The first week is three commands and a habit. Everything after it, you can just say out loud.

If you set up your second brain and read the first-week post, you learned three commands and a habit. That's the right place to start. It's enough to feel what the thing is for.

But there's a second layer the install never hands you, and it's the part that makes the brain stop feeling like software. You can talk to it. Not in commands. In plain sentences, the way you'd tell a person what you need.

I understood this the day I finished a long work session and, instead of closing the window, told the brain I was done. Let's close this session. It read back over the whole conversation, pulled out the decisions I'd made and the reasoning behind them, filed the loose ideas where they belonged, and sorted the to-dos. I hadn't opened a single file.

Here are five things you can say to your second brain, and what each one does. They follow the shape of an ordinary day. None of them are commands. There's no format to get right and nothing to memorize. They're sentences.

Subscribe now

"Good morning."

You say good morning. The brain takes it as the start of your day. It checks in on how you're actually doing, then goes to your own to-do lists and pulls the one to three things that matter most today, and helps you name what you want the day to be about.

It's a small thing, and it replaces the usual morning spent opening four apps to reconstruct what today even is. You say good morning, and the day's shape comes to you instead.

"The meeting is done."

Say this after a call. The brain opens your meeting note, reads the whole thing, and pulls out the action items. Then it does the part that's genuinely tedious by hand: it separates the tasks that are yours from the tasks that belong to other people, flags the decisions that were left open, and drops your tasks onto your to-do list, written in plain words, with enough context that they still make sense to you next week.

Meeting notes are easy to take and easy to never read again. This is the step that turns them into something that actually moves.

"What have I written about this before?"

This one is a question, and you can ask it in any words you like. What did I decide about hiring the last time I thought it through? What does my journal say about the way I handle conflict? What have I written about this client before?

Most notes apps are a place writing goes to be forgotten. You put something in and never open it again. This is the opposite. Your second brain reads across everything you've ever written, years of it, and answers you. The answer almost always holds a connection you wouldn't have found on your own, because it can read all of it at once and you can only hold a little at a time.

It's the thing people try last and wish they'd tried first.

"I keep going back and forth on this."

When you're stuck on a decision, just say so, out loud, in whatever words come. I'm torn between two things. I can't tell if this is the right call. Should I.

You don't have to ask for help. The brain hears the shape of a decision in how you're talking, and it brings in your panel: a set of advisor voices you chose when you set things up, each one thinking from a different angle. A founder. An investor. Someone who has already run the thing you're about to run. They won't all agree with each other, and that's the point of them.

What that "hearing" actually is: a rule the install put in place. It watches for decision-shaped language and opens the panel template you set up. The voices are your roster.

The first-week post told you that you could ask the panel before big calls. Here's what it didn't say. You don't have to remember to. The brain is listening for the moment you're deciding, and it brings the room to you then.

"Let's close this session."

When you finish working with the brain, don't just close the window. Tell it you're done. Let's close this session. I'm wrapping up. That's it for today. Any of those will do.

It reads back over the whole conversation and files what mattered while you were busy thinking about the work. It keeps the decisions you made, and the reasoning behind them, so the version of you in three months knows why. Sends the ideas worth writing about to the right folder. Sorts the to-dos onto the right list. And if your mind changed about something while you talked it through, it saves the shift, so months later you can see the whole arc and not only where you ended up.

Here's why this one matters more than it sounds. The thinking happens in the conversation. The good line, the decision, the thing you realized halfway through saying it. Normally all of that stays inside a chat window you'll never open again. Saying let's close this session is the difference between a brain you talk to and a brain that keeps what you said.

Four words, at the end of the day. It's the smallest habit on this list and the one that compounds the most.

That's it

None of these are commands. There's no format, nothing to learn, no special words. You're talking to the brain the way you'd brief a good colleague who happens to never forget anything: in the morning, after a meeting, when you're curious, when you're stuck, and at the end of the day.

The first week is for the three commands. Almost everything after the first week is this. Pick one sentence and use it tomorrow.

New here? Install your own second brain:

curl -fsSL https://raw.githubusercontent.com/mycelium-hq/ai-brain-starter/main/bootstrap.sh | bash

Windows: irm https://raw.githubusercontent.com/mycelium-hq/ai-brain-starter/main/bootstrap.ps1 | iex

Then: claude and type /setup-brain.

Already installed: run git pull and ask Claude to update the ai-brain-starter skill.

I'm Adelaida. I write about consciousness and AI, and I help founders install their own version of this at diazroa.com. Building all of this from Bogotá.

If you run more than one Slack, you already know the friction.

