Training · Independent Professional
By the end of this learning path, you’ll have a professional second brain: agents that prepare for appointments, draft documents, and research with sources—in your style. AI drafts; you review and sign.
Lessons
Leave knowing what you’ll build, how long it will take, where to start — and what to do when something goes wrong.
Leave with the 5 tasks in your work that could become agents — prioritized and written down.
Leave with your library of 5 professional prompts — and never write “help me with…” to an AI again.
Leave with a working triage agent — tested on a real case and ready to use tomorrow morning.
Leave with your professional knowledge base started — the fuel that makes agents 10× better.
Leave with your first project organized on your computer — typing just 6 commands, with a browser-based alternative.
Leave with a page or automation made by AI — you give the instruction and review it, without writing code.
Leave with an agent connected to a real external tool — knowing the cost before you use it.
Leave with everything you built set up in one home — with rules, memory, and a 20-minute weekly ritual.
Liberal Professional Training · Lesson 0
By the end of this lesson, you’ll know what you’ll build in this training and how much time each part takes—and you’ll leave with your first project chosen and written down.
Your time is valuable, and your schedule won’t tolerate a drawn-out course. Before investing another minute, you deserve to see the whole map: what’s on the other side, what’s expected of you, and where to start without taking the wrong door.
↓ role to study
Anyone who was working in 1995 remembers: computers arrived, and typewriters disappeared within a few years. Then came the internet, and the counter turned into email, websites, and WhatsApp. In both shifts, those who learned early worked better — those who waited had to catch up.
The third turning point is agent: a digital employee that receives a mission from you, does the work, and comes back with the result. It’s not magic or a fad — it’s the same story as the other two turning points, in the next chapter.
An attorney with her own practice experienced the first two shifts: she replaced the Olivetti with Word, and the mailbag with email. Today she spends 6 hours a week on routine drafts — exactly the kind of task the third shift can give back to her.
Before · without agents
The lawyer drafts every routine document from scratch: 6 hours a week of typing that doesn’t require her judgment.
After · with agents
The agent delivers her draft in her preferred format at 6 a.m.; she reviews, corrects, and signs it in 1 hour.
Balance: 5 hours per week back — per delegated task.
This training doesn’t end with “I get it.” Each module ends with something built and running in your work: a task map, an instruction library, a triage agent, a knowledge base — all the way to the final setup, your AI Operating System.
The ladder follows your pace: modules 1 through 4 use only the browser you already use. The more technical part comes in module 5—after you’ve had four wins—and there’s always an alternative that requires nothing new on your computer.
For a doctor in private practice, the path takes shape like this: by module 3, they already have an agent that prepares each appointment; by module 8, the entire practice has a digital home with rules and memory.
Each lesson takes 12 to 25 minutes and tells you how long it will take before you begin. Every lesson includes a 5- to 15-minute exercise using material from YOUR work — never a made-up exercise. If you only have 20 minutes today, look at the learning path and choose a lesson that fits.
An accountant with an unpredictable schedule works through the training between one month-end close and the next: one lesson per session, no marathon. What she sets up on Tuesday keeps working on Wednesday—the progress is saved on this page.
You’re not going to watch a course. You’re going to build things that work for you.
At some point, a strange message will appear on the screen. That doesn’t mean you’re “not cut out for this”: it’s a normal part of digital work for anyone, at any age. Experienced professionals know that problems are part of the workday — it’s the same here.
The process has three steps: copy the entire message, paste it into the AI, and ask “I got this error; what should I do?” The AI reads errors better than any manual and gives you the next step. No one in this training solves errors by memorizing—everyone solves them by asking.
When the attorney first saw “file not found,” she pasted the message into AI and got back: “there’s an extra space in the name — rename it like this.” Thirty seconds, problem solved, and she was never afraid of that message again.
Test yourself
An error message appeared in the middle of a step. What’s your next move?
Practice Now 0/3 done
Leave with your first project decided and written down, in ~5 minutes.
Nothing here accesses your computer or sends any data anywhere—it’s pen and paper (or a note on your phone). If you change your mind later, just cross it out and choose another.
You just decided where to start—with a name, number of hours, and success criteria. That’s more than most people do before hiring any employee.
Summary
Liberal Professional Training · Module 1
By the end of this lesson, you’ll be able to review your week and identify, using clear criteria, the 5 tasks that could become agents—with your first candidate chosen.
Everyone says “use AI,” but no one says where. Without a clear standard, you try the tool on the wrong task, the result disappoints, and you conclude, “this isn’t for me.” This lesson gives you the filter that prevents that cycle.
↓ role to study
The working definition fits in one sentence: an agent is a digital employee that receives a mission from you, uses tools to carry it out, and returns the finished result for your review. No technical jargon: mission, execution, delivery.
Notice what the sentence does NOT say. It doesn’t say the agent decides for you or signs anything. The agent executes; the judgment — what took you 20 years to build — is still yours, and that’s what makes it safe.
In the office, a doctor’s mission for the agent sounds like this: "receive the basic details of this case, summarize the history, raise the hypotheses to explore, and list the questions I can’t forget to ask at the appointment." Ten minutes before the patient comes in, the prep is on the table.
An agent isn’t a program you use. It’s an employee you direct.
Three different things often arrive wrapped in the same word, “AI.” Chat is a conversation with a consultant: you ask, it answers, and nothing happens outside the screen. Automation is a conveyor belt: it repeats the same step, the same way, without thinking — good for things that never change.
The agent sits in the middle: it receives the mission, chooses a path within YOUR rules, and uses tools to deliver. When the task requires mindless repetition, a production line. When it requires a one-off answer, a conversation. When it requires doing things your way, an agent.
