Single path · 6 short lessons
Over six sessions of 15 to 21 minutes, you’ll learn to explain the four layers of an AI system, why the terminal with an agent is the one that actually does the work, how to write rules so it works your way, and where all of this lives in the project folders.
2-minute video: the four layers, from chat to harness, and where everything lives. Watch it before lesson 1 or after lesson 2. Vertical version for Shorts/Reels.
Lessons
Leave knowing how to say, with examples from your work, where the browser chatbot helps and exactly where it gets stuck.
Leave able to sketch the chat, editor, terminal, and harness layers on a napkin, and explain why the terminal has hands.
Leave able to recognize, in a real agent session, what it read, what it ran, when it asked permission, and how it checked its work.
Leave with the first version of your agent’s standing instructions written, and the seven parts of the harness named.
Leave with five tasks from your week sorted by layer, and at least one ready to go to the terminal.
Leave with your project folder structure mapped out: one folder per project, the .claude drawer, AGENTS.md open, the knowledge wiki, and the OKF format.
Lesson 1 · Fundamentals
By the end of this lesson, you’ll be able to point to the exact moment in a real work task when the browser chatbot stops helping and starts costing you time.
You already pay for AI and still spend the afternoon copying and pasting. You paste in a report, read the response, paste in another, fix it by hand, and the next day the conversation remembers nothing. The frustration is real, and it's not your fault: it's the limit of the layer you're using.
↓ role to study
One web chatbot it's the AI you open in your browser and chat with. It's the first layer of any AI system, and it's good at what it does: thinking things through with you, drafting, summarizing, translating, and giving a second opinion in seconds. No preparation is needed. Open it, type, get a response.
Helena, operations director at a packaging manufacturer, uses it this way every day: she pastes the production meeting minutes and asks for the three risks no one said out loud. In twenty seconds, she gets an analysis that would be worth an hour of a consultant’s time. For this kind of task—a conversation with a single text—layer 1 is unbeatable.
before, without AI
Helena read the entire meeting minutes at night, marked them up by hand, and wrote the summary in the morning. An hour and a half.
with web chat
Paste the meeting notes, ask for the risks, and review the answer against what she knows about the factory. Twelve minutes.
Takeaway: about 80 minutes per meeting, for tasks involving a single text. This is where chat shines.
The consultant on the phone has a problem: they’re not in your office. They can’t see the folders on your desk. All they know is what you read to them. It’s the same with web chat. It can’t see your computer, so each file has to be copied and pasted, and the response copied back. You become the messenger between the AI and your own work.
Rafael, an architect, felt this during proposal week. He had fourteen old estimates in a folder and wanted a table comparing the price per square meter. In chat, he pasted them one by one. By the seventh proposal, the conversation had "forgotten" the first one. By the tenth, he gave up and made the spreadsheet by hand. The problem wasn’t the AI’s intelligence. It was that it couldn’t access the folder.
Notice the pattern: the task with one text works; the task with several files break. That's the ceiling of layer 1, and it shows up exactly in the tasks that take up the most of your time.
Test yourself
Camila, a dentist, wants a summary of the points to watch for in 30 patient history forms stored in separate files on the clinic computer. What happens if she uses only web chat?
The second ceiling is memory. You explain to chat how your company writes, which clients are priorities, and what report format you like. Tomorrow, a new conversation, and it’s all forgotten. It’s like a consultant who asks you to tell your company’s story all over again on every call.
Denise, a marketing consultant, keeps a document with "the rules for my writing" and pastes it at the start of every conversation. It works, but it’s a workaround: she’s the AI’s memory. When she forgets to paste it, the text comes out in the wrong tone, and she only notices after sending it to the client.
Remember these two limits: access and memory. The next layers exist to overcome these two limits.
Chat is great for talking about a text. It gets stuck when the work lives across several files and repeats every week.
Practice now 0/4 done
In about 10 minutes, take a real task from your week that involves more than one file and note where the web chat stops helping. You’ll leave with the "sticking point" written down.