You have one for your company. Another for the contractor team. A third for the friend group you trade voice memos with on Sundays. Each one has its own context, its own backlog, its own threads you haven't read.

For me it's Onde, Mycelium, and three personal workspaces. The math is constant. Which Slack am I in. Which conversation is this. Did I already reply over there.

I wanted Claude to help. Read across all of them. Surface what mattered. Draft replies in the right voice for each. Post when I said the word and not a second before.

The Slack connectors out there couldn't do it. One workspace per server instance. No safety layer on writes. The Anthropic reference implementation was deprecated.

So I built one.

Subscribe now

What you get

slack-mcp is an open-source connector that gives Claude eighteen tools across every Slack workspace you connect at once. One process, many workspaces, every send goes through a draft and a confirm before anything posts.

You can ask Claude to:

• Read the last fifty messages in #ops and pull what's still open

• Search across every workspace for the thread where the venue change came up

• Draft a reply and show it to you before sending

• React to a message with an emoji

• Edit one of your own messages with a before-and-after preview

Read tools mirror the channel to your vault on every call. So when you ask "what did Daniel say about pricing last week" three sessions from now, it's already there, searchable, no API call needed.

Why draft and confirm

Every write tool returns a draft ID. Nothing posts until you call confirm. The draft expires in an hour. You can only confirm it once.

This is the pattern I want around any AI agent touching a channel where real people read what posts under my name. Send a draft, see the preview, then commit. No "the model misread the thread and sent a half-finished sentence to the client Slack." That moment ages badly.

The trade is one extra tool call per send. The trade is worth it.

What else is in there

A prompt-injection scrubber that wraps known attack shapes (zalgo characters, fake role headers, fake code fences) instead of deleting them, so you see the attempt instead of inheriting it silently.

An audit log at every tool call. Tokens redacted. JSONL so you can grep it.

Three auth modes: cookie session (full power, occasional re-extraction), OAuth user token (permanent, no search on free plans), and bot token (limited, invited channels only). You pick.

How to install

Open Claude Code. Paste:

/plugin marketplace add adelaidasofia/slack-mcp
/plugin install slack-mcp@slack-mcp

Fill in tokens. SETUP.md walks the cookie extraction in about three minutes per workspace.

Then ask Claude to read your most active channel.

Why I built it

Same reason I built every other connector in the family. They all follow the same shape: multi-workspace where it makes sense, draft and confirm on writes that touch other humans, every read mirrored to a vault you own.

The future I want to live in is one where my AI reads everything I do, helps me move faster, and never sends anything I haven't signed off on. The connectors are how I get there.

The repo is at github.com/adelaidasofia/slack-mcp. MIT. Built on FastMCP. The full family of related connectors is linked from the README.

If you set it up and something breaks, open an issue. I read every one.

Last Tuesday I asked Claude to add a rule. Something small. The way I always do, mid-conversation, in passing, the way you'd tell a colleague "from now on let's just default to X."

Before I'd finished typing, the system flagged that I had a rule six weeks old that said the opposite. Not on the same surface. Not in the same words. Two rules, two folders, two months apart, in conflict.

I'd written both of them. I would have written a third. The system caught me before I did.

That's the small moment that made me write this post.

What's actually different now

The thing I built a few weeks ago was a memory: notes that connected, an AI that remembered who I was between sessions. The thing I wrote about two weeks ago had grown a cascade and a panel and a graph: it noticed what happened, filed what mattered, pushed back when I asked.

But all of that ran on me. I had to start the session. I had to ask the question. I had to remember to file the decision. The system would pick up if I did, and drop the thread if I didn't.

This week the system started showing me what changed without my asking.

Subscribe now

Three things your second brain can now do

I pushed all of this to the free repo yesterday. If you installed earlier, run git pull. If you haven't installed, the one-line bootstrap still works.

It notices what you said but didn't file. A new script watches your recent Claude Code sessions for sentences that pattern-match as a rule, a decision, or a lesson, and that you said out loud but never filed via /journal or /decision. Phrases like "from now on," "let's switch to," "the lesson is," "I keep doing this." It drops them into your vault as drafts you can come back to. You go through them at the end of the week. Most of mine were things I'd told Claude four times across three sessions and somehow never written down.

It notices drift. A second script tracks every file edited five or more times in the last 30 days. Repeated edits are how documents quietly rot. A study from Microsoft this April tested 19 frontier models across 52 professional domains and found that even the best ones corrupted about 25 percent of content over twenty round-trip edits. The average across all models was 50 percent. The system runs the script weekly, lists the top candidates with reasons, and asks you to look at them. It does not delete anything. It just shows you what's at risk.

