Detailed content
🚀 The Command start
You run /os-coach start <your goal> inside a folder, and an OS is born there. The text after start is your objective, in your own words — not a technical title, but an honest description of what you want to solve. The folder where you run becomes the OS.
How to read: the goal goes in, the coach asks a few questions and, at the end, writes two real files: the memory.md (your place) and the CLAUDE.md (the Identity). One command starts the entire first layer.
Why learn
Because the goal is the anchor for everything. Every future layer—substrate, rules, skills—will be scored against this sentence. A clear, honest goal now prevents a generic OS later. That’s why the coach captures your exact words, without translating it into "tech speak."
Key concepts
❓ The 2–3 questions
After the objective, the coach asks at most three short questions—and stops and waits. You answer in the next turn, and only then does it build. That's "one step at a time" in action.
Who is this OS for?
Just you, your team, or your clients? This defines the tone and who the system responds to.
What would you like to simply ask it to do?
The dream question. E.g.: "what's due in the next 10 days, and what's already overdue?".
One thing it should always do and one thing it should never do?
The first 2 defaults and your OS’s first hard refusal.
🔎 Why so few questions?
Because three short answers already capture almost the whole Identity: who you are, whom you serve, what you do, and what you refuse. The coach is deliberately lean—it prefers to build early and adjust later rather than question you for ten minutes.
Why learn
Because knowing the questions ahead of time helps you answer well. The “always/never” question is the most valuable: the “never” part becomes a hard refusal in the CLAUDE.md, the first fence around your OS. Think it through carefully.
Key concepts
♻️ Reuse responses
Here’s a subtle but important detail: the coach never asks twice. Before asking anything, it checks what you've already said—in the goal, in the memory.md and in previous answers—and reuses it. It never asks “who is this for?” if you’ve already said.
💡 Objective + start already cover almost the entire Identity
So, when starting the Identity layer, the coach doesn’t ask again the objective and who it’s for: it already has those. Usually, all that remains is the “always/never” question. You feel that getting started was quick precisely because it uses everything.
✓ The coach reuses
- ✓The goal you gave in start.
- ✓The “who it’s for” and the dream question.
- ✓Decisions already recorded in the memory.md.
✗ The coach avoids
- ✗Repeating a question you’ve already answered.
- ✗Answer for you (“I’ll assume that...”) unnecessarily.
- ✗Dump all the questions from all the layers at once.
Why learn
Because it shows the method’s efficiency: every answer you give becomes a reusable asset instead of something that gets lost. It also tells you what to expect—if the coach repeats a question, something has gone off track (perhaps the memory.md wasn’t read).
Key concepts
🪪 O CLAUDE.md generated
With the answers in hand, the coach writes the CLAUDE.md: a file short (aim for fewer than 30 lines) with who the OS is, whom it serves, the goal, two or three defaults of behavior and two or three hard refusals. All in your voice, grounded in your answers.
🌱 New here?
O CLAUDE.md is a regular text file (format Markdown, just text with a few symbols). The harness reads it first, before everything else — that's why it's “the soul.” A default is a default behavior (“always remind me 3 days in advance”); a hard refusal is a nonnegotiable "never."
Illustrative recreation — a photographer’s CLAUDE.md (goal: never miss a deadline again)
# Quem eu sou Sou o sistema operacional de um fotógrafo de casamentos solo. Existo para que nenhuma entrega a cliente atrase. ## A quem sirvo Só a mim (operação de uma pessoa). ## Sempre (defaults) - Avisar com 3 dias de antecedência de qualquer data contratada. ## Recusas duras (never) - Nunca mandar mensagem a um cliente sem você ler e aprovar antes. - Nunca deixar um prazo passar em silêncio.
Short, specific, in the owner’s voice. Notice: no em dashes, no real client names.
✓ Lean identity
- ✓Fewer than 30 lines, cut until it hurts.
- ✓Concrete refusals, not “be careful.”
- ✓In your voice, about your objective.
✗ Bloated identity
- ✗Generic “best practices” pages.
- ✗Vague refusals you can’t test.
- ✗Copied from another domain, without your objective.
Why learn
Because the CLAUDE.md is the file the OS reads before any other: it sets the tone for everything. An overloaded soul file confuses the system. The common mistake here is making it long and generic— cut until it hurts.
Key concepts
🔒 Sensitive data protection
If you mark a field as private in any answer—a name, an email, an amount—the coach never logs this actual value in any file in the folder. Instead, it uses a stable, non-identifying identifier: "Client A," initials, a short code.
🌱 New here?
One handle (identifier) is a nickname that represents data without revealing it. Instead of writing "Maria and João Silva," the OS writes "Couple A." You know who they are; the file doesn’t. That way, even if the folder leaks, no personal data goes with it.
⚠️ The forbidden contradiction
The coach never writes a file that says “this field was excluded” and then lists the field right below. That contradiction is itself a leak. If a handle would make the OS unusable for you, it asks first to record the actual amount instead of deciding on its own.
Why learn
Because protection starts in the first layer and applies across the entire OS. Marking what’s sensitive right in the start ensures that, from skills to audits, no saveable file (such as the OS-AUDIT.md) load a real piece of data. You don’t need to police it; golden rule #8 takes care of that.
Key concepts
✅ Done-check: the stranger test
How do you know the Identity is ready? By the stranger test: someone who has never seen your project reads the CLAUDE.md and can correctly describe what the OS is for and one thing it refuses to do. If it passes the test, the layer becomes solid.
✓ Passes the test
- ✓The OS’s purpose becomes clear in one sentence.
- ✓There is at least one concrete, readable refusal.
- ✓A nontechnical person can understand it without you by their side.
✗ Not yet
- ✗A stranger wouldn’t be able to say what it’s for.
- ✗There is no refusal, or it's too vague.
- ✗It’s long and generic — it confuses more than it helps.
💡 Practical tip
Couldn't get a stranger? Reread your CLAUDE.md pretending it has never seen the project. If you get stuck on “but what is this even for?”, the layer is still in progress — cut the generic parts and let the mission and refusal speak loudly.
Why learn
Because it gives you an objective green light instead of an “I think it’s good.” The done-check helps you avoid both extremes: stopping too soon (Vague Identity) or perfecting forever. If it passes the stranger test, move on to the Substrate.
Key concepts
▶️ Copy-run: run the start
Time to build for real. You’ll create a new folder for your first OS and run the start with your real goal. At the end, there will be two files in the folder: memory.md e CLAUDE.md.
▶ Copy-run · start your OS
Objective: capture your goal and generate the Identity. Open Claude Code inside a new (empty) folder for that OS and type:
/os-coach start <seu objetivo em uma frase>
Replace: <your goal in one sentence> by what you want to solve, in your own words. E.g., /os-coach start quero nunca mais perder um prazo de entrega de cliente.
What will happen: the coach confirms the goal and asks 2-3 questions. Answer (it waits) and then writes the files.
How to verify: list the folder and check that it exists memory.md e CLAUDE.md. Open the CLAUDE.md and apply the stranger test: can you state the purpose and one refusal? If so, the Identity is solid and you can move on to module 4.3.
Tip: if you marked any name or value as sensitive, check that the CLAUDE.md use a handle ("Client A"), not the real data.
Why learn
Because it’s your first layer built with your own hands. From here on, you’re no longer reading about the method—you’re using it. And the memory.md the newly created one is what will let you continue through the next modules without starting over.
Key concepts
✅ Module summary
Next module:
4.3 — next: fix the Substrate (the behind-the-scenes) 🗂️