Nothing here changes anything on your computer or in your conversations. This is just observation. If the chosen task doesn’t get stuck, great: make a note of that too; it’s a task that belongs to layer 1.
You just pinpointed the exact limit of layer 1 using one of your own tasks. The next lessons will address that limit.
Summary
Lesson 2 · Fundamentals
By the end of this lesson, you’ll be able to sketch the four layers of an AI system on a napkin (chat, editor, terminal, harness) and state the advantage and the point where each one gets stuck.
A new AI tool appears every month, and someone on your team says, "this is the one." Without a map, you buy into the hype. With a map of the layers, you can look at any tool and know in ten seconds which floor it belongs on and what it can actually do.
↓ role to study
The second layer is the AI that lives inside an code editor. If chat is a consultant on the phone, this is an assistant sitting beside you, looking at the same screen. It can see the file that's open and suggest changes right there, without you copying anything.
It's a big leap for someone who codes. For Rafael, an architect, it's less useful: the editor was made for code, not proposals and blueprints. And even for a programmer, the assistant only sees what's on the screen. If the answer depends on twenty other files in the folder, it guesses. And the person still has to run, test, and check everything.
Advantage: it sees the open file and edits it in place. Limit: it can only see the screen, and you still do the typing.
O terminal it's a text window that exists on every computer. You type a request, the computer runs it and responds in writing. No menus, no buttons. It seems old-fashioned, and it is: it's the most direct way to talk to a machine, older than any pretty screen.
Why does this matter? Because a agent placed at this desk can do everything a person would do by clicking, only in writing and faster. It opens Rafael’s entire folder, reads the fourteen proposals, builds the table, checks the math, and lets you know when it’s done. You don’t have to copy anything. It’s the employee with the office keys, not the consultant on the phone.
web chat
Rafael pastes in 14 proposals, one by one. He loses the first ones. At the tenth, he gives up and does it in the spreadsheet.
terminal agent
Rafael writes: “read the proposals in this folder and put together a table of prices per square meter.” The agent reads all 14, builds the table, checks it, and delivers it.
Takeaway: from a wasted afternoon to 4 minutes, and the table is checked against the original files.
An employee with the office keys, no manual, and no limits is a risk. That’s why there’s a fourth layer. The harness it's everything around the agent that lets it work your way: the “how we do things here” manual, the badge that says what it can open on its own and what it needs to ask about, the recipes for repeated tasks, automatic reminders, and the doors to your calendar and email.
Camila, a dentist, doesn’t want an agent that messages patients on its own. Her harness says: "patient messages require my approval; it can edit the schedule and inventory." This isn’t programming. It’s a sheet of rules the agent reads whenever it starts working. Lesson 4 takes the harness apart piece by piece.
Test yourself
Helena wants the agent to prepare the weekly production report on its own, but never send anything to the leadership team without her reading it first. Which layer does that rule belong in?
Every useful AI task follows a cycle: understand the request, plan, do the work, review the result, make corrections, deliver. Web chat stops at "plan": it tells you what to do, and you do it. The editor handles part of the task in the open file, and you review it. The agent in the terminal completes the whole cycle because it can access files, run commands, and inspect what it produced.
Denise, a marketing consultant, asked the agent: "generate the monthly reports for the three clients using the data in this folder and last month’s template." The agent read the data, put together all three reports, noticed that one client was missing a spreadsheet, let her know, and delivered the other two ready to go. In chat, she would have spent an hour pasting data and discovered the missing spreadsheet after sending the report.
That’s why the terminal is the best layer. It isn’t the prettiest or the easiest to open the first time. It’s the only one where AI has hands and eyes at the same time: it does the work, checks it, and fixes it without you in the middle.
The chat responds. The terminal executes. The harness ensures it executes your way.
Practice now 0/3 done
In about 8 minutes, read the case, decide which layer each task belongs to, and compare your answers with the key. You’ll leave with the layer map applied to a real week.