The lawyer uses all three without mixing them up: asks a deadline in chat, lets automation remind her about hearings, and gives the agent new-case triage — because triage follows her rules, and no workflow knows her rules.
Test yourself
Send the same appointment reminder every day at 8 a.m. to that day’s patients — what does that call for?
Office won because it bundled office work: documents, spreadsheets, presentations, email—one package, for every profession. Agents repeat that move one layer up: instead of bundling documents, they bundle roles—who triages, who drafts, who researches, who responds.
The practical consequence is the same as in 1995: this won’t be optional for long. Professionals who build their agents first are better prepared to serve clients and deliver faster—for the same price as those who don’t.
An architect can see the impact across her week: one agent summarizes the project description, another finds city regulations with sources, and another prepares the construction meeting agenda. Three functions that used to take up her evenings.
Not every task is a fit. The ones that are pass a three-question test: a input is clear (can you tell what the agent receives)? The rules are known (can you explain how you decide)? A output is verifiable (can you check in minutes whether it turned out well)?
Three yeses: a strong candidate. One no: either the task needs to be better designed, or it relies so much on your judgment that you should keep handling it — at least for now.
The accountant tests the weekend reconciliation: clear input (the month’s statements), known rules (her chart of accounts), verifiable output (it balances or it doesn’t). Three yeses—a perfect candidate.
Before · the accountant’s Friday
Entire reconciliation done by hand: 4 hours classifying transactions one by one at the end of the workday.
After · with an agent
The agent prepares a draft classified according to her rules; the accountant checks the exceptions in 40 minutes.
Balance: 3h20 returned per week — in a single task that passed the three-question test.
Now the filter becomes a list. The inventory is 10 questions about your week—all variations on one question: how many hours do you spend on tasks that don’t require YOUR judgment? Each answer is a task; each task gets a score for frequency, time spent, and how much of a hassle it is.
A warning that can save you months: the temptation is to start with the MOST IMPORTANT task. Resist. A good agent starts with the most REPETITIVE one — known rules, and errors that are easy to review. The important task comes later, once you’re confident directing the tool.
Dr. Aurélio, a cardiologist, did the inventory during a lunch break. He discovered he was spending 3 hours a week writing the same referral letter and changing the names. It wasn’t the most meaningful task of his week—it was the most repetitive. It was the first one he delegated, and the one that paid for the course.
Practice Now 0/3 done
Leave with 5 prioritized tasks — your opportunity map — in ~10 minutes.
This is paper and memory: no client or patient data goes anywhere. Note tasks by type ("referral letter"), never by the name of the person you're treating.
WEEKLY INVENTORY — answer based on the last 7 days: 1. What have you drafted more than once, changing very little? 2. What do you summarize for other people (client, patient, team)? 3. What do you look up or research every week? 4. What preparation do you repeat before each appointment or meeting? 5. What document do you always review using the same criteria? 6. What question do your clients/patients ask repeatedly? 7. What do you copy from one place and paste into another? 8. What report or account statement comes back every month? 9. What do you put off because it's tedious, not because it's difficult? 10. If you hired an assistant tomorrow, what would their first task be?
You just mapped where your valuable hours are leaking away—5 tasks, ranked by number, with the first one selected. No consultant would do it differently.
Summary
Liberal Professional Training · Module 2
By the end of this lesson, you’ll be able to turn any task of yours into an 8-part work order—and you’ll leave with the first prompt in your library run and tested.
Nine out of ten disappointments with AI start with the request, not the tool. "Help me with…" gives you the internet’s average—and your time is far too valuable to spend editing mediocrity. Those who write work orders get drafts; those who write loose sentences get a draft from an intern with no context.
↓ role to study
Think of the best assistant you've ever had. Even they wouldn't get it right with the request "help me with this contract." Everything is missing: help how? Toward what goal? To whose standard? AI has the same problem—without instructions, it fills the gaps with the internet's average, and the internet's average doesn't bear your signature.
A prompt isn't a polished phrase: it's an operational instruction. The good news is, you already know how to do this—every experienced professional knows how to delegate precisely. You just need to transfer that skill to writing.
The doctor who types “write a referral letter” gets a textbook passage. The one who writes “refer to an endocrinologist, patient with these findings, in my 3-paragraph format, colleague-to-colleague tone” gets the letter he would have written himself at 11 p.m.—only at 8 a.m.
Common mistake
Copy a ready-made “miracle prompt” from the internet and use it without filling in your own context. This happens because it seems like a shortcut—but someone else's prompt carries someone else's context. The generic result "proves" that AI doesn't work, and the tool gets blamed for the request. Use templates as a skeleton, always filling in YOUR own gaps.
Every complete work order answers eight questions, in this order: role (who the AI should be), objective (what should exist at the end), context (who you are and what your standard is), rules (what never to do), input (the case material), expected output (format and tone), examples (an excerpt of your work, in your style) and quality criteria (how you will judge it).
It seems like a lot—and that's exactly why it works. You write the eight blocks once for each type of task; next time, you change only the input. It's the same principle as your contract template or care protocol: effort the first time, speed forever after.
The lawyer set up the 8 blocks for triaging lease contracts: role (“assistant at my real estate law firm”), rules (“flag missing clauses using my list”), examples (one of her old legal opinions). Today, each new contract takes a paste and a 40-second wait.
Test yourself
"You are a careful reviewer. Review this report." — of the essential blocks, what is the most serious omission here?
For anyone technically responsible for what they sign, two criteria go into EVERY work order, without exception. First: never invent facts, numbers, or references. Second: every source, citation, or piece of data the AI provides must be marked as [TO CHECK] — and verification is a step in your workflow, not optional courtesy.
