Cerebro is not a note-taking app.
It is a structured vault of everything that matters to how I work: daily notes, profiles for every person I work closely with, project histories, decision logs, and a set of automated skills that run as agents when I open a session. The intelligence layer is Claude — not as a generic assistant, but as something that has read the vault and picks up the conversation where it left off.
When I say picks up where it left off, I mean something specific. A session started today knows the outcomes of the leadership meeting last Tuesday, the tension that surfaced in a 1:1 three weeks ago, the decisions made last quarter and why. I do not re-brief it. It reads the vault. This is the difference between an AI assistant and an AI that is genuinely contextual.
Building it required skills I did not have.
What I Did Not Expect
I am a CPTO. I spend my working days thinking about product and technology strategy, org design, and how teams build things. I do not, professionally, write shell scripts.
I spent a weekend learning how to write shell hooks.
The vault runs on Obsidian — a markdown-based tool that stores everything as plain files. To add the AI layer, I needed to understand how Claude sessions are initialised with vault context, which required understanding how memory files are structured and loaded. That led to YAML frontmatter, which led to how the CLI processed note properties, which eventually led me to shell scripting and git.
Version-controlling your own thoughts. That was not on the professional development plan.
The skills required, in rough order of surprise:
- Markdown — seems trivial. Is not trivial when building a system of interconnected notes with consistent structure.
- YAML frontmatter — the metadata layer that makes notes queryable and structured rather than just searchable.
- Git — for versioning the vault across three machines without losing anything.
- Shell scripting — for the automation layer: hooks that process notes, scripts that run skills.
- Prompt engineering — for making the AI layer actually useful, which requires understanding how context is passed, what the model can and cannot reason over, and how to structure memory so it is retrieved correctly.
None of these required deep expertise. All of them required T-shaped curiosity: willingness to go just deep enough in an adjacent domain to make the thing work.
A CPTO learning shell scripting because it serves the system. That was the point — not the skill itself, but the instinct to acquire whatever the system needed.
The Moment It Clicked
About three months in, during a session where I was preparing for a difficult conversation, Claude surfaced something I had written four months earlier. A note from a 1:1. Context about what someone had been working through at that point.
I had completely forgotten writing it.
The AI had not.
That is a different thing from productivity. It is a relationship between you and your own past thinking. The vault does not just help you work — it extends what you are able to remember, which extends what you are able to care about, which changes the quality of the work you do with other people.
It took several rebuilds to get there. The first version was too structured. The second was too flat. The version I run now is the result of learning what the system needed by watching it fail.
The question I kept asking was not whether it worked. It was whether it changed anything.
Post 3: six months of evidence.
Leave a Reply