times larger than the body text, there’s just one accent color used no more than twice, and 40% of the space is empty” works because you can check it by looking.
Accountant Regina chose a post by a well-known colleague in her field, packed with big numbers and very little else. AI broke it down: a number in a huge font, a short sentence underneath, a dark background, a single shade of green, wide margins. Five lines. For the first time, Regina could say what she liked instead of “it’s pretty.”
How to choose and use the reference
With all five mechanisms in hand, AI builds a single card using your real content, sized as a vertical post for Instagram. It comes as a file that opens in your browser, and you can also ask for an image to post directly. The card has a short entrance animation and another when tapped, and nothing moves for less than half a second, so it doesn’t feel jittery.
Then comes the part that turns a pretty card into a family: AI extracts the palette. Exact colors, font sizes, spacing, animation speed. That’s the template. And using it, she makes two more cards: one with a highlighted number and one with a quote. Three pieces, one source.
Marcos, the dentist, came away with three cards: a care tip (“replace your toothbrush every 3 months”), a number (“87% of toothaches start without pain”), and a patient testimonial. Posted in the same week, for the first time all three looked like they came from the same clinic.
The last step in the request is the real test. Ask for just one change, for example "change the green to navy blue" or "make the text a little bigger." If the family exists, the AI changes the palette and all three cards change together without breaking anything. If it has to edit each card separately, there wasn’t a family; there were three similar designs.
Paulo, the real estate agent, asked, “make the real estate agency’s name smaller, in the corner.” All three cards changed in the same response. He saved the style sheet in a folder, and now whenever he wants a new card, he pastes in the style sheet and the content, and the card comes out matching the family.
Common mistake
Ask for all three cards at once, without going through the palette. AI delivers three beautiful pieces that look like they were made by three different people, because each one was decided from scratch. The order of the request is the safeguard: one card, then the style sheet, then the other two based on it. If you skipped the style sheet, the one revision test will reveal it: the three don’t change together.
Practice now 0/5 done
Leave with three cards that look like they belong together, the style guide saved, and a revision applied to all three at once. About 15 minutes.
Nothing gets posted. None of your accounts are connected. The reference is only for inspiration: the prompt forbids copying its text, logo, or photo. If a card comes out too similar to the reference, ask AI to make it more distinct. The files stay on your computer, and you decide what to publish.
Eu faço conteúdo para o Instagram sobre <seu ramo: contabilidade para pequenas empresas / saúde bucal / imóveis na sua cidade>. Monte para mim um pequeno sistema visual que eu consiga postar. Referência: <cole a imagem ou o endereço de um post que você admira> Use a referência só como inspiração. Não copie texto, logotipo, foto nem nome dela. Passo 1 — desmonte a referência. Em cinco linhas, me diga o que a faz boa: tamanhos de letra, espaçamentos, a única cor de destaque, o que se mexe e em quanto tempo. Mecanismos que eu consiga conferir olhando, não adjetivos. Passo 2 — construa UM cartão que siga esses mecanismos, como um arquivo que eu abro no navegador. Conteúdo real meu: <uma dica curta do seu ramo>. Tamanho vertical de post (4:5). Uma animação de entrada e uma ao tocar. Nada se mexe em menos de meio segundo. Entregue também uma imagem do cartão para eu postar direto. Passo 3 — transforme em sistema. Extraia do cartão a cartela: as cores exatas, os tamanhos de letra, os espaçamentos e as velocidades de animação, num arquivo separado, para que o próximo cartão saia idêntico em estilo. Depois construa mais dois cartões a partir da mesma cartela: um cartão de número em destaque (<um dado do seu ramo>) e um cartão de citação (<uma frase sua ou de um cliente>). Passo 4 — eu vou pedir uma revisão. Aplique nos três cartões de uma vez, mudando só a cartela, sem quebrar nada. Entregue: os três cartões (arquivo e imagem de cada um), a cartela, e duas linhas me dizendo como abrir os arquivos com dois cliques.
You’ve just gone from posts that looked like they came from three people to three cards that look like they came from one. The palette in the folder is what ensures the fourth, fifth, and twentieth will look that way too.
Summary
Lesson 6 of 6 · Request 05
By the end of this lesson, you’ll have a table of your money leaks, from largest to smallest, with proof for every line and no cancellations made without your decision.
No one likes opening a statement and reading it line by line. So subscriptions you don’t use keep going, a tool that does the same thing as another one stays, and the thing you swear is a priority never gets a dollar. This prompt does the tedious reading for you, but accepts nothing without proof: every leak comes with the statement line and the confirming email. And it doesn’t cancel anything. You decide.
↓ scroll to study
The request looks for three different things. Duplicate: the same thing charged twice, or two tools that do the same job. Unused: what you pay for and haven’t used in more than 60 days, while AI looks through your emails for signs of access or use. Missing: what you say is a priority and where your money disagrees with your plan. The third type is the one nobody looks for, and it teaches you the most.
To do this, you provide three things: your bank or credit card statement, the last 90 days of email (or your folder of receipts), and your priorities for the quarter in a few lines. Without your priorities, the third type doesn’t exist.