It notices contradictions. A third script reads every rule file in your vault and looks for "always X" against "never X" on shared nouns. When it finds a clash, it tells you which two files and which two lines. There is an opt-in mode that uses a small language model to catch contradictions phrased differently ("default to A" vs "we never do A" in different vocabulary). Both scripts ship with test examples so you can verify the logic on your own machine before pointing it at your real vault.

Underneath: the system learned to tell different kinds of notes apart

Same week, the system grew a vocabulary. Until now everything was just a note that linked to other notes. Now there's a quiet difference between a fact, a step-by-step how-to, and an exception you don't want to repeat. Each one carries a small "last verified" date so the system knows when something is going stale.

There are connectors that pull straight into the vault from the places work actually happens: Slack, GitHub, Notion, Linear, Gmail. There are two readers that watch resolved GitHub conversations and resolved Slack threads, and turn them into how-to documents, the way a careful teammate would write up "here's how we actually do this now."

And there's a small private service that lets other AI tools read from and write to the same memory, locked behind a key, in case you want to point them at it.

A weekly meta-pass

There's a /sunday-review skill now that runs all of the above as a single weekly pass. Detects drift. Catches conflicts. Surfaces passive captures. Reports back in plain language. Sundays the vault reviews itself and tells me what it found.

I read the report Monday morning over coffee. Most weeks it surfaces three things I would have missed. Some weeks it surfaces nothing, and that's also useful: it means the past week ran clean.

What this is now, and what it isn't

The first post was about a memory you talk to. The second was about infrastructure that fills itself. This one is about the moment where the infrastructure starts noticing its own seams.

The repo is open source. Everything I described above is in there. Free to install, free to run, free to make your own. If you want to look under the hood, the README, SECURITY.md, and the CHANGELOG are where I'd start.

The same architecture extends into a private layer I install for clients who'd rather have it built for them, configured against their actual documents and their actual team, with the real-world connectors wired in. That's at diazroa.com if it's relevant. Most readers will be better served by the free repo.

What I'm noticing about this whole thing

When I started I wanted a system that remembered me. A few weeks later I wanted a system that worked without me. This week, for the first time, I have a system that lays out what shifted while I wasn't looking, ready for me Monday morning.

That feels like a different category of tool. Not a brain you query. Not infrastructure you maintain. Something closer to a colleague who does the boring part of being careful so you can do the interesting part of being alive.

Two weeks from now I expect to be writing about something I didn't see coming. That seems to be the cadence.

If you build something with it, send me what you built. I'd love to see.

Install: https://myceliumai.co/install

Already installed: git pull and ask Claude to "update the ai-brain-starter skill."

I'm Adelaida. I write about consciousness and AI, and I help founders install their own second brain at diazroa.com. Building all of this from Bogotá.

If you write on Substack, you've heard the advice. Post Notes. Post Notes. Post Notes. Growth lives in Notes.

You know it's true. The doing part is harder.

I didn't do it either. Not because I didn't have anything to say. The opposite. Ideas show up constantly. In a sales call when an objection cracks open a better way to pitch. In a strategy meeting where a pattern jumps out across three past companies. In the shower. Talking with friends. At 11 PM, in a long conversation about a chapter I'm writing.

Those ideas had a pattern. They appeared. They sat in a chat log, a meeting transcript, a voice memo, a half-written note, a Slack thread. And they died there.

The gap between having the thought and posting it was too wide. So I closed it.

Subscribe now

What you get

Every morning, three or four Notes are waiting for you in a folder. Pulled from things you actually said and wrote the day before. In your voice. Not invented. You read them, edit a sentence or two, hit publish.

That's the whole thing. Not AI-generated. AI-surfaced.

How it works

Two pieces, both free.

The first is a Substack connector for Claude. It lets Claude publish, schedule, restack, and pull your dashboard analytics, all from inside the chat. No browser, no copy-paste, no tab-switching.

The second is what makes it magic. A system that reads everything you've already written. Twelve years of journals if you have them. Meeting notes, chat logs, drafts, postmortems. It learns your voice. At the end of each work session, it scans your day, finds the three or four moments that had the shape of a Note, and drafts them in a folder for you to review.

By morning, the loop is closed. The connector publishes what the system surfaced.

You wake up to a week that is already a content engine. You didn't write a single Note from scratch

.

Why this is different

This isn't asking you to write more. It noticed you already did.

The longer you use it, the more it has to work with. Your past becomes the engine. Your voice becomes the moat.

The part I'm in awe of: the drafts are already in things I said years ago. They were just waiting, in folders nobody opens.

