💬 The chatbot only responds
New here? One chatbot is a program that talks: you type a question, and it returns a text answer. That’s all. Think of an information desk—the attendant explains where the room is but won’t go open the door for you.
Tools like ChatGPT in default mode work this way: great for explaining, summarizing, and translating—but the result stays on the screen. If you want something to actually happen on your computer (a file created, an email sent, a site published), a chatbot on its own can't do it.
💡 Main Concept
Chatbot = text input, text output. Nothing changes outside the conversation window.
- •You ask "how do I make a sales summary?" and it explains the steps.
- •You still have to open the spreadsheet, copy, paste, and calculate by hand.
🔎 See inside
When you chat with a pure chatbot, nothing changes on your computer: the text it generates lives only inside the chat window, on a remote server. No files are created in your folder, no commands are run, and no emails are sent. That's why, after a good answer, you still have to manually copy, paste, and apply it somewhere real — the chatbot gave you the map, but didn't walk the route.
🛠️ The agent picks up the tool and does the work
One coding agent it’s different: it has access to tools — things it does itself, such as opening a file, running a command in the terminal (the text box where you give the computer direct instructions—more on that in module 0.2), or looking up information online. It decides which tool to use, uses it, checks the result, and adjusts if something went wrong.
Caption: the request comes in on the left, the agent chooses the right tool from the “drawer,” runs it, and returns something already done — not just an explanation.
🔍 Inside
- Tool: a concrete action the agent can trigger on its own — read, write, run, search.
- Cycle: think → choose a tool → use it → check the result → repeat until finished.
⚠️ Common beginner error
The first time the agent asks permission to run a command or modify a file, many people get startled and cancel everything, thinking something broke. Nothing broke — that pause is the agent checking with you before acting, exactly the behavior we want (see module 0.6). The real mistake is the opposite: approving everything in "automatic mode" without understanding what's being requested. Read the request before approving — it's quick and prevents rework.
🏢 The employee analogy
Think of Claude Code as a new employee, highly capable and very fast, but still unfamiliar with how things are done here. You give it a task (“organize these reports”), it decides how to do it (opens the folder, reads the files, creates the spreadsheet), does the work, and comes back to show you what it did. At first, you supervise closely; over time, you trust it more—just as you would a real employee.
✓ Well-guided employee
- ✓Receives clear instructions and the context it needs (the module 0.4 CLAUDE.md)
- ✓Shows the plan before acting on large tasks
- ✓Notify you when something differs from what was expected
✗ Employee without supervision
- ✗Receives a vague order (“just fix this”) and guesses the rest
- ✗Makes major changes without telling you first
- ✗No one reviews the result before publishing
💡 Practical Tip
Treat the first few weeks as a trial period: assign small tasks, review what it does, and only then trust it with bigger tasks. That’s exactly what wraps up the track in module 0.6.
The trust curve, in three phases
-
1
Week 1 — nearby: small, isolated tasks (renaming files, summarizing text), you read everything before accepting.
-
2
Weeks 2–4 — light supervision: tasks with more steps (organizing an entire folder), you review the plan but not every line.
-
3
Next—real delegation: routines that run on their own at fixed times (track 6), with permissions already calibrated.
🖥️ What Claude Code is, in concrete terms
New here? One AI model is the "brain" — the program trained to understand language and decide what to do (it's the same kind of technology behind ChatGPT). The Claude Code is a program that runs in your terminal and connects this brain to real tools on your computer: it reads and writes files, runs commands, and only does so inside the permission that you grant—it’s up to you how much it can do on its own.
You open a terminal in a folder
The program now “sees” only what’s inside that folder—the project (module 0.3).
You write what you want in Portuguese
No special syntax—just a normal sentence, the way you’d ask a colleague.
It takes action and shows the result
Within the permission limits you defined (covered in depth in module 0.6).
Caption: the folder you open is the agent’s boundary. Web searches happen only if a specific search tool is used — and even then, only within what you allow.
🔎 Under the hood: what goes on the internet and what stays on your PC
Your files stay on your computer — opening, reading, and writing in the project folder doesn’t send anything outside. What travels to the model provider’s servers is the conversation text: your instruction and the excerpts from files the agent needs to "look at" to respond. This is how the brain (the AI model) processes the request and decides what to do — the heavy processing doesn't happen on your laptop.
💰 The cost of a session in practice
New here? Token it’s the unit that measures “how much text” was processed—roughly, a piece of a word. Each message you send and each response from the agent uses tokens, and that determines the cost.
📊 Teaching estimate (not a fixed bill)
- Small task (renaming files, writing a short text): a few cents.
- Typical automation session (like the Builds in this course): in the range of R$ 0.50 to R$ 3.00.
- Long and complex session (lots of back and forth, lots of code): it can exceed R$ 10.
Approximate numbers—they vary by plan and model. The point here isn’t to memorize prices, but to stop worrying that “I’ll spend a fortune without knowing.”
⚠️ Attention
Sessions that get "stuck going around in circles" (the agent repeatedly trying and failing) use more tokens than usual. If you notice this, stop, review the request, and start a new session — it's cheaper than persisting.
✓ Conscious use
- ✓Asks for one thing at a time and checks the result before moving on
- ✓Ends the session when the task is done, instead of leaving it open "just in case"
✗ Wasted usage
- ✗Asks for the same thing in different ways, hoping something will "click"
- ✗Pastes entire giant files when only a small excerpt was needed
⚠️ Common confusion: fixed plan × pay-per-use
Many people mix up two different things: there are subscription plans (a fixed monthly amount, with a usage limit included) and there’s billing based on token usage (you pay only for what you use, with no monthly fee). The number shown above is token usage — how much that turns into in actual cost depends on which of the two models you chose. There’s no single answer; what matters here is understanding the unit, not memorizing a price.
🗺️ The map of Track 0
With the mental model in place, the rest of the track is 100% hands-on: install without fear (0.2), understand the project folder (0.3), write your first CLAUDE.md (0.4), learn plan mode (0.5), and wrap up with the security rules that apply to the entire course (0.6).
🎯 Main Concept
- •0.2 installs → 0.3 organizes → 0.4 guides → 0.5 reviews → 0.6 protects.
- •Each module builds on the previous one — don’t skip ahead.
💡 Practical Tip
If you remember only one sentence from this module, remember this: a coding agent is an employee who does, not a chatbot that explains. Every design decision in the rest of the course — permissions, CLAUDE.md, plan mode — exists to make this "doing" safe and predictable. When a new concept seems confusing later on, come back here and ask: does this help the agent act better, or more safely? Almost always, it's one of the two.
Quick check (optional): What is the key difference between a chatbot and a coding agent?