This isn't distrust of the tool; it's the same rigor you already apply to any brilliant intern. AI writes with the same confidence when it's right and when it's wrong — the judgment is yours, and now it's written right INTO the request.
An opinion containing fabricated case law can destroy, in a single hearing, the reputation the attorney spent 20 years building. With the rule in the prompt, every cited decision is marked [TO VERIFY] — she checks all three in ten minutes against the original court source and signs with confidence.
AI drafts. You sign. The order matters here.
The master template is the skeleton of the 8 blocks, with gaps marked like this: <replace this>. The rule is strict: an unfilled gap means the instruction isn’t sent. That’s what sets your instruction apart from a prompt copied from the internet.
On top of it, every independent professional builds the same library of 5 missions: case screening, draft of an opinion or report, document review, research with sources e technical response to a client or patient. Five saved work orders cover most of the repetitive routine in a clinic and office.
The accountant filled in the template for “tax deadline reminder”: the context is her clients’ profile, and the only missing inputs are the tax name and date. One template, dozens of reminders a month—each one sounding like her.
It’s worth comparing them side by side on the same real task. A vague request isn’t faster: it just pushes the work to later, when you edit a generic text while grumbling. The operational request takes two more minutes at the start and gives you back an almost-finished text.
The doctor tested post-appointment instructions: with a vague request, the answer came back with technical jargon no patient could understand, so he rewrote nearly everything. With an operational request—role, non-expert audience, warm tone, 5 items, no technical terms left unexplained—it was ready to print.
Before · "write post-appointment instructions"
Manual-style text, full of jargon. The doctor rewrites 80% — about 25 minutes of editing per document.
After · order of 8 blocks
Text in their voice, in patient-friendly language. They adjust two sentences — 4 minutes, including review.
Balance: ~20 minutes per document — paid for with 2 extra minutes spent writing the request. One time.
Was the first response 80% right? Don't rewrite the request from scratch—you'd throw away what it already got right. Just say to the AI itself: "compare it with my example and tell me what’s missing for it to be ready". Its response points to the weak block; you add a rule there and run it again.
After two or three rounds of this cycle, the prompt stabilizes. From there, it goes into the library and becomes an asset: every adjustment you made is your experience turned into permanent instructions.
The architect ran the prompt for a project description and found the text accurate but cold. She asked what was missing; the AI pointed out that there was no example of her tone. She pasted in two paragraphs from an old project description—the next version came out in her voice.
Practice Now 0/4 done
Leave with the library’s first prompt tested in a chat AI, in ~15 minutes.
Use a fictional case or an anonymized real one: change names, dates, and any number that could identify someone—professional confidentiality (OAB, CFM, LGPD) does not belong in an AI chat. The test works just as well with changed data, and deleting the conversation afterward takes one click.
PAPEL: Você é assistente do meu <escritório/consultório>, especializado em <sua área>. OBJETIVO: <o que deve existir no final — ex.: triagem deste caso com perguntas para o atendimento>. CONTEXTO: Sou <profissão>, atendo <tipo de cliente/paciente>. Meu padrão: <2 linhas sobre seu estilo>. REGRAS: Não invente fato, número nem referência. Marque toda fonte como [A VERIFICAR]. Não use dado identificável de pessoas. <suas regras — ex.: nunca prometer resultado>. ENTRADA: <o material do caso, anonimizado>. SAÍDA ESPERADA: <formato — ex.: 3 seções: resumo, pontos de atenção, perguntas — e o tom>. EXEMPLO: <cole um trecho seu, do jeito que você gosta>. QUALIDADE: Está pronto quando <critérios objetivos — ex.: cada afirmação tem fonte marcada>.
You just wrote your first professional work order—and it already gave you work back in your style. This text is an asset: saved, it works every week.
Summary
Liberal Professional Training · Module 3
By the end of this lesson, you’ll have a working triage agent—set up as a permanent document, tested with one case, and ready to use tomorrow morning.
Last lesson’s prompt disappears when the conversation ends—and explaining everything again each time gets more tiring than doing it by hand. Today you’ll write an onboarding manual for a digital employee: who it is, what it does, how you like things done. Once. Forever.
↓ role to study
The prompt you ran in module 2 is an instruction for ONE conversation. An agent is that same instruction made permanent: written in a document the tool reads before every conversation, every day, without you repeating anything. A prompt is a request; an agent is a hire.
The practical consequence: you only have to explain how you work once. After that, each use starts where the previous one stopped being painful.
The doctor from module 1 ran triage by pasting the entire prompt for each patient. As an agent, he opens the conversation and pastes only the case details—and the output comes in his format because the rules already live there.
Common mistake
Build a Swiss Army knife agent that "does everything": triages, drafts, researches, and replies to email. This happens because it seems efficient—but a multi-purpose mission dilutes the rules, and the agent becomes mediocre at all of them. A good agent has ONE mission; a library means having several, each focused on its own.
Test yourself
Yesterday, you explained your standards in a conversation. Today, you open another one and the AI has “forgotten” everything. What was missing?
The agent document is ordinary text, organized into five parts: identity (who he is and who he works for), mission (his only task), rules (what never to do — including the source screening criteria from module 2), response format (the structure you want to receive) and examples from your history (two real, anonymized excerpts in your style).
Notice: this is the 8-block prompt reorganized to apply every time. What used to be “input” comes out of the document — it arrives with each case. The rest is fixed as house policy.
In the lawyer’s version, the identity says "intake assistant for my family law firm"; the mission, "receive the initial account and return a summary, risks, and questions for the first meeting"; the rules prohibit promising results and require a marked deadline [TO VERIFY].
