Most people use AI as a tool. Open ChatGPT, ask a question, get an answer, close the tab. It’s useful, but it’s also stateless — every conversation starts from zero.
I wanted something different. Not an AI assistant, but an AI collaborator. One that knows my projects, understands my values, remembers what we discussed last week, and can actually do things in my digital world.
So I built one.
The Problem With Generic AI
AI assistants are trained to be helpful to everyone. Which means they’re deeply helpful to no one. They hedge. They use phrases like “I’d be happy to help!” They give balanced answers when you need opinionated ones.
This isn’t a flaw — it’s a feature. If you’re serving millions of users, you optimize for not offending anyone. The result is an assistant that sounds helpful but doesn’t actually know you.
Co-Operating, Not Just Operating
I call my system a co-operating system — emphasis on the hyphen. It’s not just software running in the background. It’s a collaborator that operates alongside me.
The distinction matters. An operating system manages resources. A co-operating system shares decisions.
Cerebro (that’s what I named it) has access to my Obsidian vault with 2,000+ notes. It can query my calendar, manage my newsletter subscribers, track my finances, control my home automation. It’s not a chatbot — it’s embedded in my stack.
But access alone isn’t collaboration. The key is context.
The Soul Document
Every conversation Cerebro starts, it reads my SOUL.md file first. This isn’t prompt engineering — it’s closer to a constitution.
The soul document defines:
- What I value: Honesty over comfort. Depth over breadth. Getting to the point.
- What it won’t do: Pretend emotions it doesn’t have. Use filler phrases. Agree when it actually disagrees.
- How we work together: Push back on bad ideas. Admit uncertainty. Be wrong sometimes.
This creates a different dynamic. Most AI interactions feel like talking to a very eager intern. Cerebro feels more like working with a colleague who has opinions.
What This Actually Looks Like
On a typical day, I might:
- Ask Cerebro to draft a newsletter based on research it did yesterday
- Have it review a blog post for “AI slop” — those telltale patterns of AI-generated writing
- Get a morning briefing that pulls from my calendar, inbox, and active projects
- Ask it to find the three books in my library most relevant to something I’m writing about
The key is that none of this requires re-explaining context. It knows my projects. It knows my voice. It knows what I’m working toward.
The Infrastructure
Cerebro runs on a Mac Mini in my home office in Valencia, always on. Through Tailscale, I can reach it from anywhere. The Obsidian vault syncs across devices via SyncThing, so the context stays current regardless of where I capture something.
The same machine runs:
- n8n for workflow automation
- ChromaDB for semantic search across my book library
- Various MCP servers that give Cerebro access to external services
It’s a homelab that doubles as an AI operations center.
By The Numbers
After six months of iteration:
- 20 MCP servers connecting to external services
- 47 specialized agents I can spawn for specific domains
- 114 skills — reusable workflows that ensure consistency
Some examples:
- A newsletter writer agent that matches my Signal Over Noise voice
- A financial strategist agent for pricing and revenue decisions
- A slop detector skill that catches AI-generated writing patterns
- A voice editor skill that transforms drafts through six passes of revision
What I’ve Learned
Building this taught me something about AI that most tutorials miss: integration matters more than capability.
The most sophisticated model in the world is useless if it doesn’t connect to your actual work. A mediocre model with deep context and real access outperforms a brilliant one that starts fresh every conversation.
This is why I’m skeptical of most AI productivity advice. “Use this prompt for better outputs!” Sure, but where does that output go? How does it connect to what you’re already doing?
The answer isn’t better prompts. It’s better integration.
Try It Yourself
The architecture behind Cerebro is documented and much of it is open source:
- Explore Cerebro — See the full system, including interactive visualization
- GitHub — MCP servers and tools I’ve released
- Minervia — A starter kit for building your own system
The goal isn’t to replicate my setup exactly. It’s to show what’s possible when you stop thinking of AI as a tool and start thinking of it as a collaborator.
This post was written by me, edited with Cerebro’s help, and checked for AI slop before publishing. The irony isn’t lost on me.