My second brain has been rebuilt three times, and each version shows where AI was at that moment.
The first version came after I’d piled up more than 100 voice memos in Apple Notes and never opened them again. So I built an automation in Make.com. I talked into Wisprflow on my afternoon walks, the transcript went to my email, and AI processed it and filed the idea in Notion. It worked, but I had wired every step by hand, and the AI did one job in the middle of a chain I designed.
The second version came after I read Andrej Karpathy’s LLM Wiki method. I adapted it with Claude Code routines. Now I talk into an iPhone shortcut, Claude picks it up in the cloud, and my ideas turn into linked pages in a wiki the AI looks after itself.
What I noticed, though, is that this setup is hard to build and maintain. It’s not realistic for most people who just want to talk to an AI and have everything handled for them.
Now you can get the same kind of system by pasting a single prompt into Grok Bot and letting a small team of bots set up the folders themselves.
As AI gets smarter and more capable over time, the process required to run the second-brain setup I used to have becomes easier.
Today’s guest post is from Yevheniya, who writes Mother Using AI: AI from zero to advanced, one real job at a time. Yevheniya was a software engineer before this, and every guide starts the same way: test the tool on an actual job, then write down every prompt and screenshot so you can follow along.
If you want to explore what Yevheniya writes about, start here:
In this post, Yevheniya builds a chief of staff for everything saved and never opened. It’s an OpenAI dot plus a team of three Grok bots, all working on one empty Obsidian folder. The setup starts with two links and ends with seven linked pages, an answer that shows its sources, and a note waiting every morning.
What I like most is that the post keeps what broke. One bot quietly set up a second night shift. The bot meant to check everything never ran at night, because nobody was awake to approve it. When it finally ran, it caught another bot citing a source that didn’t say what the answer claimed.
Here’s Yevheniya.
Hello 👋🏻
Stephon Anderson reads these notes and has left six comments under them since the summer. In August he wrote this under one about saving tutorials and never building:
“There’s something about the ritual of collecting resources that feels productive without requiring anything real from you.”
I read it twice, because I have a folder of saved posts from 2024 that I have never opened.
Saving was never my problem. I can save anything. I just could not find it again, and the AI I used every day did not remember any of it either.
Every new chat started at zero. I explained my work, my readers and my notes again, then closed the tab, and the next morning I explained it all over again.
People with money solve this with a person. A chief of staff who files everything, remembers what was decided last month, and tells you on Monday what is waiting for you. I never had one of those.
This week I read how three people on X built one with AI. So I built two, side by side, on the same folder, and took a screenshot of every step.
How Three People on X Built Theirs
The post that pulled me in was from Argona on October 18. It has been seen almost half a million times. Three Grok bots share one group chat. One collects everything you drop, one writes the pages, one checks the pages every night for things that disagree.
Almost half a million views in three days.
On his post, the line I keep coming back to was this:
“a second brain can sound sure while remembering the wrong thing”
Additionally, Avid set one up with one Obsidian folder and thirteen rules.
One of them is my favorite rule I have read this year: “nothing in the vault posts, pays, or sends. the final yes stays human.”
He wrote thirteen rules, and rule nine is the one I copied.
Mr. Buzzoni runs six bots between 23:30 and 06:00, while the house is asleep. His last night’s run: 7 sources in, 11 pages written, 8 flags raised.
What stuck with me was one rule he wrote himself: nothing skips the last check.
All those three use cases are built on the same idea, which Andrej Karpathy wrote down in a short file he calls the LLM Wiki. It works in three layers:
Your sources stay untouched. Articles, PDFs, links, voice notes. The AI may read them and never edits them.
The AI keeps a set of pages about them. One page per idea, linked to each other, like a small Wikipedia about your own work. Each claim points back to where it came from.
A check runs on a schedule. It looks for pages that disagree, claims that went stale, and pages nothing links to.
When you ask a question, a good answer is saved as a new page. The brain grows from use over time.
What I Built: One Folder, Two Apps, Four Helpers
So I decided to build one myself. Here’s the whole thing in a single picture:
The dot is the front door. The Grok team is the library behind it. Obsidian is the window I look through.
There are five parts in my setup, let me explain:
1. What a Dot Is
A dot is OpenAI’s always-on assistant inside ChatGPT. It launched at the end of September. It has its own computer in the cloud, it can also work on your own Mac if you connect it, and it keeps working on a job after you close the chat. You can also put it on a schedule. Mine came with a name, Alfred, and a small pixel face.
2. What Grok Bot Is
Grok Bot is xAI’s desktop app for bots that work for you. There is one main bot, and you can ask it to create others, each with its own job and its own memory, and put them in group chats together. They can run commands on your Mac when you approve them, and they can run on a schedule too.