Every major AI tool has a place to save permanent instructions: in Claude, it’s called Project, in ChatGPT, Custom GPT, in Gemini, Gem. Different names, same idea: a folder where your agent document is saved, and every conversation opened there starts out knowing the rules.
The creation screen looks like something for a technician, but it’s an intake form: one field for the name, a large field to paste in the manual. You fill it out once and you’re done—there’s nothing to “program.”
The accountant created hers in a Claude Project: named it “Triage — business clients,” pasted a document into the instructions field, and saved it. The next day, her partner used the same folder—and the agent responded according to the firm’s standards, not the person asking.
The 4 Steps—in Any Tool
A new employee doesn’t take over the entire client portfolio on day one — they go through a probation period. It’s the same with an agent: before using it for real, give it three cases you know well and compare each answer against your quality criteria from module 2.
Three cases, not one: the first might be right by luck. The flaws show up in the second and third — the tone that slips, the missing question, the risk it didn’t catch.
Dr. Marta, a family lawyer, tested the triage agent with three old accounts she remembered in every detail. The first was flawless. In the second, the agent suggested a deadline without marking [TO VERIFY] — caught red-handed. In the third, it forgot to ask about minor children. Two adjustments noted, fifteen minutes — and she knew exactly what to fix before the first real case.
When a dry run catches an error, the fix isn’t to delete everything: open the agent document and add ONE rule that prevents that error. Did the agent get the tone wrong? Add an example of the right tone. Did it forget a question? Add it to the list of required questions.
Every correction turns your experience into a permanent policy — today’s agent is better than yesterday’s, and tomorrow’s inherits everything. That’s how “a prompt that worked” becomes an employee who improves over time.
The architect’s agent described finishes with empty adjectives ("elegant," "sophisticated"). She added the rule "describe the material, dimensions, and catalog reference—never use an adjective without supporting details" and pasted in a paragraph from an old project specification. The problem never came back.
Practice Now 0/4 done
Leave with Project 1 working — a prepared appointment in less than 10 minutes — in ~15 minutes.
Test with a FICTIONAL case or an old anonymized case (names, dates, and numbers changed) — never put identifiable client or patient data into AI. Did anything go wrong? The document is text: correct it and save it again; nothing breaks.
IDENTIDADE: Você é o assistente de triagem do meu <consultório/escritório> de <área>. MISSÃO (única): receber os dados básicos de um caso e devolver a preparação do atendimento. REGRAS: Não invente fato, número nem referência — marque tudo que precisar de conferência como [A VERIFICAR]. Não use dado identificável. Não prometa resultado. <suas regras>. FORMATO DE RESPOSTA, sempre nesta ordem: 1) Resumo do caso em 5 linhas; 2) Hipóteses ou pontos a explorar (com grau de confiança); 3) Perguntas que não posso esquecer de fazer no atendimento. EXEMPLOS: <cole 1–2 trechos seus, anonimizados, do jeito que você escreve>.
You just hired your first digital employee: one mission, your rules, tested. Tomorrow morning, it will prepare for your first real client meeting—in minutes, not half an hour.
Summary
Liberal Professional Training · Module 4
By the end of this lesson, you’ll have started your professional knowledge base—the structure mapped out and the first file written with 5 true things about your work.
Twenty years of experience trapped in your head don’t work while you sleep. The agent you built is good—but it responds like a talented intern because it doesn’t know your templates, protocols, and arguments. Today, that knowledge starts moving out of your head and becoming fuel.
↓ role to study
Think of the best intern you've ever had: smart, fast—and useless during the first week because they didn't know how things worked. AI is that intern every day, forever, UNLESS you give it what you'd give a partner: your templates, your protocols, your way of doing things.
This set of files has a name: your knowledge base. It’s the multiplier — the same agent, with the knowledge base, responds like someone who’s worked with you for ten years.
The doctor’s intake agent was good after module 3. When it got access to his first-visit protocol, it became great: the suggested questions started following the sequence he had refined over twenty years of practice.
Before · agent without a knowledge base
Accurate but generic triage: textbook questions, manual-like tone, no memory of the practice’s patterns.
After · agent + knowledge base
Triage using the practice’s protocol: their sequence of questions, their alerts, the team’s vocabulary.
Balance: the same agent, ten years more experienced — the entire difference lies in the files it reads.
The knowledge base is written in Markdown — and before the word worries you: Markdown is a paper binder, not code. Folders are dividers, files are cards, headings are colored tabs. Anyone who has organized an office archive can understand how it works in a minute.
Six symbols do it all: # form title · ## internal tab · - list item · **text** highlight · [name](address) link to another card · and vertical bars for a table. Nothing else — if it seems more complex, you're doing more than you need to.
The lawyer opened the “modelo-parecer-locação.md” file: # in the model name, ## in each section of the opinion, a list of the clauses she always checks. Readable for her, the intern — and the agent.
The knowledge base’s minimum structure has eight folders: context (who you are), projects, clients (always anonymized), processes, prompts (the module 2 library), agents (the module 3 documents), references e decisions. Don’t add more folders in the first month — too much structure is a way to procrastinate.
For independent professionals, two are worth their weight in gold: processes, where your customer service protocols become reference sheets that agents consult, and decisions, where your theses — why you do things this way and not another — are recorded with the rationale.
In the accountant's foundation, "processes/monthly-close.md" describes the checklist she spent years refining; "decisions/tax-regime-clinics.md" records the position she defends and its exceptions. Her agents cite both instead of giving opinions off the cuff.
In your profession, what’s most valuable first
It goes into the knowledge base: models, protocols, theses, writing patterns, the organization’s vocabulary, decisions with justifications — everything that is YOUR knowledge and doesn’t identify anyone. It’s material you’d share in a course for colleagues without hesitation.