It's analysis on paper: no tool is opened. If your answer differs from the answer key, read the explanation: for decisions about the layer, the difference usually comes down to “how many files” and “whether it repeats or not.”
The case. Helena, operations director, listed four things for the week: (1) reply to a delicate email from a late supplier; (2) cross-check the 12 shift spreadsheets to find unjustified machine downtime; (3) every Monday, prepare the previous week’s production report using the same template and the files generated by the factory; (4) the process engineer wants a quick adjustment to a calculation file open on his screen.
(1) Web chat. One text, one answer, tone matters: this is the best use of layer 1. (2) Terminal. Twelve files, cross-checking, and verification: only an agent with access to the folder can close the loop. (3) Terminal + harness. It's a task involving several files that repeats every week using the same format: the agent runs it, and the model and the rule “don't send it without me reading it” are written into the harness. (4) Editor. One file, open on the programmer’s screen: layer 2 gets it done right there.
You just applied the layer map to a week of work, task by task.
Summary
Lesson 3 · Tool
By the end of this lesson, you’ll be able to read a real agent session in the terminal and point out, line by line, where it read files, where it ran commands, where it asked for permission, and where it checked the result.
The first time someone opens a terminal, the black screen is intimidating, and they close it. That puts off exactly the people who would benefit most from it. This lesson takes the fear out of it by showing what appears on the screen and what each line means, so you know what you’re looking at when someone on your team, or you, sits down there.
↓ role to study
Everyone’s computer has a program called Terminal (on Mac) or Windows Terminal. It opens an almost empty window with a blinking cursor. That’s all. Nothing has been activated or changed. It’s a blank page waiting for a written request.
Rafael opened his for the first time expecting “programmer stuff.” He saw the cursor and typed the command his nephew had suggested, without really understanding it. The computer responded with a list of his folders. The black screen is only scary until it responds for the first time.
What you’ll see next is a minimal recipe: a one-word request and the computer’s response. You don’t need to memorize anything, just recognize the question-and-answer format.
request and response
$ ls
Propostas Plantas Contratos Fotos-obra
With the terminal open in the right folder, one word calls the agent (Claude Code in this example). From there, you write your request in everyday Portuguese, as you would to an employee. The difference from chat is that the agent is inside the folder: it can read the files without you having to paste anything.
Denise, a marketing consultant, does the monthly report this way: she opens the client’s folder, calls the agent, and writes "put together the September report using the August template and the data in the spreadsheets in this folder." The agent lists what it plans to do, reads the files, and gets started.
Common mistake
Call the agent in the wrong folder. It can only see the folder where it was opened and the folders inside it. If you opened it in the entire "Documents" folder, it will search everything and take a long time; if you opened it in an empty folder, it will say it couldn’t find anything. Before you call it, check the folder name in the computer’s response.
in the right folder
$ cd Clientes/Padaria-Nova $ claude > monte o relatório de setembro no modelo do de agosto, com as planilhas desta pasta
Vou ler: agosto-relatorio.docx, setembro-vendas.xlsx, setembro-redes.xlsx Lendo 3 arquivos… Montando relatorio-setembro.docx
in the wrong folder
$ claude > monte o relatório de setembro no modelo do de agosto
Não encontrei nenhuma planilha nem relatório de agosto nesta pasta (Documentos). Quer que eu procure nas subpastas? Isso pode demorar.
The agent reads files on its own. Before creating, changing, or deleting anything, it asks, and you answer yes or no. It’s the same logic as with a new employee: they can look things up freely, but they ask before signing anything. In lesson 4, you’ll see how to change this behavior through the harness; by default, it asks.
Camila, a dentist, likes this. When she asked the agent to "reorganize the X-ray folder by patient," it showed her the plan, said it would move 212 files, and waited for her approval. She read the plan, saw it would sort by last name instead of first name, corrected it in one sentence, and then approved.