How to try it

The Substack connector alone: github.com/adelaidasofia/substack-mcp

The full system, including the part that surfaces your ideas every morning: github.com/adelaidasofia/ai-brain-starter

If you've never installed a Claude connector before, the README walks you through it. Nothing to compile, nothing to learn.

If it works for you, reply and tell me. I read every one.

The official Google connector for Claude has two problems.

It supports one account. And it dumps the full email body every time you search.

If you're a founder running two mailboxes — work and personal — that's already a dealbreaker. And if you're asking Claude to search your inbox ten times in a session, you're burning tokens on content you didn't ask for.

So I built my own.

google-workspace-mcp is an open-source MCP server that gives Claude 61 tools across Gmail, Calendar, Drive, Docs, and Sheets. Multi-account from the start. Token-efficient by design — search returns compact shapes (id, from, subject, snippet), and full bodies are opt-in.

Subscribe now

The architecture is simple. OAuth tokens live in the macOS Keychain, not in any file or repo. Every tool takes an optional account email that overrides the default — so I can ask Claude to search my Onde inbox for a client thread and my personal inbox for a flight confirmation in the same session, with the same tool.

Here's what a Drive search returns versus the raw API — just metadata, no 40KB document blob:

• id, name, mime type, modified date, owner

• Nothing else until you ask for it

The v2 update added Drive, Docs, and Sheets, which meant a re-OAuth for both accounts. I found three bugs in the process: apostrophes in search queries broke the Drive query syntax, the accounts index was not gitignored (emails would have leaked to the repo), and the tool count in the README was wrong. Fixed all three before pushing.

I run this for Onde — event planning for Latin American corporates — where Claude needs to pull meeting transcripts from Drive, calendar meetings, update budget sheets, and draft follow-up emails without me switching context between apps. It has become the connective tissue between Claude and everything Google

The full tool list (61 tools across 5 services):

Gmail (10): search, read, send, draft, reply, labels, archive, trash, send-as identities

Calendar (7): list calendars, list events, create, update, delete, free/busy, respond to invites

Drive (18): search, read, list folder, create folder, upload, move, rename, share, trash, permissions, comments

Docs (9): create, read, append, insert at index, find-and-replace, export, suggestions

Sheets (14): create, list tabs, add tab, read/write range, append, clear, batch read/write, named ranges, conditional formatting, data validation

Setup takes a few minutes the first time — most of it is GCP OAuth consent screen configuration. Each account you add after takes about 90 seconds.

The repo is at github.com/adelaidasofia/google-workspace-mcp. SETUP.md walks through the full GCP flow. MIT license.

If you're building a Claude setup that touches Google at all, this is worth the setup time.

Here are the four things that actually make your ai-brain-starter useful. Three commands and a habit. If you do them in the first week, you'll feel what it's for by day eight. If you skip them, the brain sits in your Obsidian folder doing nothing.

1. Map your whole brain

The first thing is to run /second-brain-mapping on every single file in your vault. All of it.

Your brain already has thousands of notes in it: journal entries from years ago, article clippings, book notes, half-finished ideas. Mapping reads every file and figures out what's connected to what. Names, concepts, people, places, decisions. It builds a map you can query.

Why it has to be everything: a partial map has blind spots. The magic happens when you ask a question today and the brain pulls a note from 2019 that answers it. That only works if 2019 got mapped. Skip files and you get blind spots in the places you forgot you had.

Run it in small batches

If your vault has a few thousand files, maybe ten thousand, don't feed it all in one go. It's like dumping a library on someone's desk: the AI either collapses or skims half of them. Small batches means one chunk at a time (~100 files), read properly, finished, then the next. You end up with every file mapped cleanly, not half-done with gaps. Bonus: if something breaks, you redo one small piece, not the whole thing.

The skill stages this for you. Your job is to be patient on the first big run. It takes time. It's worth it.

Subscribe now

Optional advanced move: MiniMax

Tell your brain to use MiniMax for mapping. Cheaper model, does the boring work just as well, big savings on every mapping run. Setup means signing up for MiniMax and pasting an API key into your settings.

Why it matters: mapping your whole vault is a lot of reading. Using the smartest AI model for it is like hiring a Harvard professor to file papers. It works, but you're paying for brilliance you don't need.

The default model works fine if you skip this. You'll just pay more in tokens.

2. Journal every night

Second thing: journal every night. Run /journal before bed, or whenever your day ends for you. Write what happened, what mattered, what you're sitting with.

Why every night: your brain can only show you patterns if it has data to look at. One journal entry is a moment. Thirty is a picture of your month. A year of nightly entries is the most honest record of your life you will ever have.

