I’d been using AI the way most people do: as a very fast research assistant. Ask a question, get an answer, close the tab. Start over tomorrow. I used custom GPTs at work. I had Claude and ChatGPT and Gemini in rotation depending on the task. All useful. But every conversation started from zero. The AI never knew who I was, what I was working on, or what I’d decided last Tuesday.
I’d gotten a taste of something better with Google’s personalization. Gemini picks up on your search history, your YouTube habits, your web behavior, and starts to feel like it knows you. That was exciting. But it was passive knowledge, things Google already had. I couldn’t shape it. I couldn’t tell it what mattered to me or correct it when it was wrong.
Then I tried something different, and in about eight days I accidentally built what I can only describe as a second brain.
I’m a marketer, not an engineer. These days I run marketing for a nonprofit. My days are meetings, Slack, Google Docs, and email. I’m not writing production code for a living. But I am drowning in context: brand guidelines, campaign performance, board prep, team decisions, research findings, revenue data. The gap between what I know and what I can hold in my head at any given moment has been getting wider for years.
What I built closed that gap. It started with a single text file.
The file that changed everything
It actually started with Claude Cowork, Anthropic’s collaborative interface. I was pulling in documents and having it create things I’d never built before, anything beyond a basic Word doc. I was working in a standard folder on my Mac, just files and subfolders, and even that was new territory. Then I stumbled into a thread about using Obsidian as a file store for AI work, and something clicked: it’s all Markdown. Plain text files. I’d tried Obsidian before as a note-taking system and found it obtuse. But with Claude as the one actually reading and managing the files, Obsidian stopped being a note-taking app I had to learn and became infrastructure I could build on. It took the basic folder I’d started with and made it something much richer.
That led me to Claude Code, Anthropic’s command-line tool, which took things further. Code works inside your actual file system, unlike a chat window. It reads your documents, remembers your preferences, operates on real files. But the important part isn’t the tool. It’s a file called CLAUDE.md.
CLAUDE.md is a markdown file that sits in your project directory. Claude reads it at the start of every conversation. You can put whatever you want in it: who you are, what you’re working on, how you like to communicate, what mistakes to avoid. It’s not a prompt. It’s not a system instruction someone else wrote. It’s your own accumulated knowledge about how you work, written in plain English.
Mine started small. A few lines about my role, my email accounts, a note about our brand voice. But every time Claude and I worked on something and I corrected it or refined an approach, I’d add that to the file. Brand rules went in after a few rounds of getting the tone wrong. Workflow preferences went in after I got tired of re-explaining them. Over days, this file grew into something that felt less like instructions and more like institutional memory.
CLAUDE.md turned out to be just the beginning. As the system grew, it sprouted more files: a MEMORY.md index pointing to topic-specific files (brand positioning, email strategy, revenue data, research notes), a STATE.md tracking active work so Claude knows what I’m in the middle of, and individual memory files that Claude creates and updates as we work together. There are about twenty of these now, covering everything from our Klaviyo email setup to our Google Workspace conventions to how I like my writing to sound. It’s a file system that teaches Claude my entire working world.
That’s when I realized how different this was from every other AI tool I’d used. ChatGPT and Gemini are conversations. This was accumulating into something else.
What actually changed
Looking back, there were three shifts. I didn’t plan them. Each one just happened because the previous one made the next one obvious.
The first was persistence. Most people are stuck in the ask-and-forget loop. You talk to AI, you get an answer, the context evaporates. With Claude Code and a CLAUDE.md file, the context stays. Not because of some magic memory feature, but because you’re literally building a knowledge base that your AI reads every time it starts working. It knows your brand. It knows your team. It knows what happened in last week’s meeting. You stop re-explaining yourself.
The second was making the knowledge base mine. This is where it splits from custom GPTs. I’ve used custom GPTs at work. They’re useful, but they’re someone else’s curriculum. The knowledge base is defined once, maintained by whoever built it, and you operate within those bounds. My system is the opposite. The corpus changes every day because I’m the one changing it. A decision made in a meeting on Tuesday is in the memory by Tuesday afternoon. A lesson learned about our email platform sticks around for next time. It’s not a reference document. It’s a living record of how I work.
The third was going from read-only to read-write. This is the part most people haven’t considered. Custom GPTs and chatbots answer questions. My system does things. It syncs meeting notes into my knowledge base. It extracts action items and decisions automatically. It sends me a morning briefing that pulls from Slack, my calendar, revenue data, and my own notes. It publishes reports to Google Drive. It runs audits on our email platform. It doesn’t just know about my work. It participates in it.
How small automations compound
None of this was planned. I didn’t sit down and architect a system. I solved one problem, and solving it revealed the next one.
