Detailed content
🧱 First principles: remove assumptions, count the atoms
Every problem is a a pile of assumptions on top of atoms. Thinking from first principles means removing assumptions, counting the atoms, and rebuilding. Most people reason by analogy — they copy what others do with small variations. That's why almost no one makes big leaps.
A way of reasoning that rejects "because I said so." You break down the problem until you reach truths you can verify yourself, then rebuild from there.
It’s how you move from “copying with slight tweaks” to solutions that seemed impossible. The only real limit is physics, not convention.
Assumption vs. atom · recommendation disguised as law · "what’s the real constraint here?"
🔋 The $600 battery analogy
In 2012, people said a battery cost $600 per kWh and that this was an "insurmountable floor." Someone broke the battery down into the metals it contains, checked the price of each on the exchange, and arrived at ~$80 per kWh. The "floor" was conventional wisdom; the raw materials told a different story.
💡 Guiding question
Everything that seems fixed — price, deadline, number of people, “the right way” — tends to be a recommendation disguised as a rule. What actually holds up (load-bearing) is the physics: energy, the cost of the atom, time. In any project, ask: what is the real constraint here?
🗑️ The 5-step protocol (the order matters)
Before building, trim the scope. The #1 mistake in automation is automate something that should never have existed. Skipping a step makes you optimize what shouldn't exist — that's why the order is sacred.
A five-step engineering process for simplifying any system or idea before building.
If you optimize before deleting, you waste effort polishing junk. Deleting first saves everything that comes after.
Requirement with an owner · delete ~10% back · optimize only afterward · state it until it gets boring.
Question every requirement
Each requirement must have the a person's name behind it, never "legal asked for it." A requirement without an owner is suspicious.
Delete the part or step
If you don't later need to put back at least ~10% of what you removed, you didn't delete enough.
Simplify and optimize
Only then to delete. Optimizing first is the classic mistake.
Speed up the cycle
Now that the system is lean, make each cycle faster.
Automate
Finally, so you don't have to touch it again.
🧠 Mental step-by-step process
Break everything into parts → what is given? what is substantive? what is a constraint? → reduce it until the essential question is laid bare → state the problem until it gets boring → only then write code.
🧩 The Universal Stack: Think + Store + Create + Scrape + Connect
Beginners think they need dozens of confusing tools. They don't. Almost everything built with AI comes down to five functions. Knowing this removes the paralysis.
The minimum set of tools for building almost anything. Surprisingly short.
You add tools only when one of the five functions calls for them—never just because they're trendy.
Think · save · create visuals · scrape · connect.
Think
An AI that reasons and executes (Claude Code).
Save
A place for data (a database with a ready-to-use API).
Create visuals
Image, asset, and polished screen generation.
Scrape data
Fetch public information from the web (ready-made scrapers).
Connect
Automations that connect the pieces (the AI itself, or fine-grained visual control).
That’s all
You almost never need more. Combine the five and build.
🍳 The stack as a kitchen analogy
Building is like putting together a complete kitchen: the chef who thinks (the AI), the pantry for save ingredients (the database), the dish for plate pretty (the visuals), the trade fair for search fresh ingredients out there (the scraper) and the brigade that makes everything run on its own (the automations). If one station is missing, the dish doesn’t go out.
🛣️ Autopilot (AI)
- ✓Claude Code handles most of the process.
- ✓Ideal for most of the journey, without you having to steer.
🎛️ Take the wheel (visual automation)
- ✓Visual tools offer fine-grained control over specific sections.
- ✓It’s good to know how to drive the car in both modes.
🎯 Who you're looking for + scrape with the scraper store
Without a defined audience, no scraper is useful. First define the Ideal Customer Profile (ICP) — exactly who you want to reach—and only then start collecting data.
Define the niche + location + signals (e.g., "roofers in São Paulo with a social media presence") before opening any tool, then scrape public data.
Lead generation is a numbers game: think hundreds, not dozens. And you’re paid for the value you create and how much of that value you can capture or replicate.
ICP · volume game · API key · “actor” (each scraper) · multichannel.