3. An Empty Obsidian Folder, in One Minute
Obsidian is a free notes app. Everything in it is a plain folder of text files on your own computer, which is why both AIs can read and write to it.
Download Obsidian from obsidian.md and open it.
Click Create next to “Create new vault.”
Name it Second Brain. Click Browse and pick your Desktop.
Click Create.
You get an empty folder with one welcome note in it. Leave it empty. The bots build the rest.
4. Grok Bot Builds the Library: Three Bots, One Prompt
I opened the main Grok Bot chat and pasted Argona’s prompt, with two changes: I named the three bots, and I told it where the folder is.
Use this prompt:
“Build me a second brain with three bots in one group chat: Reader, Librarian, Checker. Keep my sources untouched in raw/, maintain wiki/ yourself, file good answers back as pages and run a nightly check for contradictions. Everything should live in my new Obsidian vault at ~/Desktop/Second Brain on this Mac. Show me the structure first. Do not create anything until I confirm.”
The last two sentences matter more than the rest. It showed me the plan and waited.
I clicked Yes, build it. Then Grok Bot asked if it could run a command on my Mac to make the folders.
There are three buttons on that card. Allow once, Always allow, and Never.
“Always allow” lets every bot run commands on your whole computer from then on. I pressed Allow once, every single time. In the first hour, that was eleven times. It is boring. I would still do it again, at least for the first week.
About eight minutes after I pasted the prompt—most of it spent waiting for me to press a button—the team was there.
Now the bot and the folder are ready.
5. The Dot Becomes the Front Door
The Grok bots live in an app on my Mac. The dot lives inside ChatGPT, the app I already open all day. So the dot gets one job: catch everything I throw at it during the day, and write me a short note every morning.
Here’s the prompt I used:
“I’m building a second brain in a new Obsidian vault at ~/Desktop/Second Brain on my Mac. A Grok Bot team (Reader, Librarian, Checker) looks after wiki/ and checks/. Your job is the front door. 1) Anything I send you during the day (links, voice notes, screenshots, PDFs), save it untouched as a new file in raw/ with today’s date in the file name. 2) Every morning at 8:00 Istanbul time, read wiki/log.md and the newest file in checks/, then write daily/YYYY-MM-DD.md: what was added yesterday, any contradictions waiting for me, and one question worth asking next. Never edit wiki/ or checks/. Show me the plan first. Do not create anything until I confirm.”
Change Istanbul to your city. If you only use one of the two apps, delete the Grok sentence and the dot still works as a front door on its own.
I said yes. Four minutes later:
Two Links In, Seven Pages Out
Now that the whole setup is ready, we can proceed to execute this.
I sent the dot two links: Karpathy’s LLM Wiki file and Argona’s post.
The first thing I learned: the dot did exactly what I said. “Save it untouched” meant it saved the link and nothing else. Two tiny files with a web address in each.
Then, my Bot Reader carried the work. I did not have to tell it anything. It was already watching the folder.
Here’s what it noticed:
Two links became seven pages: one page for each source, and five pages for the ideas inside them.
Each claim has a link back to where it came from.
The Bot Reader also told me what it could not get.
The X post links to a longer article, and only the title of that article came through. It wrote that on the page instead of guessing.
The Moment I Trusted It
While Reader was still working, I asked the group a question:
“@Librarian I’m not technical and I have one hour a week. Based on what’s in the wiki, what is the smallest version of this second brain I could start with, and what should I not bother with yet? If the answer is worth keeping, file it in wiki/answers/.”
Librarian had looked at the folder a few minutes earlier, when it was still empty. So it said the wiki had nothing on this yet, and it would not save an answer with nothing to point to.
Then Reader stepped in, in the same chat, and told it the pages were there now.
This is the screenshot I would show someone who is afraid of AI making things up. One bot said, “I don’t know yet.” Another bot corrected it in front of me. I could see the whole thing happen.
After I approved it, Librarian wrote the answer and saved it as a page. The answer: drop in a few sources you care about, ask questions, and skip the extras for now. In the graph on the right, you can see it is already linked to the other pages.
What Went Wrong: Two Night Shifts
Checker Bot has one job. Every night, it reads the pages Reader and Librarian wrote and looks for pages that disagree, claims that went stale, and pages nothing links to. It’s the only bot checking the other two’s work.
When the main Grok Bot built the team, it scheduled Checker’s nightly check for 3:00. Then Checker set up a second nightly check for itself at 23:00.
So Checker had two night shifts. It would read the whole wiki at 23:00, read it again at 3:00, and leave two reports in the checks folder. The dot only reads the newest report each morning, so the 23:00 report would never reach me.
I only noticed because I read every message. I told Checker to keep one, and it deleted its own, which left the 3:00 check. Keep that time in mind, because it comes back later.
If you build this, open each bot’s chat once on the first day and read what it says it did.
