🏛️ Capstone: your executive assistant

The executive assistant lives in its own folder, separate from work projects—because its memory is about you, not about a specific piece of code.

Mixing everything in a single folder makes the context confusing and causes the agent to lose track of the boundary between “your life” and “your project X.”

Dedicated folder, personal context vs. project context.

Six fixed folders: contexto/, regras-e-skills/, projetos/, registro-de-decisoes/, referencias/ e arquivo/ — each with a unique role and no overlap.

A fixed structure avoids the question “where do I save this?” every time — you (and the agent) always know where to look.

Separation by function, predictable location.

The CLAUDE.md in this setup is short — an index that says "read contexto/negocio.md to learn about the company," instead of pasting everything into it.

A huge CLAUDE.md takes longer to load and is harder to maintain—pointing to other files follows the same principle as Learning Path 0.4, now applied to the entire house.

Index file, on-demand loading, lean beats huge.

contexto/ stores who you are and the business; regras-e-skills/, what the agent can do on its own; projetos/, what’s in progress; registro-de-decisoes/, the history of “why”; referencias/, supporting material; arquivo/, what’s already past.

Without a concrete example, the structure stays abstract—seeing what goes in each folder clears up any doubts when it’s time to create yours.

Function → folder mapping.

A prompt ready for Claude Code to create the six folders and the CLAUDE.md index all at once, in the folder you choose.

This is the first practical step in the learning path—without it, there’s nowhere to save the result from Phase 2.

Copy-and-run example, result verification.

With the house empty but organized, the next step is to fill it contexto/ for real—that’s what Phase 2 (Life) does.

An empty house doesn’t help anyone — the real value comes when it contains real context.

Structure vs. content, phase transition.

Shallow context ("I'm an entrepreneur, I have a company") leads to generic answers; deep context (how you make decisions, what you delegate, how you prefer to be notified) creates an assistant that seems to really know you.

It's the difference between a decorative assistant and one that truly saves you time.

Context depth, garbage in garbage out.

Questions about your role, your work routine, and what you're trying to achieve — becomes contexto/quem-e-voce.md.

Everything else in the interview builds on this block — without it, the assistant doesn’t know who it’s working for.

Profile, routine, goals.

What your company does, who’s on the team, who makes which decisions — becomes contexto/negocio.md e contexto/time.md.

Without this, the assistant can’t distinguish a decision you made from one that needs someone else’s approval.

Organizational structure, decision owner.

What you want to get off your plate first, and what you never want to delegate under any circumstances — becomes contexto/prioridades.md.

It's this list that guides the choice of the first skill in Phase 3—without a clear priority, you build the wrong skill.

What to delegate, what to never automate.

Direct or detailed tone, when to notify and when not to interrupt — becomes contexto/comunicacao.md, the same kind of rule covered in Track 6.6.

If an assistant interrupts you the wrong way at the wrong time, you’ll quickly turn it off—even if it’s technically correct.

Communication tone, notification window.

A prompt ready to ask the agent to lead the interview one question at a time, saving the answers to the right files at the end.

This is the exercise that actually fills the house—without running it, the files stay empty.

Copy-and-run example, verification (read the generated .md files).

Get the list of contexto/prioridades.md and choose the recurring task that takes up the most of your time today — not the easiest one to automate.

A first skill that solves something you feel every day builds confidence to keep going; a trivial one becomes just an exercise with no real use.

Recurring need, real pain point, quick win.

Directly apply the 6-step framework from Track 3.3 to turn the chosen task into a skill in regras-e-skills/.

You already know how to do this — the capstone is where that skill shows up again, this time with a real purpose.

Reusing a framework, project vs. global skill.

Install and authenticate an MCP (a protocol that gives the agent “hands,” covered in Track 2) that reads and creates calendar events.

An executive assistant that can’t see your calendar is decorative—it needs to know when you’re available and when you’re not.

Calendar MCP, authentication, read/write permission.

A second MCP connected to your task manager (or a simple file projetos/tarefas.md, if you don’t use an external app).

Together with the calendar, it gives the assistant a complete view of “what to do” and “when to do it” — the foundation of any real executive assistant.

Task MCP, single source of pending items.

Reuse the scheduled routine from Track 6.7 directly to run the new skill at a fixed time, without you having to trigger it manually.

This is what turns a manual skill into an assistant that works on its own—the heart of the “digital employee.”

Reusing a schedule, autonomous execution.

A prompt ready to launch the new skill for the first time and ask for a summary of what it did—your check that it "really worked."

Seeing it in action is what separates "I read about this" from "I built this."