Never enter any folder or tool: name, CPF, address, identifiable medical record, case number with parties’ names, client financial data. Professional confidentiality (OAB, CFM) and LGPD aren’t technology details—they’re your professional responsibility, and you can’t outsource it. Turn a real case into a case type: "patient, 58, with high blood pressure" instead of the name; "tenant behind on rent since March" instead of the person.
The doctor turned ten memorable appointments into ten anonymized sample cases for the reference folder. The agent learns the clinical pattern the same way—and no patient can be identified here or in any data leak.
Test yourself
The lawyer wants the agent to learn from a real eviction case. What can go into the knowledge base?
There are two ways the knowledge base can reach the agent. Paste: you copy an excerpt and paste it into the conversation — fast, focused, good for what changes from case to case. Attach: you add the file to the Project (the same folder as module 3) — permanently, and the agent starts consulting it in every conversation.
The rule of thumb: recurring work belongs in the Project; case-specific details get pasted in as needed. Client intake protocol? Attach it. Today’s client account? Paste it in.
The lawyer attached three files to the Triage Project: the new-client checklist, the summary template, and the legal-theories file. From then on, each triage cited HER checklist — without her pasting anything beyond that day’s account.
Your experience can only work for you once it’s out of your head.
Practice Now 0/3 done
Leave with the structure designed and “meu-trabalho.md” written with 5 true lines, in ~12 minutes.
Use any notes app or Notepad—you don’t need anything new on your computer yet (the move to actual folders is in module 5). And the standing rule: not a single line identifies a client or patient.
# Meu trabalho - Sou <profissão> e atendo <quem, em que tipo de caso>. - Minha semana tem, sempre: <as 3 tarefas mais repetidas>. - Meu padrão de qualidade: <como você reconhece um trabalho pronto>. - Eu nunca: <seus limites — ex.: prometo resultado, uso termo X>. - Vocabulário da casa: <3 termos seus e o que significam>.
You just started the most durable asset in this training: your experience in files that any agent—today or ten years from now—can read.
Summary
Liberal Professional Training · Module 5 · with a route that avoids the black screen
By the end of this lesson, you’ll be able to open the black screen, create the home for your knowledge base with 6 commands, and save its first version—or do the same in your browser without installing anything on your computer.
This is where most courses lose people—not because of the actual difficulty, but because it feels startling. Making a request through the app is like ordering from the menu; the black screen is talking directly to the chef: faster, more power, the same kitchen. You’ll get there with four wins under your belt—and a shield you learn to use before your first command.
↓ role to study
Rule for this lesson: learn to defend yourself BEFORE you attack. The shield is the procedure from lesson 0, now official: if a strange message appears on the black screen, select the line, copy it, paste it into AI, and ask, “this happened — what should I do?” Thirty seconds, and you have an answer.
With the shield equipped, the worst-case scenario on the black screen is: you read a sentence in English and ask what it means. No one can break the computer by typing the commands in this lesson — they create and list folders, nothing more.
The accountant got stuck on the first command: she mistyped it and got a curt sentence in English. She pasted it into the AI and read, “you typed ‘pdw’; the command is ‘pwd’,” corrected it, and moved on. The fear lasted exactly one question.
Before you look at the black screen for the first time, here’s a translation: what you’re about to see is a written request counter. You type a short request, press Enter, and the kitchen responds. To open the counter: on a Mac, search for "Terminal" using Spotlight; on Windows, find "PowerShell" in the Start menu—both already come with your computer.
Six requests cover your routine: pwd ("where am I?"), ls ("what’s here?"), cd name ("go into the folder"), mkdir nome ("create a folder"), open your text editor, and — the sixth and most important — ask the AI for the rest. Seriously: the sixth command is a question.
The doctor pasted into the black screen the answer the AI gave him to “how do I see what’s in my documents folder?” It worked on the first try. He didn’t memorize anything—memorizing is its job.
Common mistake
Trying to memorize commands the way you memorize verb conjugations. This happens because that's how people learned computing in 1998—but today's goal is to understand what you're ASKING FOR, not memorize the spelling. The AI does the memorizing; you lead. Got a letter wrong? The step 01 shield takes care of it in thirty seconds.
the right request
pwd
/Users/voce
common error (letter swapped)
pdw
command not found: pdw ← "I don't recognize that request": one letter is out of place. Copy it, paste it into AI, and fix it.
In module 4, your knowledge base lived in a notes app. Now it gets a permanent address: a parent folder called my-base, with the eight sections inside. The request is the same one you’d make to a secretary: "create a folder with these sections" — only in writing, and done in two seconds.
One detail that prevents the most common misstep: folder names can’t contain spaces — use hyphens. “my folder” with a space becomes TWO folders; “my-folder” with a hyphen becomes one. And remember: when the black screen gives no response, that means success — it only speaks when something went wrong or when you ask it a question.
The lawyer created the workspace in three requests and checked with “ls”: all eight folders listed on screen, in order. Then she saved the “meu-trabalho.md” file inside contexto using the same old Notepad — a new editor wasn’t a prerequisite, and still isn’t.
the right request (hyphen)
mkdir minha-base
(sem resposta — silêncio é sucesso: a pasta existe)
common error (space in the name)
mkdir minha base ls
base minha ← the space created TWO folders. Ask AI for help deleting them, then redo it with a hyphen.
The architect knows the process by heart: sheet R00, R01, R02 — you never draw over the only copy. The computer has this for any folder: it’s called Git, and for you it’s just this: a dated version stamp you can roll back to if something gets worse.
