Before any tool, framework, or automation, a cognitive rewire needs to happen. This module installs the mental operating system that makes all the difference between using AI and operate with AI.
Illustrative diagram — The Default Shift: every task goes through the question "to what extent can AI help here?"
O Default Shift is the most important habit change in this course. Before carrying out any task the old way, your first question needs to be: "how could AI do this?". If not 100%, at least the first 30%. The real question isn’t "can AI do this?" (binary, blocking) — it’s "to what extent can AI be applied here?".
Imagine you need to update links in 300+ YouTube descriptions. Old way: open each video manually, edit, save. Hours of mechanical work.
With Default Shift active: you ask, "How far can AI take this?" The answer: 100%. While you get some water, Claude Code writes a script that processes all 300 videos through the YouTube API. Work that would take hours becomes minutes.
Tenho 300+ vídeos no YouTube onde preciso trocar o link
"exemplo.com/antigo" por "exemplo.com/novo" em todas
as descrições. Você consegue escrever um script
que faça isso via YouTube Data API v3?
Liste o que você precisa para começar (credenciais,
escopos de OAuth, etc.).The Default Shift starts with the prompt. You stop thinking “what am I going to do?” and start thinking “what am I going to delegate?”
Most people approach AI with a binary question: "can AI do this or not?". This frame is limiting because it forces an answer that rarely reflects reality. The right question is graded: maybe AI delivers 80% quality, and you finish the remaining 20% in 10 minutes instead of working 2 hours from scratch.
Treat the Default Shift like learning to type. At first, you need to consciously think about each key. With practice, it becomes automatic. Set aside 2 weeks to actively question each task before carrying it out. The conscious effort now builds the habit for good.
You used AI to generate the first draft or structure. You avoided the "horror of the blinking cursor." You do 90% of the work, but you don’t start from zero.
AI does half the heavy lifting. You direct, review, refine, and make strategic decisions. At least 2x productivity.
You define the problem, and AI handles it completely. Your role is to check quality and iterate on the prompt if needed. The goal of every mature workflow.
One of the most common traps: you tested something with AI 3 months ago, it didn’t work, and now you carry that "this can’t be done with AI" as truth. The problem? Models get better faster than our memory of past failures. What was impossible in March may be trivial in July.
At one point, creating high-quality infographics with AI was considered "impossible" — models generated incoherent images, incorrect text, and absurd proportions.
Three months later, a new model produced professional infographics with a well-structured prompt. The work that “couldn’t be done” became trivial. The mistake wasn’t the failed attempt—it was not trying again.
Create a recurring event on the calendar: "AI limitations retest". Every 4–6 weeks, pick 3 tasks you marked as "AI can’t do this" and test them again. You’ll be surprised how often the answer changes.
You don’t automate one work. You automate one piece of work. Function Breakdown is the process of breaking down your role into functions, and each function into micro-tasks — then identifying which of those micro-tasks AI can perform. The magic happens when you start chaining these pieces together.
The task "create a YouTube video" seems too big to automate. But when you do the Function Breakdown:
Each of these micro-tasks can have its own automation. You don’t need all of them at once— starts with a.
What do you do at work? Not the tasks—the functions. "Create content," "serve customers," "analyze data," "manage projects."
For each role, list 5–10 specific, repeatable activities. The more granular, the better the chance of automation.
The most frequent and time-consuming microtasks are the first candidates. Automate one, validate the result, then move on to the next.
Over time, individual automations start connecting. One’s output becomes the next one’s input — and you’ve created a real workflow.
AI generates outputs quickly. Too quickly for anyone who doesn’t question them. The Curiosity Rule is a discipline: never accept an output without asking why. Ask for 3 alternatives. Question the logic. Understand what was built. Because "if you built something and can’t explain how it works, you built a liability, not an asset."
Dark code is code (or an automation or prompt) that works but nobody knows exactly why — not even you. When it stops working, you're stuck. When you need to adjust it, you don't know where to start.
Dark code is a liability disguised as productivity. In the short term, you gained speed. In the long term, you created a blind dependency.
# Após receber qualquer output técnico:
"Explique cada parte desse código/automação
em linguagem simples. O que acontece se X
mudar? Quais são os pontos frágeis?"
# Pedindo alternativas:
"Dê 3 formas diferentes de resolver este
problema. Para cada uma, explique os
trade-offs e me diga qual você recomenda
e por quê."
# Validando a recomendação:
"Por que essa abordagem é melhor que [X]?
O que eu perderia se escolhesse [Y] em vez?"Reframe the relationship: AI isn’t a machine that sells answers. It’s a mentor with vast knowledge but no context about your business. You question the mentor. You ask for explanations. You disagree and ask it to defend its position. This protects you from dark-code and accelerates your learning.
No one tells you about the Dip. In the first 1–2 weeks of actually working with AI, your productivity drops by about 20%. New workflows are slow. You test things that don’t work. You learn what no ask. It’s uncomfortable; it feels like a regression.
In 2 weeks, the baseline doubles. It doesn’t increase by 20%: fold. The workflows you tested and broke are already working. You’ve learned what to ask. The investment in temporary discomfort buys permanent productivity.
But you have to push through. Anyone who gives up in the Dip — “this doesn’t work for me” — stays stuck at their old performance level forever, without ever seeing the payoff.
Performance is normal or slightly above average. You’re motivated and trying new things. It feels easy.
Performance drops ~20%. Old workflows have broken, and new ones aren’t flowing yet. This is where most people give up.
The new workflows are running. You’re operating at a sustainable pace. Performance won’t return to normal—it will go far beyond it.
Most people who give up on AI gave up during the Dip. Not because AI didn't work — but because they didn't know the Dip is temporary and unavoidable. Knowing it's coming completely changes how you face it.
Real learning doesn’t happen with the first successes—it happens with the first failures. The goal isn’t to avoid failing: it’s to reach your the first 10 errors in the fastest and safest way possible. Each well-documented error is worth more than 10 successes you didn’t understand.
Quick: don’t wait for perfect conditions to get started. Test today, with the tool you have, on the task at hand. Iteration beats planning.
Safe: never test automations in production without review. Have a test environment, back up the data, and review the outputs before applying them. Speed without care leads to disasters.
Keep a simple file (it can be in Claude Code itself) with entries in this format:
Data: 2026-06-01
Tentei: automatizar relatório de vendas em PDF
Resultado: Claude gerou tabela com dados fictícios
Por que errou: não forneci os dados reais no contexto
Aprendizado: sempre incluir os dados junto com o prompt
Próximo passo: testar com CSV colado no contextoThis record turns every failure into transferable learning — for you and others.
1.2 — Method (How to Decide) — Learn the decision framework for choosing when, how, and how much to use AI in each context of your work.