What appears on screen is always the same format: what it wants to do, what it’s about, and the question. Here’s an example.
permission request
> reorganize a pasta Radiografias em subpastas por paciente
Plano: criar 48 subpastas (uma por paciente, pelo nome no arquivo) e mover 212 arquivos. Nenhum arquivo será apagado. Posso mover os arquivos? [s/n]
The part that sets the agent apart from chat comes at the end: after doing the work, it checks the result. Did it count the files moved? Does that match the plan? Did the report open? If something doesn’t add up, it says so instead of delivering something polished but wrong. Helena calls this “the employee who checks before bringing it to my desk.”
In the weekly production report, Helena’s agent warned: “Thursday night’s shift has no spreadsheet; I put together the report without it and marked the gap in red.” In chat, that gap would have gone unnoticed until the board meeting.
What to look for in an agent session
Practice now 0/3 done
In about 8 minutes, read the transcript below, mark where the agent read, planned, asked permission, executed, and checked, then compare your answers with the key.
This is reading; nothing is executed. If you're already comfortable, you can then open your own terminal and type just the first line of the lesson (the list command), which doesn't change anything. This practice doesn't require it.
session. Rafael wrote in the Proposals folder: "make a table with the client name, area, and price per square meter for each proposal in this folder." The screen showed, in this order: (a) "I found 14 files: proposta-alves.pdf, proposta-bento.pdf…"; (b) "I’ll extract the client, area, and amount from each one and calculate the price per m²"; (c) "Can I create the file tabela-propostas.xlsx in this folder? [y/n]"; (d) "Creating tabela-propostas.xlsx… 14 rows"; (e) "I checked: 14 proposals, 14 rows. proposta-lima.pdf doesn’t include an area; I left the cell blank and highlighted it in yellow."
(a) read · (b) planned · (c) asked for permission · (d) executed · (e) checked. If Rafael said "no" at (c), nothing would be created: the folder would remain exactly as it was, because permission comes before any changes. Also notice that step (e) pointed out a gap (the proposal without an area) instead of making up a number: that's the behavior you want from an agent, and it's something chat can't do.
You just read an entire agent session and knew what each line meant. The black screen isn't a mystery anymore.
Summary
Lesson 4 · Fundamentals
By the end of this lesson, you’ll be able to name the seven parts of the harness and write the first version of your agent’s standing instructions in Portuguese, with at least three rules and one limit.
An agent without a harness is a brilliant employee on their first day: they don’t know how your company writes, what they can sign, or who comes first. You end up repeating everything every session, which is the same problem as chat. The harness turns a generic agent into your agent, and its most important part fits on one sheet.
↓ role to study
The first part of the harness is a text sheet the agent reads every time it starts working. It’s called fixed instructions. In it, you write once what you repeat in every conversation today: how your company writes, which clients take priority, what report format you use, and what it must never do.
Denise turned the document "rules for my writing," which she used to paste into chat by hand, into this sheet. Now the agent starts out knowing: "direct tone, no slang, short paragraphs, never promise campaign results." She doesn’t have to paste anything anymore. It’s the memory between conversations that was missing in Lesson 1, solved with a sheet.
no fixed instructions
Denise explains the tone and rules in every session. When she forgets, the text comes out wrong and she finds out after sending it.
with fixed instructions
The rules are on the sheet. Every session starts by reading them. Changed a rule? Edit the sheet; it applies from then on.
Takeaway: zero repetition per session, and the wrong tone no longer reached the client.
The agent acts through tools: reading and writing files, running tasks on the computer, searching the internet, opening a website. These are its hands. Without tools, it goes back to being a chat. The third piece, permissions, is its badge: it defines what it can do on its own and what it needs to ask about. Reading is almost always allowed. Deleting, publishing, sending, paying: only with your yes, or forbidden outright.
Camila decided on the clinic agent’s badge in five lines: it can edit the schedule and inventory; it can organize the X-ray folder but never delete anything; patient messages require her approval; medical records are off-limits. It’s not a security system she bought. It’s a written list the agent follows.