You won’t memorize its commands—you’ll ask: “I’m in the my-workspace folder; give me the commands to save a first version labeled ‘first version’.” The AI gives you two lines; you paste them; the history begins. What matters to understand: before changing an important record, stamp it.
The architect stamped the knowledge base before rewriting the inspection protocol. The new version was worse — and going back to the previous one cost a single sentence to the AI: “restore the folder to yesterday’s stamp.” No work lost, no cold sweat.
The version stamp, without memorizing anything
the stamp (AI wrote it, you pasted it)
git add . git commit -m "primeira versão"
[main a1b2c3] primeira versão 9 files changed ← carimbado: 9 arquivos guardados nesse ponto de restauração
common error (skipping the first line)
git commit -m "primeira versão"
no changes added to commit ← "nada foi colocado na caixa antes de lacrar": faltou a linha do add. Cole as DUAS linhas, em ordem.
So far, AI has been talking to you "from the outside"—it only knows what you paste in. Once your knowledge base lives in a folder, there’s a better way: talk to AI INSIDE the folder, where it can read your notes on its own, with your permission. It’s the difference between describing your office over the phone and having someone come in.
That’s the topic of the next module — for today, just know that the home you’ve just built is exactly what makes this possible. An organized folder-based knowledge base is a prerequisite; the rest is an invitation.
When the doctor opened the first conversation inside minha-base, he asked, “what do you know about my work?” — and AI replied by citing the meu-trabalho.md file, the protocol, and the two decisions he had written. Nothing pasted in; everything read from the workspace.
Test yourself
You want the AI to review your protocol file using your recorded decisions too. Which conversation is best for that?
Practice Now 0/4 done
Leave with the knowledge base on your computer, versioned — in ~15 minutes.
These commands only create and list folders — nothing gets deleted, and your computer is completely safe. Stuck at any step, or prefer not to work on your computer today? The browser route (claude.ai/code) delivers the same result without installing anything new on your computer — it’s the official route, not a backup plan.
pwd ls mkdir minha-base cd minha-base mkdir contexto processos decisoes prompts agentes referencias projetos clientes ls
You just made it through the black screen: your workspace has an address, eight folders, and a restore point. Yesterday’s fear has turned into six written requests.
Summary
Liberal Professional Training · Module 6
By the end of this lesson, you’ll have an AI-built page of instructions for your practice or office, created under your direction—checked against your completion criteria, without writing a line of code.
You’ve never laid a brick—and yet you’ve directed renovations: plan in hand, giving instructions, checking the wall. Building with AI is exactly that kind of work. What changes today is that the construction crew works inside your folder, charges nothing per hour, and asks your permission before putting up each wall.
↓ role to study
So far, AI has responded in text, and you’ve copied it wherever you needed. The development agent cuts out the gofer: it works INSIDE your project folder—reads the notes, creates files, builds pages—and always asks your permission before each change.
"Development" sounds intimidating, but the honest name would be "building": you describe what you want to exist, it proposes, you approve, it executes. Permission is the safety key — nothing changes in the folder without you saying yes.
The lawyer wants an FAQ page for the firm — those ten questions every new client repeats on the phone. She doesn’t know how to build pages. She doesn’t need to: she needs to know how to describe the page and recognize when it’s good. She’s been doing that for twenty years.
The tool for this lesson is Claude Code — the version of AI that works inside folders. There are two ways to use it: through the black screen (in the minha-base folder, the opening request is one word: claude) or through the browser at claude.ai/code, pointing to your project — the same route without the black screen from module 5.
A well-crafted first request always has three parts: context ("read the my-work.md file first"), objective ("create a page with the 5 frequently asked questions") and definition of done ("it’s ready when it opens in the browser and every answer sounds like me"). Five lines are enough.
The doctor opened the project and asked: "read my context; create página-orientações.html with what the patient needs to know before the first appointment; ready to open in a browser, with every technical term explained". Four minutes later, they were reviewing the result.
Common mistake
The novel-length request: twenty lines describing the whole dream at once. This happens because it seems thorough—but a long request dilutes the goal, and AI delivers a little of everything, none of it complete. A good request has a goal, input, and done criteria in five lines. Ask for the rest in the next turn, one layer at a time.
opening the project in the folder
cd minha-base claude
Welcome. I’m reading your folder… › what are we doing today? ← the AI can now see your files — your request can refer to any of them
Every serious project has a project brief — the document that says, “this is the standard for this build.” In your project, that brief is an instruction file in the folder’s root (the usual name is README, “read me”): who you are, what this project is, and what the AI should always respect.
Without it, every new conversation starts the explanation over from scratch — and the building ends up with a floor in every style. With it, any conversation, today or next month, builds to the same standard.
The architect recognized the mechanism right away—it’s her project brief, pointed inward. The project’s readme says: “residential architecture firm; technical yet warm tone; never promise a construction timeline; every page follows the firm’s identity.” The AI read it—and stopped inventing styles.
The nonnegotiable rule for this lesson: never accept without testing. AI builds quickly and makes mistakes confidently — the perfect pair for a professional who checks methodically. The cycle has three beats: ask (five lines), check (against the done criteria, always opening the actual result), adjust (pointing out the specific deviation).
Notice that this is the module 3 dry run, now applied to a piece of work: you don’t reread the intent; you open the page and click. What meets the criteria stays; what misses them goes back with precise instructions.
The accountant asked for a table of tax deadlines for her page. It looked great—with one wrong deadline. She didn’t argue with the tool: she pointed out, “the deadline in row 3 is wrong; check with me before publishing dates,” and added the rule “tax dates: always mark [TO VERIFY]” to the readme. The error became policy.
