TL;DR (aka future GT, read this first): I spent 15 years in sales and marketing building knowledge systems for other companies. When I started building my own (ICB, courses, games), my scattered notes across Notion, repos, and my head couldn't keep up. So I built a second brain — not a note-taking app you fill yourself, but a file system operated by an AI agent that reads my code, extracts concepts, generates writing angles, and drafts content in my voice. It runs without me. Here's how it works, what it produces, and where it falls apart.
I have never been good at taking notes.
Not in school. Not at HCL. Not at Publicis Sapient. Not even at ISB, where everyone else had color-coded Notion dashboards and I had... a vague sense of what was happening.
What I was good at was building systems that other people would use to organize things. Sales methodologies at HCL that the top 150 accounts ran on. Customer health scores at Hevodata for 500 accounts and $3.5M in ARR. Marketing funnels at ekincare that took 35 clients to 201.
The irony wasn't lost on me, but I didn't fix it. I just kept building for other people and telling myself I'd organize my own knowledge later.
Then I started ICB. Solo founder. Next.js, Supabase, scrapers, SEO, content, newsletters — the whole stack. And suddenly "later" was here.
I had knowledge spread across:
- 4 GitHub repos (ICB itself, the scrapers, the content ops, the video pipeline)
- Medium drafts I half-wrote at 2AM
- A Notion page I opened exactly once
- My head (least reliable database, no backups)
Every time I sat down to write a post, I'd spend 30 minutes just finding what I already knew. Then I'd write something generic because the specifics were buried somewhere I couldn't reach.
What I built instead
I built GTBrain.
Not an app. Not a SaaS product. A folder system with an AI agent operating it.
The structure is stupidly simple:
00-inbox/ → Raw material lands here
01-sources/ → Repos I've ingested, articles I've saved
02-concepts/ → Reusable ideas extracted from sources
03-learnings/ → What I actually learned building things
04-writing-angles/ → Angles worth writing about
05-outputs/ → Drafts, ready for review
06-style/ → My voice profile (derived from approved writing)
07-systems/ → How this whole thing works
The agent (a Hermes instance I call GT's personal editor) does the work. Every night, it:
- Reads my recent code commits and ingested material
- Extracts concepts — not summaries, but the principle underneath
- Checks what's already in the brain (no duplicates)
- Generates writing angles with audience, hook, and proof
- Connects orphan notes to the graph so nothing stays isolated
- Drafts posts in my voice
I don't feed it. It feeds itself.
What happened
The cross-project connections surprised me.
The agent found a pattern I hadn't seen: I build infrastructure before output. The gt-courses LMS existed before a single student graduated. The ICB video pipeline was built before a single video was published. GTBrain itself — here you are, reading this post on a system I spent weeks building before writing a word.
The agent extracted this as a concept file called "The Infrastructure-First Trap." Then it generated a writing angle: "I spent more time building the blog pipeline than writing posts. I'd do it again. No, I wouldn't."
That's genuinely useful. I would not have connected those dots myself — not because I'm dumb, but because I'm inside the projects. The agent sees the forest.
It writes in my voice. Mostly.
The first drafts were... polite. Too polite. They sounded like an AI writing about a builder, not a builder talking to another builder. I had to teach it my voice by feeding it approved posts and extracting patterns: specific about tools and numbers, self-deprecating, ends with a question not a sales pitch.
Now it's close enough that I edit, not rewrite.
The system runs without me.
The cron job fires at 1AM IST every night. I wake up to a new concept or draft in my inbox. I review it, approve it, or kill it. The agent also fixes broken links and orphan notes while I sleep — the kind of housekeeping I'd never do myself.
Where it falls apart
-
It can't tell me what to write. The angles are good, but the agent doesn't know what's timely — what conference I just spoke at, what client conversation is still fresh, what thing broke in production yesterday. That's still me.
-
The voice work is fragile. Feed it one bad Medium post and it learns the wrong patterns. The style profile needs curation.
-
It amplifies my biases. If I'm obsessed with SEO one week, the brain generates everything through an SEO lens. There's no external perspective — it's a mirror, not a mentor.
-
It works because I'm a builder who writes. If I were a writer who doesn't ship code, there'd be nothing for the agent to ingest. The pipeline runs on my repos, not my thoughts.
The lesson
I spent 15 years in marketing building systems for other people's knowledge.
The thing that finally worked for my own wasn't a better note-taking app or a smarter framework. It was giving the job to software that doesn't get bored, doesn't forget, and doesn't need a clean desk to start.
The bottleneck in writing isn't writing. It's connecting what you already know.
The specific tools (Next.js, Supabase, whatever) don't matter. But the architecture does: a system that reads your actual work, extracts the pattern, and hands you a draft — that's worth more than any writing course or content calendar I've ever seen.
What's your knowledge debt look like? The gap between what you've built and what you've written down.