It started with Obsidian, a note-taking app that stores everything as plain text files. I wanted a central place for meeting notes, so I set up a sync from Granola (my meeting recorder) into Obsidian.
But once meeting notes were flowing in automatically, I realized I was still pulling out action items manually and tracking them in my head. So I built an extraction step: Claude reads the meeting notes and identifies who committed to what, with priorities and due dates, formatted so Obsidian can query them.
Now I had action items in one place, but I wasn’t checking them consistently. So I built a dashboard, a single page showing everything that’s open, sorted by priority and date. That became my home screen.
Then decisions started getting lost the same way action items used to. So I added a decision log. Same pipeline, different output. Decisions extracted from meeting transcripts, queryable, linked back to source.
Then I wanted to start each day oriented instead of scrambling through Slack. So I built a morning briefing that synthesizes overnight Slack activity, today’s calendar, revenue and email campaign data from Shopify and Klaviyo, and open action items into a single email that arrives before I’ve made coffee.
Each of these took hours, not days. Each one was simple on its own. But together they created something that feels different from any individual AI tool. The meeting notes feed the action items. The action items feed the dashboard. The dashboard and action items feed the briefing. Each piece makes the others more useful. That’s the part I didn’t expect.
Once I had enough of these workflows, a pattern emerged: I was explaining the same multi-step process to Claude repeatedly. So I started writing skills, which are basically saved recipes. /triage runs my morning and evening ritual: pull reminders, show priorities, flag what’s overdue. /sync-meetings runs the full Granola-to-Obsidian pipeline. /weekly-review synthesizes the whole week. I have about fifteen of these now. Instead of explaining what I want each time, I type a slash command and the whole workflow runs. It’s the difference between cooking from memory and having the recipe written down, except the recipe also does the cooking.
How it all fits together
Here’s what the system actually looks like. Every piece on the left feeds into the brain in the middle, and the brain produces what shows up on the right.
The left side is where information lives. The middle is where it’s processed and remembered. The right side is what surfaces to me without asking. Most of the arrows run automatically on schedules — cron jobs on a small cloud server that cost about five dollars a month.
What a Tuesday looks like now
I wake up to a briefing email. What happened in Slack overnight, what’s on my calendar, where revenue and email campaigns stand, which action items are overdue. I didn’t ask for it. It just runs.
I open my Obsidian dashboard. Every open task across every meeting from the past month, sorted by priority. I knock a few out before my first call.
During meetings, Granola records and transcribes. After the meeting ends, my system syncs the notes, extracts action items, pulls out decisions, flags anything relevant to active work, like if someone mentioned the board deck or a campaign I’m running.
When I need to draft something, whether it’s an email, a strategy doc, or a board update, Claude already has the context. It knows our brand voice, our messaging framework, our recent decisions, and our revenue numbers. I’m not pasting background into a chat window. It’s already there.
End of the day, I do a quick triage. What got done, what rolls to tomorrow. Friday afternoon, a weekly synthesis: what was decided this week, what commitments are coming due, what changed.
I’m not doing less work. I’m losing less of it. Decisions don’t disappear into meeting transcripts nobody re-reads. Action items don’t fall through cracks between tools. Context doesn’t evaporate between conversations.
Starting from zero
You don’t need to build all of this. You don’t even need most of it.
Start with the memory layer. Create a CLAUDE.md file, or whatever your AI tool’s equivalent is, and write down who you are, what you’re working on, and how you like to work. Update it as you go. This one thing, just giving your AI persistent context about you, changes the experience more than any prompt trick or model upgrade. It’s the difference between talking to a stranger every time and picking up where you left off.
Then solve your most annoying daily friction. For me it was scattered action items. For you it might be something else: manually summarizing meetings, digging through Slack every morning, reformatting the same report every week. Pick one. Automate it. Live with it for a few days.
Then watch what it reveals. Once meeting notes flow automatically, you’ll notice decisions aren’t being tracked. Once action items are in one place, you’ll notice you’re not checking them often enough. Each solved problem lights up the next one.
The temptation is to plan the whole system upfront. Don’t. What I have now looks intentional in retrospect, but it wasn’t. It emerged from solving one small problem at a time, each one building on the last. The architecture came from the work, not the other way around.
What’s different now
The biggest change isn’t productivity, though that’s real. It’s how I relate to AI as a tool.
I used to treat it like a calculator for words. Pick it up, use it, put it down. Now it’s closer to a colleague who’s been sitting in every meeting, reading every document, keeping track of everything I’ve said I’d do. It doesn’t replace my judgment. It extends my memory.
That’s what I mean by “second brain.” Not a fancier chatbot. A system that knows my work, grows with it, and works alongside me, in the background, compounding while I focus on the parts that actually need me.
Eight days. A text file. And I’m still building.
P.S. Of course, Claude wrote most of this after we interviewed each other.