Detailed content
🛡️ Model vs. Environment — Iron Man’s Armor
Claude Code brings AI to inside your computer, instead of just chatting with it on a website. Think of it this way: the the model is Tony Stark (the brain, the intelligence) and the the environment is the Iron Man suit — the body that lets the brain read files, run commands, and actually work with things.
The model is the intelligence; the environment (or IDE) is where that intelligence acts—reading, executing, and creating things on your machine.
Without this concept, everything feels like confusing magic. With it, you understand why the same AI runs in different places and why choosing the environment matters.
Model · environment · IDE · "one task, one window".
✓ The model (Tony Stark)
- ✓It’s the brain: it reasons, writes, and decides.
- ✓Alone, it’s brilliant, but it can’t act in the world.
- ✓You can switch models without switching environments.
✓ The environment (the armor / IDE)
- ✓Gives the model “form”: reads and creates files.
- ✓Runs commands and connects other tools.
- ✓There are several “suits of armor”—you can switch whenever you want.
💡 Don’t get stuck on the term "IDE"
IDE stands for “integrated development environment.” For now, think of an IDE simply as "the place where the assistant works". Nothing else.
💻 Installing Claude Code
Time to take AI out of the browser and put it on your machine. Installation is a prerequisite for everything: without installing, there’s no practice. An important note before you begin — Claude Code requires paid plan; the free plan doesn't include it.
Download and install the official Claude Code app and connect your account.
It’s the required starting point. Everything that follows assumes the app is installed and you’re logged in.
Installation · sign in with your Claude account · paid plan required.
Visit the official website
Open claude.com/claude-code — the official installation source.
Download for your system
Choose the Windows, macOS, or Linux version. It works on all three.
Install and open the app
Follow your system’s standard installer and open the app.
Sign in with your Claude account
Use the same account as the site and confirm that it has paid plan.
🧭 The Interface Inside
Before giving commands, get to know the screen. Knowing where each button is prevents you from getting stuck right at the start: conversation history, model selector, permission mode, usage indicator, and context window.
A tour of the screen: conversation panel, model selector, plan usage indicator, and context window.
Finding the controls is what gets you past "I don't even know where to start."
Conversation panel · model selector · usage indicator · context window · microphone · attach files.
🔎 Where to look first
- •Project history usually sits on one side of the screen.
- •The model selector and permission mode are near the typing box.
- •The plan and context usage indicator appears next to the message field.
🔐 Permission modes
The permission mode controls how much freedom you give the assistant — how much it can do on its own without asking you. It’s your safety lock: too much freedom too soon is a recipe for surprises.
How much the assistant acts on its own, from most cautious to boldest.
It’s the difference between experimenting with peace of mind and accidentally messing up important files.
Four modes: ask before acting · accept edits · plan mode · ignore permissions.
✓ Safe to get started
- ✓Ask first: confirms every edit and every command. Ideal for unfamiliar projects.
- ✓Accept edits: edits on its own, but still asks before running sensitive commands. Recommended for getting started.
- ✓Plan mode: doesn't execute; it thinks and proposes a plan first.
✗ Dangerous for beginners
- ✗Ignore permissions: doesn't ask anything. Powerful and dangerous.
- ✗Use only in an isolated environment, when you know exactly what you're doing.
- ✗Never on a "real" computer with important data.
💡 Pro tip
Start at "accept edits". It's safe enough to build fluency without having to approve every step, while still protecting you on the most sensitive commands.
🧠 Context Window and “Rot”
Imagine a brilliant person with an extremely high IQ, but with amnesia. The context window is how much of the conversation the model can remember accurately. The longer the conversation, the worse it remembers — a phenomenon called "context rot" (context rot). Hence the golden rule: one task, one window.
The limit on how much of the conversation the model remembers well—and why long conversations reduce quality.
Managing context separates those who get reliable results from those who receive “confidently wrong” answers.
Context window · tokens · context rot · hallucination · "one task, one window".
📊 In practice: watch out for hallucinations
- •The model can hallucinate (make things up with conviction) between ~1% and ~10% of the time.
- •It performs worse in long conversations—the fuller the window, the greater the risk.
- •Protect your context: start a new conversation for each task.