The real payoff is weekly and monthly insights. When you run /weekly or /monthly, the brain reads your recent entries and shows you what's been firing: what's been on your mind, what's been repeating, what's shifting. You can't see your own patterns from inside. Someone, or something, reading them back to you is the only way.

If you skip nights, the pattern has holes. Missing a day is fine. Missing a week means the weekly insight misses most of the week. This is the habit that compounds

3. Ask the panel

The third is a mindset, not a command. Before you take an AI answer and run with it, especially for something strategic, ask: what would the panel think?

The repo comes with an advisory panel, a list of named voices (founders, writers, critics, people with different lenses). When you ask the brain to run a panel on a decision, it brings in three to five of them and makes sure at least one dissents.

Why it matters: when you're excited about an idea, you can't see its blind spots. A good friend pokes holes before you spend money or time on something. The panel is five of those friends, on demand, with different specialties. Not everyone has five brilliant friends available at 2 AM.

Use it before any decision that costs money, time, or relationships. I've killed ideas I was in love with because one voice on the panel said something I couldn't unhear. That's the whole point.

That's it

Four things, mostly. Mapping, MiniMax (optional), nightly journaling, and the panel.

If you do the first three in your first week, you'll feel what the brain is for by day eight. If you keep the habit, the weekly and monthly insights start to feel a little uncanny. And if you use the panel before big calls, you'll make fewer bad ones.

The install is the easy part. This is the actually-using-it part.

A week ago I published a post about how I turned Claude Code and Obsidian into a system that actually knows me. 5,000 notes connected, 12 years of journals compressed, an AI that remembers who I am between sessions. I open-sourced the whole thing. People started DMing me screenshots of their vaults. Strangers asking how to adapt it for their work.

Then I kept building. And somewhere around day three, the thing I built stopped being a second brain. It started behaving like an operating system for optimizing how I think and work.

If you installed it after the last post, run git pull. It's changed a lot since then. If you haven't installed it yet, this is what you're building toward.

Nothing stays in the chat anymore

The biggest problem with version one was invisible: conversations still leaked.

You'd make a decision during a session. Surface an insight. Think of something worth writing about. Then the session would end, and all of it stayed in a transcript you'd never reopen. The vault was smart, but the workflow around it still had holes.

So I built a cascade that runs automatically every time a session ends. Nine things happen without you doing anything:

It saves journal seeds verbatim. Files writing ideas to the right draft folder. Routes to-dos to the right list. Logs decisions with the reasoning behind them. If your thinking shifted on something, it captures the shift so you can see the arc over time. Stages sharp lines as social posts. Tracks time. And if the session changed any rules or scripts, it checks that nothing broke.

Nine lanes. Every session. You never think about it.

Without the cascade, every session is a standalone event. With it, every session leaves something behind for the next one to build on.

Subscribe now

Meetings just process themselves

I used to take meeting notes and then spend 20 minutes afterward filing action items, updating contacts, and making sure nothing fell through the cracks. Now I say "the meeting is done" and the system runs a full enrichment cascade.

It pulls the transcript, adds a summary, a decisions table, and action items, then pushes changes to every file the meeting touched. Contacts get updated. To-dos split into your list and your team's. Strategy docs get revised. The decision log fills in. One sentence triggers the whole thing.

If you're someone who has four calls a day and ends every evening re-filing the same information across three apps, this is the part that changes your Tuesday. Not the graph. Not the panel. This. Twenty minutes back per meeting, every meeting, without touching it.

The graph shows you the shape of your own thinking

In the first post I mentioned the knowledge graph briefly. It was early. Since then I ran 22 extraction stages across the entire vault: 5,000+ files, 9,700 nodes, 44,500 edges, 1,096 communities.

The top node in my graph is "Money." It connects to 1,326 other concepts. "Colombia" connects to 1,233. "Love" to 1,053. These aren't tags I assigned. They're patterns the system extracted from 12 years of writing. I didn't know my vault had that shape until the graph showed me.

That's the thing that's hard to describe until you see yours. You ask "what does my vault say about fear of failure?" and instead of searching through thousands of files, you get an answer in seconds with connections you wouldn't have found on your own. Notes cluster into communities and show you themes you hadn't named. Your own writing, organized in a way you never could have done by hand.

Whatever you've been writing about for years, the graph will show you what it's actually about.

The system starts patching itself

This is where it gets weird.

I built something called the Instinct Engine. After enough sessions, it starts detecting patterns: recurring friction, repeated mistakes, things you keep rediscovering instead of writing down. At the end of each session, it silently checks for signals. Did this task take too many steps? Did you correct the AI on something that might come up again? Did the session go down the wrong path before finding the right one?

