My AI Brain Has 4.1 Million Words. Here's How I Built It.

The First Session: 90 Minutes to a Working Brain

I started with Claude Code on a Sunday afternoon.

No plan. Just frustration with bland AI outputs.

I asked: "Can you scrape my LinkedIn posts and store them in a way that helps you write like me?"

Claude didn't push back. She asked questions.

"What else can I scrape?"

I sent her: - YouTube videos I'd appeared in - Podcast transcripts - My Medium articles - My Bear notes (a personal note-taking app I use constantly) - My Substack posts - My tweets - My Startmate blog articles

She asked for a plan. I said yes. She outlined what she'd scrape and how she'd structure it.

Then I just kept saying yes.

Gave her access to: - My Granola meeting notes (every founder call, every investor chat) - My email archives - My connection data - My LinkedIn messages (31,000+ DMs)

Ninety minutes later, I had a SQLite database with everything.

Every piece of content I'd ever created, searchable in milliseconds.

The Token Problem (And How Claude Fixed It)

The first version worked.

But it was expensive.

Every time I wanted to write something, Claude had to load the entire context. 826,000 words. Millions of tokens.

I literally asked: "How would you structure things to make them more efficient from the token perspective?"

Claude restructured the entire brain.

Instead of one massive file, she split it into: - worldview.md - My positions on 26 topics (fundraising, leadership, hiring, community, startups) - michael-batko-voice-guide.md - Writing style guide with DO/DON'T examples - best_of - A curated table of my 276 best pieces across 20 topics - FTS5 full-text search - So Claude could query only what she needed

Now when I write, Claude doesn't load everything. She searches for relevant context, pulls only what matters, and writes.

Token cost dropped by 80%. Speed increased dramatically.

If You Only Had 3 Sources

People ask: "Where do I start?"

If you could only scrape three sources, here's what I'd choose:

1. Meeting notes (richest in context)

Your meeting notes contain how you explain things to different audiences, your actual advice (not polished for public consumption), your questions (which reveal your thinking), and your reactions to ideas.

If you use Granola, Otter, Fathom, or even just Google Docs - scrape it all.

2. Personal notes (if you use a notes app)

I use Bear. 924 notes.

These are unfiltered thoughts. Ideas before they're polished. Questions I'm wrestling with.

If you use Notion, Obsidian, Roam, Apple Notes - this is gold.

3. Emails (especially sent emails)

Your sent emails show how you communicate with different people, what you prioritise, and your voice when you're not performing for an audience.

Start with these three. You'll have enough to make AI sound like you.

The Technical Build

Here's what the Batko Brain actually is:

Database Structure: - SQLite database (batko.db) - 57 tables, 103,000+ rows - FTS5 full-text search for instant queries - Indexed by source, date, topic, and content type

Sources Breakdown:

Weekly Sync: runs Sunday 8pm via launchd, scrapes new content, rebuilds the FTS5 index, and emails me a summary.

Monthly Deep Analysis: topic trend analysis, voice consistency check, best performing content identification, and worldview updates based on new writing.

It Writes Better Than I Do

Here's the creepy-good part.

Every single time I write a blog post now, Claude uses my voice automatically, links to my batko.ai blog posts (and dynamically chooses the highest viewed ones), structures arguments the way I would, uses phrases I actually use.

And honestly? It writes better than I would write it.

Not because it's smarter. Because it has access to every good sentence I've ever written.

It can pull my best metaphors, my clearest explanations, my strongest examples - and combine them in ways I wouldn't think of in the moment.

You're not replacing yourself. You're giving AI access to your best self, every time.

What's Still Broken

I'm Australian with an Austrian accent. Yes, both.

Whispr Flow, my transcription tool, struggles.

It mishears words. Misinterprets sentences. Sometimes produces complete rubbish because it doesn't capture my accent.

So voice-to-text still needs manual cleanup.

The brain is only as good as what goes in. And if the input is garbled transcription, the output suffers.

Everything else? Works. The weekly sync runs automatically. New posts, notes, and meeting transcripts flow in without me touching anything.

But transcription is the weak link I haven't solved yet.

How to Build Your Own (90 Minutes)

Step 1: Choose your 3 sources - Meeting notes (Granola, Otter, Fathom, or Google Docs) - Personal notes (Notion, Obsidian, Bear, Apple Notes) - Sent emails (Gmail, Outlook)

Step 2: Export everything - Most tools have export functions (Settings > Export Data) - For LinkedIn: use their data download tool - For email: IMAP access or export to .mbox

Step 3: Ask Claude to structure it - Open Claude Code - Say: "Scrape everything I've created and structure it so you can easily access it" - Let Claude ask questions and keep saying yes

Step 4: Build the voice guide - Ask Claude to analyse your writing patterns - Extract your most-used phrases, sentence structure, preferred transitions - List words you never use

Step 5: Create the worldview doc - 10-20 topics you write about most - Your position on each (with examples from your writing) - Link to source content for each position

Step 6: Automate the sync - Weekly cron job to pull new content - Auto-update the database - Email yourself a summary

I spent 90 minutes on the first version. I spend 10 minutes a month maintaining it.

You don't need 826K words. You don't need 57 tables. You need your best writing (50+ pieces), a voice guide (10-15 rules), and a worldview doc (10-20 positions).

That's enough to make AI sound like you.

The rest compounds from there.