The First Morning Note
The dot’s second job is the morning note. Every day at 8:00, it reads the wiki’s log and the newest Checker report, then writes me a short note: what was added yesterday, any contradictions waiting for me, and one question worth asking next.
I didn’t want to wait until the next morning to see if it worked, so I asked the dot to run the note once, right away, as a test.
It is careful to the point of being funny. The note looks at yesterday, and yesterday was October 20. Everything in the wiki was written that afternoon, on October 21, so the log had nothing from the 20th. The dot wrote that nothing was recorded, and added that this “does not establish that nothing was added that day.”
The contradictions part came back empty for a similar reason. Checker’s first night shift at 3:00 was still hours away, so the checks folder had no report yet. The dot said it had no evidence either way and didn’t guess.
Five Mornings Later
I am writing this part on October 26.
The dot did its job every day. On October 22, 23, 24 and 25 there was a new note in the daily folder a few minutes after 8:00. I did not open ChatGPT once to make that happen.
Here is what I wanted from it, and what I got:
What changed yesterday. On the 22nd it listed every page the bots had written. On the 23rd it said nothing new had come in, and it did not pad the note to look busy.
What is waiting for me. This is the line I care about most. And every morning it said the same thing: the checks folder is empty, so it cannot tell me if anything disagrees.
One question. A different one each day. On the 23rd: What is one real question you want the two captured sources to help you answer this week?
The empty checks folder was the second thing that went wrong. Checker never ran at night, not once.
When I opened its chat, the reason was sitting there. Checker has to ask permission before it reads files on my Mac, and at 3:00 nobody was awake to press Allow once. The request waited, then expired. Five nights in a row.
So this morning I ran it by hand. I pressed Allow once three times and it wrote its first report.
No contradictions yet, which makes sense with two sources. But it caught something I had missed. Two lines in Librarian’s answer cite a source that does not say what the answer says. One of them claims “the three-bot group chat is already running” and points to Argona’s post, which is about Argona’s setup, not mine.
That is the job I wanted it for. A second brain can sound sure while remembering the wrong thing, and this one now tells me where.
There are two ways to make it run on its own. Leave the computer awake and press Always allow, which gives every bot that permission for good, so I am not doing it yet. Or move the check to an hour when you are at your desk to press Allow once. I am trying the second one next week.
Dot or Grok Bot: Which One Does What
After a few days with both Dot and Grok Bot on the same folder, here’s what I’ve learned:
Both are paid. A dot comes with ChatGPT Pro, which starts at $120 a month, and Pro dots are not offered yet in the UK, Switzerland or the European Economic Area. Grok Bot starts at $30 a month, on Cursor’s Premium plan. Those are the prices on both official pages on 21 October 2026.
If neither is possible right now, Obsidian is free, and the two prompts above also work in other AI apps that can read a folder on your computer. You lose the schedule and the morning note. You keep the habit, which is the part I care about.
Your Task This Week
Pick one topic you actually work on. One. Not your whole life.
Three that work well:
A project at work. Meeting notes, the client’s emails you forward to yourself, the reports you keep half-reading.
Something you are learning. The videos, threads and articles you saved about AI, or anything else you keep meaning to study.
Your writing. Ideas, drafts, other people’s posts you liked, what your readers said.
Then:
Create the empty vault (Part 3).
Paste the Grok Bot prompt from Part 4. If you do not have ChatGPT Pro, this is all you need. Grok Bot alone does every job in this post, and you can drop files into the raw folder yourself.
If you do have a dot, paste the prompt from Part 5 too, so it catches things from your phone and leaves the morning note.
Drop in three things you saved and never opened.
Ask one question about them.
And then use this prompt in the group chat:
“@Librarian I’m working on [your topic]. Based only on what’s in the wiki, what are the three ideas that come up most, and which one should I act on first this week? Cite the page for each one. File the answer in wiki/answers/ if it is worth keeping.”
What you want the bots doing after that: Reader turns each new thing you drop into pages, Librarian answers your questions from those pages, Checker tells you when something disagrees or has no source. You add things and ask questions. That is your whole job.
Final Thoughts
The folder on my Desktop has 23 files in it now. Apart from Obsidian’s welcome note, the bots wrote every one, and I have read every one.
Stephon called collecting “productive without requiring anything real from you.” For the first time, my saved things are being read by something. Every morning it tells me what changed and asks me one question back.
I still have not answered the first one.
Checker needing 'Allow once' three times before its first report is the detail that tells the real story. The system works, but the permissions friction is the tax on every new agent you add. Did you end up giving it standing permission, or is the ask-every-time step doing useful work as a review gate?
rule 9 is pretty much how I run my own bots, nothing posts or pays without my yes. wild that the Checker caught another bot citing a source that didn't say it