🕷️ Step by step: scrape leads
- Define the ICP: niche, city/region, and signals. Write it down before opening any tool.
- Create an account with a scraper service (e.g., Apify, on apify.com) and sign in.
- Go to Settings → API & Integrations and copy your API key.
- Connect the store to Claude Code (via a connector/MCP or by pasting the token into the configuration).
- Describe what you want in natural language. The AI finds the best "actor", runs it, and returns the data.
⚖️ Cost vs. time
There are free and paid options. Remember that your time has a cost too. If you have more capital than time, platforms that already provide validated leads (such as Apollo, apollo.io) save steps. If you have more time than capital, build the workflow piece by piece with a scraper + a validator.
✅ Validate, enrich, and package the result
A large list is worthless if the emails bounce—and data in a terminal isn't a result. You validates the contacts and then packages in a format you actually use: the screen where you open it and can call the contact right away.
Validate email addresses (do they exist? are they active?), enrich contacts, and turn raw data into a usable HTML/dashboard.
An invalid email hurts your sending reputation and burns your list. Validation is cheap compared with the damage.
Bulk validation · "desired outcome" in the prompt · scheduled update · deploy + repository.
🔎 Validate and enrich
- •With a prospecting platform (e.g., Apollo, apollo.io) is already built in — you can skip this step.
- •On a tight budget, use a batch validator (e.g., Anymail Finder, anymailfinder.com): searches, validates, and finds missing email addresses, and has an API.
- •Give Claude Code the key and ask it to validate the scraped list—without manually uploading a file.
📊 Package it into something usable
- •Ask for an HTML page/dashboard with the fields that matter (name, email, social media, an interesting fact).
- •Open locally to review.
- •If it's recurring, host it and schedule updates (e.g., every 7 days)—set it up once and forget about it.
🍳 Guided hands-on example
Goal: a usable list of 10 roofers in São Paulo, with name, email, social media, and an interesting fact—to start conversations.
- Define the ICP: "roofers in São Paulo with an online presence".
- Connect the scraper store to Claude Code (copied API key).
- Ask for the scraping in Portuguese. The AI chooses the best "actor" and brings back the data.
- Validate the emails with the validator key and receive the cleaned version.
- Ask for a beautiful HTML page with the 10 contacts and open it locally.
- Assess the savings: what is the cost per lead? How many do I convert from that?
Notice: you used four of the five pieces—think, scrape, validate, and create visuals. Only “store” was missing; use it when you want to persist the list in a database.
🧠 High Agency + the Clarification Meta-Prompt
High agency is the belief that you can solve problems, find workarounds, and make things happen, even when the obvious path is closed. And when you don’t know the way forward, the best move is to have AI interview you first.
The mindset of someone who solves problems without waiting for permission, paired with a prompt that unlocks any poorly defined problem.
~80% of the time, the problem isn’t technical—it’s a lack of clarity about the bottleneck. The meta-prompt solves this by having the AI clarify your intentions before suggesting a solution.
Constraint as recommendation · act without permission · state the objective + desired outcome · think through the solution together.
📞 The high-agency gut check
If you were stuck and had one call to make to someone who would definitely answer and help you, who would you call? That person is your high-agency role model. The goal is to become that person: treat constraints as recommendations until proven otherwise.
💬 The move that unlocks everything
Whenever you get stuck, send the meta-prompt "ask me questions to clarify" (it’s ready in the prompts section below) and answer the questions the AI asks. Often, the clarity that comes from the questions is already half the solution.
Going deeper (optional): why questions come before answers
Asking for a solution to a poorly defined problem forces the AI to guess—and it will take a shot. When you turn that around and ask it to interview you, each question exposes a hidden assumption. In the end, the real bottleneck usually fits in one sentence. Then the solution is almost trivial.
Self-check (optional): in the 5-step protocol, what comes BEFORE simplifying and optimizing?
✏️ Exercises
1. Break it down using first principles (medium)
Take a “fixed rule” from your work (“this takes 3 days,” “this needs approval from someone”). List the assumptions and separate what’s physical from what’s convention. Criterion: identified at least one “rule” that was only a recommendation and described a viable alternative.