Notice how the tool and permission complement each other: the tool defines what it can do; permission defines what it’s allowed to do.
A task you repeat deserves a recipe. A skill it's this: the step-by-step process for “creating the monthly report,” written once with the format, order, and checks. After that, you just say “create the monthly report” and it follows the recipe. It's a reusable sewing pattern: cut the same dress from any fabric.
One hook it's a trigger: “whenever you finish a report, check the totals against the spreadsheet” or “let me know before publishing anything.” Helena has one: “at the end of every session, list what changed and what remains pending.” She doesn't ask; it just happens.
Test yourself
Rafael wants the agent to automatically check whether the price per square meter is within the office’s range every time it puts together a proposal. Is that a skill or a hook?
The sixth part is connections to other systems. The technical term is MCP, but think of them as outlets: one for the calendar, one for email, one for the team spreadsheet. Plug it in, and the agent can see that system just like it can see the folder.
The seventh is memory and division of work. The agent can retain what it learns about your preferences between sessions, and, for large tasks, call other agents to work in parallel, like a manager assigning an audit to analysts. Helena uses this at month-end: one agent handles the shift spreadsheets, another handles maintenance spreadsheets, and the main one pulls everything together and checks it.
You don't need to connect all seven pieces on day one. Fixed instructions and permissions already change everything. The rest comes in when a task calls for it.
Put the pieces together and the cycle from lesson 2 changes in quality. The request comes in, and the agent has already read the house manual. It plans within the boundaries of its badge. If the task has a recipe, it follows the recipe. It executes with the hands you’ve authorized, using the connections you’ve enabled. It checks its work, and the end-of-session trigger lists what changed. Next time, it remembers.
For Camila, the practical difference is that the clinic’s agent acts like a receptionist who’s been there for ten years, not a temp on their first day. Same AI model. What changed was the sheet around it.
The model is the same for everyone. The harness is what makes it yours.
Practice now 0/4 done
In about 12 minutes, use your usual chatbot to draft the fixed instruction sheet for your future agent: three rules for how it should work and one limit on what it must never do. You’ll leave with the sheet ready to paste in when the agent exists.
It's just text. It doesn't turn on any agent or change anything on your computer. If the answer is generic, it's because it lacks your context: add a real example and ask again.
Você vai me ajudar a escrever as instruções fixas de um agente de IA que vai trabalhar comigo. Sou <sua profissão e contexto, ex.: dentista com clínica de 3 cadeiras>. As tarefas que ele vai fazer: <2 ou 3 tarefas reais, ex.: organizar agenda, montar relatório mensal>. Como eu gosto que as coisas sejam feitas: <tom, formato, prioridades>. O que ele nunca deve fazer sem me perguntar: <ex.: enviar mensagem a paciente, apagar arquivos>. Escreva a folha em português, com estas seções: "Quem eu sou", "Como trabalhamos aqui" (3 a 5 regras curtas), "Pode fazer sozinho", "Só com minha aprovação", "Nunca". Frases curtas, sem jargão.
You just wrote the most important piece of the harness. Once the agent exists, it's the first thing it will read.
Summary
Lesson 5 · Practice
By the end of this lesson, you’ll be able to look at five tasks from your week and say, in seconds and with a reason, which belongs in chat, which in the editor, which in the terminal, and which belongs in the harness.
Without a decision rule, the tendency is to use chat for everything, because it’s already open. Then the big tasks stay manual and AI becomes just a faster copywriter. The rule in this lesson fits in four lines and gives you back the hours lost to tasks involving multiple files.
↓ role to study
The whole course comes down to one question for each task: how many files does it involve, and do you repeat it? The answer points to the layer. The rule below is the one Helena put on her office wall, and it works just as well for Camila’s practice or Rafael’s firm.