The cycle, with every request
It will happen: in some long conversations, the AI starts getting worse with each response, changing things that were already good. It’s not you—it’s a limitation of the AI, and there are three ways out, from the lightest to the heaviest: narrow the request (half the size, a single goal), give a concrete example (show an excerpt done the right way), or restart the conversation (the work remains; only the conversation went sour).
The module 5 version stamp is your safety net at this point: if the work itself got worse, go back to the stamp and start again from the good point. Nothing is lost except a few minutes.
Sônia, an accountant, experienced the full case: on her sixth attempt to adjust the page, each correction broke another part. She stopped, restarted the conversation with a request three times smaller — “just the header, in this example, nothing else” — and pasted a section from her old site. The page was done in two rounds. What she learned: you can’t sweeten a sour conversation; you replace it.
Practice Now 0/4 done
Leave with the FAQ page (or pre-appointment instructions) built and reviewed, in ~15 minutes.
The agent only works inside the project folder and asks permission before each change—and module 5’s version stamp guarantees a way back. For the content, use questions your audience commonly asks: no real cases or names go on the page.
Leia antes: contexto/meu-trabalho.md. OBJETIVO: crie o arquivo pagina-orientacoes.html com as 5 perguntas que meus <clientes/pacientes> mais fazem, respondidas no meu tom. INSUMO: as 5 perguntas são: <liste-as aqui>. CRITÉRIO DE PRONTO: abre no navegador; título com o nome do <escritório/consultório>; cada resposta em linguagem leiga, sem promessa de resultado e sem dado de pessoa real.
You just directed a digital build from start to finish: you asked for a plan, checked against clear criteria, and made precise adjustments. That’s how you build without laying a brick.
Summary
Liberal Professional Training · Module 7 · free tier, no card required
By the end of this lesson, you’ll have connected an agent to your complete knowledge base—it will answer questions with citations from YOUR files—and you’ll know how to estimate the cost of any connection before using it.
Your agent is good, but it works locked in a room: it knows a lot and can’t reach anything — not even your calendar or your twenty years of records. Today, it gets outlets. And since you’re paying the electric bill, the cost of each connection appears BEFORE the first use, not on the invoice.
↓ role to study
Imagine hiring the best consultant in the country and locking them in a room with no phone, no files, no windows. That's your agent today: brilliant and isolated — everything goes in and out through your hands, copied and pasted. It works, but you're the office runner.
Tools are what change the category: with access to your calendar, your knowledge base, a search — always with your permissions — the agent stops just answering and starts taking action.
The lawyer feels the limit every Monday: the agent triages new cases perfectly, but she still checks the calendar by hand, case by case, to suggest a meeting time. The consultant knows how; it just can’t reach the calendar.
Twenty years of knowledge shouldn’t depend on you being in the room.
Every serious digital service has a API — in plain language, a standard outlet. The whole mechanism fits in one line: you send a request in the agreed format, the service processes it, and the response comes back in the same format. Request, service, response. That’s all.
The outlet exists so programs can use services without opening screens—just as a wall outlet works for any device, from any brand, with the right plug.
The accountant has been using adapters for years without calling them that: when her system looks up a company ID and the company’s data appears filled in, it sent a request to the tax authority’s adapter and got a response in the agreed format. No screen was opened; no human typed anything.
The outlet has one direction: you make a request when you want. The doorbell (the technical name is webhook) reverses this: when something happens in the real world — a form is answered, a payment is confirmed, a message is received — the service rings your agent’s doorbell on its own.
It’s the difference between calling the clerk’s office every hour and having the clerk’s office call you when the document is ready. The second option gives your mornings back.
In the office, the patient fills out the pre-visit form at 9 p.m.—and the doorbell rings: the intake agent receives the answers and has the prep ready before coffee. The doctor didn’t check anything; they were notified.
Before · you keep watch
Check incoming forms 5 times a day, “just to see if they came in” — most checks turn up nothing.
After · the doorbell rings
When you arrive, the agent has been notified and has already prepared everything; you find the work ready at your first glance of the day.
Balance: zero empty checks — monitoring becomes an alert, and your attention returns to the person in front of you.
The piece that brings it all together is missing: the MCP — the universal adapter. With it, AI uses the outlets ON ITS OWN, within your permissions: checks your calendar, searches a trusted site, reads your files—and you see every requested use on screen, approving it as you did in module 6.
And here’s the climax of this module for anyone who makes a living from expertise: the most valuable tool your agent can get isn’t a calendar or search — it’s YOUR KNOWLEDGE BASE, fully connected. It’s called a “project with files”: the agent consults your entries before answering and cites which file it used. Your twenty years of experience, answering on your behalf.
The lawyer connected the knowledge base and asked, as a test: “what is my position on rent reviews in older contracts?” The answer came in three paragraphs — citing “decisões/revisao-aluguel.md”. Nothing made up: her position, with the file’s location.
Test yourself
"When the client pays, the agent should be notified to issue the receipt." — what component handles the NOTIFICATION?
Every outlet has a meter. Some connections are free up to a certain volume (the “free tier” — and ALL the exercises in this course use it, with no credit card); others charge per use, a few cents per request. The professional’s rule: estimate first, use later — ask the AI itself, "how much would it cost to do this 100 times a month?" and decide with the number in front of you.
Security follows the same logic as always, now with a wider reach: a connected agent can access whatever you connect — so confidential data should not go into any connection, and permissions are granted per tool, never “all at once.”
The architect wanted an agent that could search for material prices in real time. Before connecting it, she asked about the cost: the free tier covered 200 searches per month—three times her usage. She connected with her eyes open; the bill never surprised her.