Accountant Regina, who expects every client to stay organized, had never looked at her own statement this way. She provided 90 days of credit card transactions and her list of priorities: “grow my presumed-profit clients.” AI found two invoice-issuing subscriptions that did the same thing, a course paid for in January with no access since March, and zero reais spent on anything related to the priority she herself had written down. Three types, three findings, one afternoon.
The first instruction in the request is “read everything before saying anything.” AI builds the complete list of recurring charges, one-time expenses above an amount you choose, and every subscription mentioned in email. Only then does it start pointing things out. This prevents a premature finding that appears on the first page of the statement and disappears after reading the third.
Next comes the rule that makes the result trustworthy: every finding comes with proof. The exact line from the statement, the date, the amount, and the email or other evidence that supports the conclusion. If the AI can’t find proof, it marks the item “unconfirmed” and doesn’t guess. Dentist Marcos received a “unused” finding about the clinic’s scheduling program; the proof showed that the access emails belonged to the receptionist, not him. It wasn’t a leak. Without the proof, he would have canceled the program the clinic uses every day.
Common mistake
Treating a marked line as money saved. The table says “R$ 2.400 per year in leaks,” and you already feel relieved. But a flagged line isn’t a real back in your pocket. Money is saved only when you confirm a charge, cancel it yourself, and it disappears from the next statement. The request separates “confirmed” from “unconfirmed” and doesn’t let AI cancel anything for that very reason.
The result is a table, ordered from the biggest leak to the smallest. Each row includes: the leak, its monthly cost, its yearly cost, the proof, and what to do (cancel, consolidate, keep, or start). At the bottom are two totals that never get combined: confirmed and unconfirmed. "Start" is the column for the third type: money you should be spending on your priority but aren’t.
Paulo, the real estate agent, got 9 line items. The three biggest were an ad listing site he hadn’t opened since May, a digital signature tool duplicated by the one from his real estate agency, and zero dollars in the “start” column for professional photos, even though “better listings” was his top priority. Confirmed: R$ 310 per month. Unconfirmed: R$ 89. He knew exactly what order to act in.
The last step in the request is for the AI to draft the cancellation emails and the text for the one purchase you should make, and send nothing. You read, you decide, you press the button. It’s the same discipline that ran through all six lessons: the AI prepares, the evidence comes with it, and the decision is yours.
And this is where the course comes full circle. The one rule applies to all five prompts and any others you come up with: choose the hardest problem you can still verify. Give the AI the files and tools, and a clear definition of “done.” Then open the result yourself. Dentist Marcos sums it up this way: “I don’t understand how it does it, and I don’t need to. I understand what it gives me because I can open it.”
What to do with each row in the table, in order
Practice now 0/5 done
Leave with the table, from largest to smallest, with proof on every row and confirmed and unconfirmed totals kept separate. About 15 minutes, plus the time it takes to download your statement.
Before you paste the statement, delete the account and card numbers and replace people’s names with initials. The AI won’t cancel, send, or buy anything: it only builds the table and writes drafts that stay in the conversation. If you don’t want to share emails, skip that part: findings marked “unused” will be labeled “unconfirmed,” and that’s fine. Start with 30 days of statements if 90 seems like too much.
Faça uma auditoria de para onde vai o meu dinheiro. Encontre os vazamentos. Prove cada um. Vou te dar: um extrato (banco / cartão / maquininha), meus últimos 90 dias de e-mails (ou a pasta de comprovantes) e minhas prioridades para este trimestre. Passo 1 — leia tudo antes de dizer qualquer coisa. Monte a lista de toda cobrança recorrente, de todo gasto avulso acima de R$ <150> e de toda assinatura mencionada nos e-mails. Passo 2 — encontre três tipos de vazamento: - DUPLICADO: a mesma coisa cobrada duas vezes, ou duas ferramentas que fazem o mesmo serviço. - SEM USO: coisas que eu pago e não uso há mais de 60 dias (procure nos e-mails sinais de acesso ou de uso). - FALTANDO: coisas que eu digo que são prioridade e nas quais não estou gastando; onde meu dinheiro discorda do meu plano. Passo 3 — para cada achado, me dê o comprovante: a linha exata do extrato, a data, o valor e o e-mail ou evidência que sustenta. Se não encontrar evidência, marque "não confirmado". Não chute. Passo 4 — uma tabela, do maior para o menor: vazamento → custo por mês → custo por ano → comprovante → o que eu devo fazer (cancelar / juntar / manter / começar). Totais no rodapé, com "confirmado" e "não confirmado" separados. Passo 5 — escreva os rascunhos dos e-mails de cancelamento e o texto da única compra que eu deveria fazer, mas não envie nada. Eu decido. Regra: uma linha marcada não é dinheiro economizado. Só uma cobrança confirmada e cancelada é. Minhas prioridades deste trimestre: <escreva em 2 ou 3 linhas>
You’ve just done the reading you’d been putting off, with proof in every line and without handing the decision to anyone. The course ends here, and the rule remains: the hardest problem you can still verify.
Summary