If it finds something, it suggests turning it into a permanent rule or a concept note. The system patches itself based on how you use it.

In my first week, this produced 26 rules. Small ones: never count words in your head, use a calculator. Always check the clock before writing a timestamp. Re-read a file before editing if more than five minutes passed. None of them are revolutionary alone. But each one eliminates a class of error. After a week, dozens of failure modes just stopped happening. Your rules will be different from mine. They'll be yours.

85 voices that push back

The advisory panel started because I kept making decisions alone and regretting them. Now it's 85 named voices across 13 domains: founders, psychologists, economists, therapists, writers, strategists. Each one has a specific lens and a reason they're in the room for this particular question.

A hook detects decision language ("should I," "torn between," "weighing options") and pulls the relevant voices in automatically. You don't have to remember to ask for a second opinion. The system notices you're deciding and brings the room to you.

Last week the panel caught me about to repeat a pricing strategy I'd already tried and dropped. The marketplace founder voice flagged the pattern. The procurement perspective reframed the question. I changed course before making the same mistake a second time.

You configure your own panel. Your mentors, your thinkers, the voices you wish you could call before a big decision. They won't always agree with each other. That's the point.

One line installs everything

The first post said "set up yours in an afternoon." True, but it assumed you could troubleshoot plugins. For everyone who isn't technical, that was a wall.

Now it's easy: https://myceliumai.co/install

Claude interviews you and builds your system from your answers. I tested this on a friend who has never used a terminal. Working vault in under an hour.

What this actually is now

If you came here because you want a better note system, the second brain part still works. The vault is still Obsidian. The notes still connect. The memory still persists between sessions. Everything from the first post is still the foundation.

But it's not just that anymore. It remembers what you decided. Captures what you learn. Processes your meetings. Challenges what you assume. Tracks your patterns over time. And it gets smarter every session without you doing anything extra. The vault fills itself. The rules pile up. The graph gets denser. The panel pushes back harder because it has more to work with.

I wrote in my journal last week that it feels like building a machine, and the machine is starting to hum. That's the best way I can describe it. Not a tool I open. A system that's running whether I'm paying attention or not, and every time I come back, it's a little smarter than when I left.

That's not a second brain. That's infrastructure for thinking.

Two ways to start

Free. The repo is open source and the install is one command. A community is forming around it: people sharing vault setups, suggesting features, building on top of it.

https://myceliumai.co/install

Built for you. If you want me to build your vault, your workflows, your panels, optimized for how you actually work, I'm taking a limited number of clients at launch pricing. diazroa.com

I'm Adelaida, I write about consciousness and AI, and I'm still building this from Bogota.

I’ve been trying to organize my brain for over a decade.

I signed up for Roam Research the first week it launched. I’ve used Notion, Evernote, Apple Notes, Google Keep, scattered Google Docs across five accounts. I have notes from seven companies I’ve started. Journal entries going back to 2014. Book highlights. Meeting transcripts. Strategy docs. AI conversations. All of it scattered — some in Obsidian, some in Google Drive, some on my desktop in folders I named things like “IMPORTANT” and “Sort Later.”

I feel like I’ve lived five lives. And no system I tried could keep up.

I always knew the answer was to connect everything. I could see it in my head — all the threads, all the patterns. But I never had the time to actually build the system. There was always a pitch to prep, a client to call, a fire to put out.

Then I spent a week with Claude Code and did what I’d been wanting to do for years.

5,000 notes. 12 years of journals. Two books in progress. A startup to raise for. Thousands of contacts. All connected, all compressed, all navigable. And the AI doesn’t just search it — it knows me. It knows my startup, my emotional patterns, my team, my writing voice, what I decided last week and what’s still unresolved. The difference in output quality is unreal.

After I built mine, a friend sent me a post by Andrej Karpathy, the guy who built Tesla’s autopilot AI and co-founded OpenAI, describing almost the same approach. Raw data compiled into wikis, operated on by AI, all in Obsidian. It was exciting to see I was headed in the right direction.

Subscribe now

Here’s how I did it.

The problem: not just memory, but burn

Before this weekend sprint, I didn’t have a vault. I had notes in Roam Research. Notes in Google Docs. Notes in Apple Notes. PDFs and markdown files uploaded to Claude Projects. Strategy decks in Google Drive. Journal entries in one app, meeting notes in another, investor research in a third.

Every time I opened an AI tool, I started from zero. “I’m Adelaida, I run Onde, we’re raising $600k, my Co-Founder Sergio just joined, we have Accenture as a client...” Fifteen minutes of context-setting before any actual work. And even with Claude Projects where I’d uploaded files, I was burning through usage just getting it to find things.

