eccher.
PT
AI in practice

How I built my second brain and why it left me calmer rather than more productive.

· · 6 min read
Translated from Portuguese. Read the original

Building my second brain let me stop carrying the system in my head; productivity wasn't the gain.

I already used Obsidian, kept my notes in order, and remembered things well. I was already organized. My problem was the cost of keeping everything organized by hand, a cost that keeps demanding attention even when you're not using the system. Manual organization demands attention even when the system sits idle.

We usually mistake being organized for remembering things well, and we assume what's missing is a tool, that all it takes is swapping Obsidian for Notion, or Notion for Todoist. Keeping information organized by hand doesn't scale, even when you're good at it. The tool isn't the problem.

I knew it was possible to pair AI with Obsidian. I'd read about it and had pictured the design. What I lacked was the push to go from "this could be done" to "I'm going to do it now." That push came from a video by Brad Bonanno explaining how he had structured his own second brain. One look at his workflow and I understood exactly what had been holding me back: I was trying to replace the manual work with a single smart agent, when what works is the opposite: several small agents, each with a clear responsibility, running on predictable schedules.

The build itself took two days. I've been using it for two and a half months.

The architecture

The system lives in an Obsidian vault, organized by areas of life (work, personal projects, personal life), and each area has its own subfolders for people, projects, insights, and to-dos. Every file carries standardized metadata in YAML frontmatter, which lets me filter, search, and run automated processing on the content.

The entire vault is a git repository. It syncs across my computers through Obsidian Sync, and at the same time every file is versioned, with each change recorded. Combining Obsidian and git is what gives me both the fluidity of a notes tool and the auditability of a codebase.

The agents run on a local server I keep at home. They have scheduled jobs that fire at fixed times, read the vault, process the content, and write back to the vault. It works as background infrastructure instead of a chat, running quietly and surfacing only when I ask for something or when it hands me a briefing.

Three routines run every day.

Input has to be cheap, or it doesn't happen

Throughout the day I write a single note in my daily journal. It has no template or rigid structure, and I don't decide where each thing goes. The note is split into sections by area, covering what happened in each context: meetings, decisions, conversations, stray ideas, frustrations, perceptions, open items. What matters is getting it down, and the organizing comes later.

This design is deliberate. If I had to decide on the spot which folder each piece of information belongs in, the friction would be enough to make me stop writing things down. The rule is simple: everything goes in one place, split by area, with no taxonomy. The cost of input has to be close to zero, because any friction in capture brings down the whole system.

The nightly routine is where organization actually happens

Every night, an agent reads the day's note and processes each section. For each area, it identifies what belongs in a project's profile and updates that profile, keeping the earlier history intact and adding the new context with a date. It also identifies decisions that need to be recorded with their rationale, insights that become files of their own, open items that become to-dos in the matching area, and events that become incident records when something got stuck.

At the end, it updates a central file, the active context, which describes the state of everything across my areas on a single page. That file feeds the next day's briefing. Once it's processed, the day's note goes to the archive. It has done its job.

The most common objection here is that this looks fragile: if the agent misclassifies something, the system gets corrupted. The trouble with that objection is that it assumes the previous state was reliable, and it wasn't. When I organized by hand, I sometimes misclassified things too, only without a trace. With git, every decision the agent makes is recorded and reversible. Mistakes became auditable commits instead of invisible mess.

Output arrives before the question

In the morning, before anything else, the briefing runs automatically. Another agent reads the active context and gives me at most ten items, ordered by urgency, with the day's focus highlighted. Instead of a to-do list, I get an assessment of what matters today, given everything that's open across all areas.

That briefing replaces the effort of remembering my priorities. It runs every day, based on information updated the night before, and gives me a clean start. The cost of figuring out what to do first moved out of my head and into the infrastructure.

Versioning is capital that builds up without effort

The choice of git is the least obvious part of the system, and probably the most valuable over the long term.

Usually, when we organize information, we organize state, meaning what's true today. A project's profile describes the project now, and the page for a to-do describes what needs to be done now. When the state changes, the previous one is lost, and all that remains is the current version.

With git, every change is recorded. Every development in a project, every decision made, and every to-do completed or discarded leaves a trace. A year from now, a project's profile will show me where it stands and also how it got there. Decisions that seemed small at the time form a pattern when you see them in sequence.

This is the kind of capital that only becomes visible over time. It already starts building up now, though, at no extra cost, because versioning is automatic.

What this actually gave me

The most concrete consequence was personal rather than professional.

Projects I'd been putting off for months started getting off the ground. The time on my calendar hasn't changed. They got moving because I can keep track of more things in parallel without each one needing attention even when I'm not there.

Before, a personal project paused for two weeks required mental reconstruction before I could continue. Now its state is saved for reference, and when I come back, I pick up where I left off.

I'm describing the design in broad terms because the specific details (frontmatter structure, agent prompts, scheduled jobs) vary too much to fit in a single piece.

If you're building something similar, have already built one and made different choices, or want to start and don't know where, send me an email or find me on LinkedIn. I want to hear where you are in the process.