2. Map your idea to the universal stack (medium)
Choose one of your projects and fill in the five functions: think / store / create / scrape / connect. Criterion: all five addressed, with a justification for each item included or left out.
3. Run the Lead Workflow (practical)
Connect a scraper store to Claude Code and scrape 10 real leads from your ICP in HTML. Criterion: HTML file open in the browser with 10 contacts, each with a name and a contact channel.
4. Apply the 5-step protocol (medium)
Take an automation you have (or are planning) and go through the 5 steps in order. Criterion: at least one step/requirement deleted before any optimization, with justification.
5. Use the meta-prompt on a stuck problem (easy)
Choose something you don't know how to solve and send the clarification meta-prompt. Answer the questions. Criterion: you state the real bottleneck in one clear sentence after the switch.
⌨️ Ready-to-use prompts
Always start with clarification meta-prompt — is the most important. The full database is in the Library → Prompts.
⭐ Clarification meta-prompt (the most important)
This is the result I want: [describe the desired result]. Before suggesting any solution, ask me questions to clarify my intentions. How should I think about this problem? Then let's think through the solution together.
Original (EN):
This is what I am trying to achieve: [desired outcome]. Ask me questions to clarify my intentions before proposing anything. How should I best think about this problem? Then we'll think about the solution together.
Break it down from first principles
Quebre este problema em primeiros princípios. Liste as suposições,
separe o que é restrição física real do que é só convenção
("sempre se fez assim"), e reconstrua a solução a partir do que
não pode ser duvidado: [problema].Original (EN):
Break this problem down to first principles. List the assumptions,
separate real physical constraints from mere convention
("this is how it's always done"), and rebuild the solution from
what cannot be doubted: [problem].Apply the 5-step protocol
Aplique o protocolo de 5 passos a este processo, na ordem: (1) questionar cada requisito — cada um precisa do nome de uma pessoa responsável; (2) deletar partes; (3) simplificar e otimizar; (4) acelerar o ciclo; (5) automatizar. Processo: [descreva].
Map my idea to the universal stack
I have this idea: [idea]. Map it to the universal stack — thinking, saving, creating visuals, scraping data, connecting with automations. For each of the five functions, tell me which tool to use or whether I can do without it.
Scrape leads
Quero alcançar [nicho] em [cidade/região]. Use uma loja de raspadores, encontre o melhor "actor" e me traga, de cada lead: nome, e-mail, redes sociais e um fato interessante. Me dê 10 para eu iniciar conversas. Pode entregar numa página HTML bonita.
Package the result in a recurring dashboard
I want a dashboard with this data: [fields]. My intended outcome: [what I want to see and do]. Generate it in HTML, organize by [criterion], and explain how to host it and schedule updates every 7 days.
⚠️ Common errors
- ✗Optimize before deleting. You simplify and automate a step that should never exist. Always question it and delete it first.
- ✗Confuse convention with physics. Accepting “it can’t be done in a week” as a rule. It’s almost always a recommendation—ask what the constraint is real.
- ✗Scraping without an ICP. Collecting data without knowing who you’re looking for creates a large, useless list. Define your audience first.
- ✗Skip email validation. A list full of bouncing emails damages your sending reputation. Validate it or use a source that has already been validated.
- ✗Asking for a solution without clarity. Ask AI to "solve" a poorly defined problem. Use the meta-prompt and let AI interview you first.
✅ Module summary
Completion checklist
- ☐ I can explain in my own words what first-principles thinking is and gave an example of my own.
- ☐ I can distinguish, in a real problem, between a physical constraint and a convention.
- ☐ I can recite and apply the 5-step protocol in order.
- ☐ I can name the 5 parts of the universal stack and map one of my ideas to them.
- ☐ I connected a scraper store to Claude Code and ran a scrape for real leads.
- ☐ I validated the emails on a list.
- ☐ I have the clarification meta-prompt saved and have used it at least once.
Next module:
2.4 — Design Systems: Give Everything You Build a Consistent Visual Identity