The problem wasn’t just that AI had no memory. It was that every workaround I tried — uploading files, re-explaining context, copying and pasting between tools — was slower and more expensive than just doing the work myself.

What changed everything wasn’t giving AI more files. It was giving it a connected system. When your notes link to each other, when concepts reference people who reference meetings who reference decisions, the AI doesn’t have to read everything. It follows the links. It finds what it needs in seconds instead of scanning hundreds of files. The context is richer, the answers are more accurate, and you stop burning tokens on “let me search for that.”

That’s the difference between a folder of files and a knowledge graph. One is storage. The other is navigation.

The architecture: three layers of memory

The breakthrough wasn’t better organization. It was compression. I stopped thinking of my vault as one flat archive and started thinking of it as three layers:

Layer 1: Operating context (5 notes, read every session). These answer: who am I, what am I working on, what happened last, what’s unresolved. Start Here, Current Priorities, Open Loops, Last Session, and CLAUDE.md (the memory file). With these five files, Claude goes from blank slate to fully oriented in under 10 seconds.

Layer 2: Domain summaries (7 notes, read when relevant). Each major area of my life has one compressed summary: my startup, my writing, my home, my psychology notes, my school vision, my book notes, my business ideas. 200-400 words each. I also have 28 monthly journal summaries that compress 2,325 entries into one note per month. If I ask “what was I going through last October?”, the AI reads one note instead of twenty.

Layer 3: Deep archive (5,000 notes, read only when needed). Journals, AI chats, meeting transcripts, book notes, historical material. Stays large. Stays untouched. Only gets opened when a specific question requires it.

The optimization isn’t at the archive level. It’s at the summary level. Compress first, search second

Claude code prompt:

You can install the ai-brain-starter from https://myceliumai.co/install

The memory file: teaching AI who you are

The single highest-value thing I did was create a CLAUDE.md file at the root of my vault. This file loads automatically every session. It contains who I am, what I’m working on, my vault structure, key terms, people, rules for how the AI should behave, tool routing, and efficiency rules.

This alone changed everything. Claude stopped being a smart stranger and started being a colleague who was in the room yesterday.

The accountability rules

This might be my favorite part of the whole system.

By default, AI is a yes-machine. It agrees with you, validates you, avoids friction. That’s useful for autocomplete. It’s terrible for thinking.

So I added 15 rules to my memory file:

1. Correct me if I’m wrong.

2. Stop me if I’m gossiping.

3. Check me when I’m stubborn.

4. Tell me the truth even when it hurts.

5. Tell me when I’m self-sabotaging.

6. Call me out when I’m making excuses.

7. Remind me who I said I wanted to be.

8. Don’t let me settle just because it’s easier.

9. Check my ego every time.

10. Tell me when I’m overthinking everything.

11. Call me out if I’m playing the victim.

12. Don’t let me stay comfortable if it’s keeping me stuck.

13. Tell me when I’m the problem.

14. Call me out when I’m avoiding what I need to face.

15. Tell me when I’m out of alignment with my values.

The moment I added these, every conversation changed. Claude started pushing back on my self talk that was off. It told me when I was avoiding a hard conversation with a team member. It reminded me of a decision I’d made the week before that I was about to contradict.

AI is useless when it’s nice. It’s useful when it’s honest.

Delegation: AI managing other AIs

Here’s what most people miss: your AI assistant should delegate too.

I have paid accounts in Perplexity, Manus AI, Gemini Pro, Granola, n8n, PostHog, Linear, HubSpot, Framer, and more. Each is better than Claude at something specific. So I built a routing table into my memory file. (got most of these pro subscriptions through one subscription Lenny’s Newsletter)

When I ask Claude to research 30 angel investors, it tells me: “This is a Manus task.” Then it writes a structured task brief. I send the brief to Manus. Manus does the research autonomously. I paste the result back. Claude cleans it up and integrates it into my vault.

In one session, I had Manus running four tasks in parallel: CRM standardization across 243 contacts, monthly journal summaries for 2 years, domain summaries for 7 life areas, and a graph cleanup audit. Meanwhile, Claude was doing the work that needed judgment — fixing data quality issues, adding wikilinks, building dashboards.

Your main AI is the orchestrator. Your other tools are the workers. Route the right work to the right tool.

Automation: the vault fills itself

I set up a script that pulls meeting notes from Granola and auto-saves them to my vault as properly formatted Obsidian notes with frontmatter, attendee lists, and transcripts. It runs twice a day. I never manually import a meeting note again.

Anything that happens regularly should happen automatically. Meeting notes, journal prompts, decision logs. The vault should fill itself.