The decision rule
Denise, a marketing consultant, applied the rule on a Monday. Her list had these tasks left: respond to a new client’s proposal (one text: chat). Put together the monthly reports for three clients using the spreadsheets in each folder (multiple files and repetitive: terminal, with the recipe in the harness). Review a client’s captions for the week (one text at a time: chat). Check that all reports sent during the quarter match the invoices (multiple files and cross-checking: terminal).
At the end of the week: two tasks stayed in chat, where they were already a good fit. Two moved to the terminal, and those were exactly the ones that took up Friday afternoon. Nothing went to the editor, because Denise doesn’t program. The rule doesn’t take chat away from anyone: it gives the terminal back the work that never should have been copied and pasted.
Once the rule becomes a habit, the pattern is always the same: chat stays in your daily routine, the terminal takes on the heavy tasks, and the harness grows by a line each week with what you’ve learned you always want done the same way. In a month, the fixed-instructions sheet already reflects how your operation works.
In your profession
Practice now 0/4 done
In about 10 minutes, list five real tasks from your upcoming week, run each one through the four-line rule, and mark at least one as a candidate for the terminal. You’ll leave with your first layer plan.
It's a list on paper or on your phone. Nothing is opened or configured. If no task ends up in the terminal, reread line 3 of the rule: almost every week, there's at least one task involving several files hidden in an “afternoon of organizing.”
You just decided, based on clear criteria, where each hour of your week should go. The circled task is your first job for the terminal.
Summary
Lesson 6 · Foundation
By the end of this lesson, you’ll be able to sketch the folder structure of one of your projects (one folder per project, the hidden rules drawer, the instruction sheet, the recipes, and the knowledge base) and explain what each part does.
In previous lessons, the harness appeared as “a sheet the agent reads.” But a sheet lives somewhere. If you don’t know where, you leave the agent loose in the entire Documents folder, lose rules between projects, and accumulate knowledge in conversations that disappear. The answer is folder organization, and it fits in one diagram.
↓ role to study
Remember lesson 3: the agent only sees the folder where it was opened and the folders inside it. That becomes an organization rule: one folder per project, all inside a parent folder called, for example, "projects". So, "projects/project1" is project 1's office. The agent opened there sees all of project 1 and nothing else. It doesn't search your computer, mix up clients, or take long.
Rafael had everything in an "Office" folder, with dozens of mixed-up projects. When he asked, "build the proposals table," the agent read proposals from three different projects. He switched to "projects/vila-nova-project," "projects/centro-project," and the problem disappeared. Each folder became an office with walls.
one folder for everything
The agent reads 40 projects to answer about one. One client’s rules leak into another’s. It takes longer and makes mistakes.
projetos/projeto1
The agent opens the right project, reads only what matters, and that project’s rules stay inside it.
Takeaway: scope defined by the folder. Switching projects means switching folders.
Each project has a folder that starts with a dot: ".claude". The dot at the beginning is a convention: folders like this are hidden from the regular listing, like a drawer of procedures that stays closed but everyone knows where it is. That’s where the project’s harness lives: the sheet of standing instructions (the CLAUDE.md file you wrote in Lesson 4), the permissions file, the "skills" folder with recipes, and the "agents" folder with subagents.
Inside "projetos/clinica," Camila has a ".claude" drawer with a sheet of rules ("never message a patient without approval") and a recipe called "fechamento-do-mes." When she opens the agent in the clinic folder, it reads the drawer before doing anything. In the new office’s folder, there’s another drawer with different rules.
What follows is the real screen from a listing that shows the drawer. Read it as a folder map, nothing more.
the hidden drawer appears with -a
$ cd projetos/clinica $ ls -a .claude
CLAUDE.md settings.json skills/ agents/ $ ls .claude/skills fechamento-do-mes/ confirmar-agenda/
The “.claude” drawer belongs to Claude Code. Other agents (Codex, Cursor, Gemini) have their own. To avoid writing the same sheet five times, the industry agreed on an open standard: a file called AGENTS.md in the project root, and a ".agents/skills" folder for the recipes. Claude Code reads both. Codex does too. One sheet, several employees.