Common mistake
Connect everything at once on the first weekend: calendar, email, payments, search. This happens when you get excited about what you've discovered—but every new connection is one more thing to check when something seems off. One working, verified integration is worth more than five half-finished ones. Connect the one that hurts most; stabilize it; only then add the next.
Practice Now 0/4 done
Leave with the knowledge base connected to the agent and 3 answers citing the right files, in ~15 minutes.
Everything here runs on the free tier — no credit card, no hidden charges. The files have already been anonymized since module 4, so nothing confidential is sent anywhere. And if a response mentions something that is NOT in your files, you’ve caught generic memory in the act — that’s what the step 4 test question is for.
You just heard your own experience answer, with a reference to the file and everything. Starting today, your twenty years of experience can serve clients even when you’re not in the room.
Summary
Liberal Professional Training · Module 8 · the final setup
By the end of this lesson, you’ll have built your AI Operating System using 100% real material—with ten folders, each containing one of your deliverables—and scheduled your 20-minute weekly routine.
So far, you’ve hired employees (agents), set up the files (knowledge base), and plugged things in (connections). One thing is missing: headquarters—the place where it all becomes ONE operation, with an org chart, rules, and memory. Without headquarters, each piece works and grows outdated on its own; with it, the whole operation improves every week.
↓ role to study
The AI Operating System is a structure of ten numbered folders: 01-identity (what the operation is), 02-context, 03-rules, 04-agents, 05-prompts, 06-tools, 07-processes, 08-projects, 09-memory e 10-audit. The numbering isn’t for show: it’s the reading order — whoever arrives (you six months from now, a partner, a new AI) understands the operation by reading in sequence.
If the dividers in module 4 were one person’s file, this is the headquarters of an operation: identity first, rules near the top, memory and auditing closing the loop.
The doctor named his in folder 01: “Consultório Dr. A. — operação com IA: triagem, orientações e pesquisa, sob revisão médica em 100% dos casos”. One sentence—and any new AI conversation already knows where it’s stepping.
The good news about the final module: you’ll create almost nothing today—you’ll organize what you have. The task map from module 1 goes into 02-contexto. The library of service orders goes into 05-prompts. The agent documents go into 04-agentes. The process and decision files go into 07 and 02/08. The page from module 6 goes into 08-projetos. The connections from module 7 are documented in 06-ferramentas.
The completion criteria are countable: at least one REAL file in each key folder. Real means written by you, about your work, in this course.
The lawyer finished organizing everything in twelve minutes — it all already existed. The only folder that started empty was 09-memoria; she opened a “diario.md” file and wrote the first line: “system set up today; first review scheduled for Friday.”
Common mistake
The showcase OS: ten beautiful, numbered, organized folders—and empty. This happens because organizing things feels like progress without the risk of writing. But a folder without a real file isn't ready, and an empty system dies in the second week. If a key folder is missing material, the fix isn't to make content up: go back to the corresponding module and do the practice you left unfinished.
The 03-regras folder is the shortest and most important: what this operation must never do. You already know the core section—what never goes into the AI: identifying information, professional confidentiality, client documents without anonymization. Now it gets a permanent home and a new section: the regulations for your profession.
For the lawyer, the section cites the confidentiality requirement in the OAB Statute and states: "only anonymized legal filings may be entered; AI is never cited as a source of law." For the doctor, CFM and LGPD: "AI does not take part in clinical decisions; administrative and communication use only, subject to review." Once written down, this stops being an intention and becomes a policy that every AI in the practice reads.
When the attorney’s intern started using the firm’s agents, she didn’t need a lecture: her first conversation cited the rules in folder 03 — the house rules apply to humans and agents, and they’re written in one place.
09-memory is the operation’s diary: decisions made (“I changed the triage agent to ask X”), errors found, ideas for the following week. 10-audit is the accountability trail: every source check you did is recorded — what, where you checked, when.
It may seem bureaucratic; it's the opposite. The day someone asks, "Do you use AI? And who makes sure it's used responsibly?" the answer won't be a speech—it will be a folder with dates.
Dr. Marta, the lawyer from the tabletop test, experienced this day three months later: a corporate client asked in a meeting how the firm monitored AI use. She opened 10-auditoria on screen: forty-two case-law checks recorded, each with a date and a verified source. The contract was renewed that same week — and the folder became a sales argument.
A system without maintenance dies — not suddenly, but from neglect. The antidote takes 20 minutes a week, on the same day, at a scheduled time in your calendar like any professional commitment. Three steps, always the same.
Review the memory (what did the system get wrong or right this week?), update projects (put changes in the files), tighten ONE prompt (the week’s worst performer gets a new rule). Twenty minutes — less than a useless meeting — and on Monday the system is smarter than when it left on Friday.
The accountant scheduled hers for Friday at 5:40 p.m., her last appointment of the week. By the third ritual, she noticed the pattern: every prompt refinement saved rework the following week. She never skipped her ritual again—not out of discipline, but for profit.
The weekly ritual, always the same
Practice Now 0/4 done
Leave with the Operating System set up — ≥1 real file in each key folder — and the ritual scheduled, in ~15 minutes.
It’s a furniture move, not a content change: you only create folders and move files that already exist—nothing gets lost, and the version stamp in module 5 preserves the previous state. Prefer the browser? The claude.ai/code path builds the same structure without a black screen.
cd minha-base mkdir 01-identidade 02-contexto 03-regras 04-agentes 05-prompts 06-ferramentas 07-processos 08-projetos 09-memoria 10-auditoria ls
You just built your AI Operating System—with material you wrote yourself, rules for your profession, and a scheduled routine. The training ends here; the operation starts now.
Summary