Quick context hygiene
/context # vê quanto da janela já foi usada /compact # resume a conversa, preservando o essencial /clear # zera tudo e começa do zero (refresh total)
⌨️ Slash commands and plan mode
Slash commands are commands you type with / that control the conversation and context — your everyday hygiene tools. And for larger tasks, there's the plan mode: the assistant reasons and proposes a plan before before running anything.
Quick commands to clean up, summarize, measure, and switch things—plus a mode that plans before acting.
Save time and tokens, and plan mode reduces errors and rework on difficult tasks.
/clear · /compact · /context · /model · /cost · /init · /memory · plan mode · declared intent.
Essential everyday commands
/clear apaga a conversa e começa do zero /compact resume a conversa preservando o essencial /context mostra quanto da janela já foi usada /model troca o modelo em uso /cost mostra uso de tokens e custo da sessão /init pede para o assistente conhecer a pasta/projeto /memory abre e edita os arquivos de memória (CLAUDE.md)
🗺️ The trick that works in plan mode
Always state your real intent. Something like: "My stated intention is X. I’m asking you for this because what I want to be true is ___." Making your intent explicit greatly improves the response.
Going deeper (optional): /compact vs /clear
Use /clear when you switch tasks completely — you want a clean slate. Use /compact when you're still on the same task, but the conversation has gone on too long: it summarizes what matters (decisions and objective) and frees up space without losing the thread. Small difference, big impact on quality.
Self-check (optional): you finished a task and are about to start another, very different one. Which command should you use?
✏️ Exercises
1. Installation complete (easy)
Install Claude Code and sign in. Criterion: the app opens connected to your account, and you can send a message.
2. First file (easy)
Ask the AI to create a folder with two files .md (as in the guided example). Criterion: the folder and both files exist on the computer and have content.
3. Context Hygiene (medium)
Start a conversation, run /context, then /clear, and run /context again. Criterion: you see the context drop to near zero after the /clear.
4. Plan mode (medium)
Enable plan mode and ask for a plan to "organize my Downloads folder"—without executing it. Criterion: the assistant returns a step-by-step plan without changing anything.
5. Declared intent (medium)
Redo exercise 4, including "My stated intention is...". Criterion: the plan visibly aligns more closely with what you want.
⌨️ Ready-to-use prompts
Now you can run these for real. They cover everything from the first file to plan mode and context hygiene. The complete collection is in the Library → Prompts.
Your first command (creates files)
Create a folder called "cafe-gelado" on my desktop and write two .md files inside: "por-que-cafe-e-otimo.md" and "por-que-cafe-nao-e-otimo.md". Three short sentences in each.
Plan mode, no execution
My Downloads folder is a mess. Without running anything yet, give me a step-by-step PLAN to organize it by file type. I’ll approve it before you take action.
Declared intent (canonical)
My stated intention is: I want to understand, not just receive something ready-made. Explain what the Claude Code "context window" is in 3 simple sentences, as if I had never programmed before.
Original (EN):
My stated intention is ___. I'm asking you to do this because the thing I want to be true is ___.
Get familiar with a project
Familiarize-se com esta pasta (rode a inicialização) e me diga, em 5 bullets, para que serve este projeto e quais são os arquivos mais importantes.
⚠️ Common errors
- ✗Thinking the free plan is enough. That’s not enough — Claude Code requires a paid plan. Without one, you can install it but can’t use it.
- ✗Let the conversation turn into a monster. Mixing ten tasks into one window leads to context rot. Use
/clearbetween tasks. - ✗Start with “ignore permissions.” On a real computer, that means asking it to delete or mess something up. Start with "accept edits".
- ✗Blindly trust. The model can hallucinate. Review important results—especially in long conversations.
- ✗Fill the
CLAUDE.mdof text. More than ~200 words starts to clutter the context. Keep it short and direct.
✅ Module summary
Completion checklist
- ☐ Claude Code is installed and connected to my account (paid plan).
- ☐ I understand the difference between the model (Tony Stark) and the environment/IDE (the armor).
- ☐ I can locate the model selector, permission mode, usage indicator, and context window.
- ☐ I created my first folder and files by asking AI.
- ☐ I used /context, /clear, and /compact at least once.
- ☐ I tried plan mode with a stated intent.
- ☐ I know that “one task, one window” is the golden rule.
Next module:
1.3 — Memory system: how CLAUDE.md gives the assistant lasting context