Copy-and-run example, proof of value.

Your life changes, your business changes—the context you wrote in Phase 2 gets outdated if no one reviews it. Phase 4 is the habit that prevents that.

An assistant with outdated context starts giving wrong answers without warning—silently getting worse, not breaking.

Living context, silent obsolescence.

A short checklist: what changed in the business, on the team, in priorities—and update the files for contexto/ corresponding.

15 minutes a month is cheap; letting 6 months of unrecorded changes pile up is expensive to fix later.

Review checklist, fixed cadence.

Every three months, take another look at contexto/prioridades.md and decide: can you build one more skill? Connect one more MCP?

The home grows one room at a time—trying to expand everything at once goes against the spirit of the previous tracks.

Incremental growth, recurring prioritization.

Whenever you make an important decision with help from the assistant, a short line goes into registro-de-decisoes/: date, decision, reason.

Months later, “why did we decide that again?” becomes a 10-second search instead of a mystery.

Append-only record, searchable history.

Unlike a chatbot that forgets everything with each new conversation, your home files are your assistant's permanent memory—they persist from session to session.

It's this persistent memory that sets an executive assistant apart from a generic chatbot—the central theme of module 7.6.

Persistent memory vs. session memory.

A prompt ready to ask the agent to review the entire home folder and flag anything that seems out of date—your guided monthly review.

Having the prompt ready reduces the friction that might make you skip the monthly routine in the rush of everyday life.

Copy-and-run example, assisted review.

If your home folder exists only on your computer, a hard drive failure or a spilled coffee can wipe out months of interviews, decisions, and adjustments.

Context is the most valuable asset in this entire track—protecting it is just as important as creating it.

Single point of failure, digital asset.

Turn the home folder into a Git repository, with a commit at each monthly review — the same Git used to deploy Track 4.

Along with a backup, you get a history: you can see exactly what changed in each revision.

Git, commit, version history.

The home contains personal and business data—the remote repository (GitHub, GitLab, or similar) has to be private, always.

Accidentally publishing a folder with personal context exposes sensitive information that can’t be taken back.

Private repository, sensitive data, secret (recall Track 4.4).

A scheduled routine (Track 6.2) that automatically commits and pushes every week — even if you forget to do it manually.

Backup that depends on someone remembering fails; scheduled backup doesn’t depend on anyone remembering.

Backup automation, scheduling, resilience.

If an interview or a new skill messes up a file, Git history lets you go back to an earlier version with a simple command.

Knowing you can undo changes makes it less scary to experiment with changes to the house.

Revert, checkpoint, safety to experiment.

A prompt ready to initialize Git in the home folder, create the remote private repository, and make the first commit.

Secure the house before moving on to the final capstone of module 7.6.

Copy-and-run example, check via git log.

The complete home folder, with real context (not a fictional example), at least one working skill, connected MCPs, and the repository backed up — all four phases together.

Seeing the complete list of what "done" means helps you avoid finishing the capstone with one phase only half complete.

Delivery checklist, integration of the four phases.

Four questions: Is the context specific (not generic)? Did the skill solve a real pain point? Is the schedule running on its own? Does a backup exist?

Without an objective criterion, it’s hard to know whether the capstone is finished or still missing something essential.

Evaluation criterion, guided self-assessment.

Three concrete differences: (1) your project context—it knows who you are without you having to explain again; (2) real actions—it edits, schedules, and runs commands instead of just describing them; (3) memory that persists across sessions—the house doesn’t forget.

Closes the loop opened in Track 0.1 ("agent ≠ chatbot") — now with a concrete example you built yourself.

Persistent context, real action, memory across sessions.

A configured environment (T0), the WAT framework internalized (T1), tools via MCP (T2), custom skills (T3), deployment off your laptop (T4), published interfaces (T5), scheduling and loops (T6)—and now an assistant that brings it all together (T7).

Naming what was built reinforces that this wasn’t abstract theory — these were seven real capabilities, stacked together.

Retrospective, accumulated capabilities.

Three possible directions: go deeper into a specific learning path that felt too shallow, add one more skill per month in Phase 4, or share what you’ve learned by teaching someone else.

Finishing the course without a next step lets momentum die—naming concrete options keeps the project truly alive.

Continuity, deeper learning, teaching as reinforcement.

A prompt ready to ask the assistant itself to run the four evaluation criteria on your home folder and point out anything missing, if anything.

This is the final exercise for the entire course—the ultimate proof that the assistant understands itself.

Copy-and-run example, self-assessment, course wrap-up.