Vercel released a tool that copies ready-made recipes into the right folder for each agent and maintains an open catalog of recipes. That’s what makes a skill portable: a little folder with a SKILL.md inside that any agent can understand. Helena approved standardization at the factory for a simple reason: if the team switches agents, the rules and recipes still apply.
Practical rule: write general rules in AGENTS.md; keep only what’s specific to that agent in ".claude". In large projects, each subfolder can have its own AGENTS.md, and the agent uses the nearest one.
Test yourself
Denise uses Claude Code today, but her new client works with Codex. Where should she write the agency’s tone rules so she doesn’t have to maintain two copies?
Rules say how to work. What’s missing is what the project knows: meetings, decisions, data, references. Andrej Karpathy, one of the best-known names in the field, proposed a simple pattern that became a reference: two folders. In "raw," you put everything that comes in (minutes, PDFs, emails, clippings), without changing it. In "wiki," the agent writes simple text pages based on the raw folder, linked to each other, one per topic, and keeps them up to date. It’s a LLM wiki.
Three operations, always the same: ingest (a new meeting record came in, and the agent updates the affected pages), to ask (you ask, and it answers by citing the pages) and review (it checks for contradictions and orphaned pages). Helena built the factory wiki in two weeks: each production meeting record goes into raw, and the "machine downtime" page is always up to date, with the date and source for each line.
The difference from chat is the same as always: the conversation disappears; the wiki stays in a folder, as text you can open and edit, and the agent in the next session finds everything there.
A wiki works for one project. When knowledge needs to circulate across teams, companies, and different agents, it helps to have an agreed-upon format. Google Cloud published one: the OKF, Open Knowledge Format. It's minimal by design: a folder of text files, each with a short header at the top listing its title, type, description, and author; an "index.md" file in each folder listing what's there; and a "log.md" with a history of what changed.
Two details matter for anyone who makes decisions. First, the header records who wrote each page: a person (“human:helena”) or an agent (“agente/versão”). That way, you can place different levels of trust in what people have confirmed. Second, the index lets the agent see what exists before opening everything, which saves time and keeps it from reading what it doesn’t need to. Rafael uses the header to mark “verified by: human:rafael” on the cost pages he checked.
You don’t need to adopt OKF on the first day. The takeaway is the same as in previous lessons: knowledge in a folder, in plain text, with an owner and a date, that people and agents read and update.
Rules in the drawer, knowledge in the cabinet, one room per project. The agent walks in and already knows where everything is.
Practice now 0/4 done
In about 12 minutes, use your usual chatbot to sketch the folder tree for the project you circled in lesson 5: the project folder, the rules drawer, and the knowledge base. You’ll leave with a drawing ready to use when you sit down at the terminal to create the folders.
It's a text-based design. No folders are created; nothing changes on your computer. If the answer includes folders you don't understand, ask “explain each folder in one sentence, as if I were organizing a closet.”
Quero organizar as pastas de um projeto em que um agente de IA vai trabalhar comigo. Projeto: <nome e o que é, ex.: relatório mensal de produção da fábrica>. Arquivos que ele vai usar: <ex.: 12 planilhas de turno por semana, atas de reunião em PDF>. Regras que ele deve ler sempre: <cole a folha de instruções fixas da aula 4, ou resuma>. Desenhe a árvore de pastas em texto, começando em projetos/<nome-do-projeto>/, com: 1) a pasta oculta .claude (com CLAUDE.md, settings.json, skills/ e agents/) e um AGENTS.md na raiz com as regras gerais; 2) uma base de conhecimento no padrão wiki (pastas raw/ e wiki/) com 3 páginas iniciais sugeridas; 3) um index.md na raiz listando cada pasta com uma frase. Explique cada pasta em uma frase, sem jargão. Diga o que eu NÃO devo colocar dentro dessa pasta.
You just designed your agent's office. Creating the folders is a one-line request to it in lesson 3.
Summary