CRM that actually works

I had hundreds of contact files with inconsistent metadata. No way to query “show me all active investors” or “who needs follow-up.”

I ran a standardization pass: a script that added structured fields (relationship, status, priority, last interaction) to every contact. Then a correction pass to fix the mistakes. Now I have working Dataview dashboards for active contacts, investors needing follow-up, and monthly summaries with emotional tracking.

The vault went from a filing cabinet to a queryable database.

The graph: pruning matters more than planting

My vault had 1,846 unresolved links, 7,000+ orphan files, and 4,600+ dead ends. The graph was noisy.

The fix wasn’t adding more notes. It was fixing malformed links, adding aliases, creating the 10 genuinely missing concept notes, adding outbound links to dead ends, and deleting 1,400 junk files (including a node_modules folder that somehow ended up in the vault).

The graph is a map. A noisy map is worse than no map.

The daily journal

This one came from an insight I didn’t expect.

After Claude analyzed all 2,325 journal entries, it told me something I’d never noticed: when I have gaps between journaling, it usually means I’m having a good stretch. Life is flowing, I’m busy, things are working. But it also means I process the hard stuff in detail and barely log the good stuff. Twelve years of data on my worst days. Almost nothing on my best ones.

So we built a journal skill specifically to fix that. I type /journal, Claude asks me about my day, follows the thread when something interesting comes up, and pushes when I try to stay surface-level. If I say “good day,” it doesn’t let me off the hook: “What specifically made it good? I want to capture this one.” If I’m clearly avoiding something, it says so.

Then it saves the entry in my voice. First person, stream of consciousness, messy and real — the way I actually write in my journals, not the way AI writes.

That’s the pattern I keep seeing: Claude surfaces an insight from my data, and then we build a skill to act on it. The journaling gap led to the journal skill. The realization that I carry people too long led to accountability check-ins. The discovery that my gym streaks correlate with my most productive weeks led to tracking it in every entry. The system doesn’t just store information — it generates insights that change behavior, and then those changes get built into the system.

Each journal entry also gets reviewed by an advisory panel — 60+ thinkers I selected across wealth, psychology, relationships, health, creativity, and strategy. Names like Naval Ravikant, Brene Brown, Esther Perel, Peter Attia, Rick Rubin, Marcus Aurelius. Based on what came up in the conversation, Claude picks the 3-5 most relevant voices and gives me their perspective in 1-2 sentences each. It’s not a gimmick. It actually feels like getting pushback and ideas from people who think very differently than I do. One night the panel caught me romanticizing a pattern I’d repeated three times before. Another night it connected a money anxiety to a relationship dynamic I hadn’t seen.

I’ve been journaling for 12 years. I’ve never been as consistent as I am now. Because it doesn’t feel like homework. It feels like a conversation with a friend who knows me and a panel of mentors who challenge me.

You don’t need 12 years of journals to start. You need one entry. The system compounds from there.

Subscribe now

The surprise: insights I never asked for

Once the vault was connected, I asked Claude a question I hadn’t planned: “Now that you know every area of my life, what questions have I not asked that I should be asking?”

What came back blew my mind.

Claude had mapped my entire emotional life by floor — using the High-Rise framework from my book — cross-referenced with whether I slept well, whether I worked out, who I was with, and what I was working on. It told me which people consistently drain me and which ones correlate with my highest floors. It showed me that this year I’ve written more than twice as much from the higher floors than in any previous year. It identified my happiest days and what they had in common. It flagged patterns in my relationships that I’d been living inside for years without seeing.

It told me my sister is the most stable high-floor presence in 12 years of journals. I sent her that. She deserved to know.

These aren’t things I asked for. They’re things the system surfaced because it could see across 5,000 notes at once. No human could read 2,325 journal entries and cross-reference them with my CRM, my calendar, and my sleep patterns. Claude can. And when every note links to every other note, the insights aren’t just summaries. They’re connections I couldn’t make myself.

The team vault

I also built a separate vault for my startup team — synced through Google Drive so everyone accesses the same files. It has its own CLAUDE.md, its own dashboards, its own CRM entries. When anyone on the team opens Claude in that vault, they get the same context: who our clients are, where the raise stands, what happened in the last meeting, what the priorities are.

But I still work from my personal vault. Because it knows how my brain works — not just the company, but my writing, my journals, my patterns, everything. It can make connections across all of it, so even when I'm thinking through a business problem, the extra context makes the thinking better in ways I didn't expect. The team vault syncs Onde-related files from mine automatically. I get the full picture. They get what they need.

How I actually use it

I almost never open the Obsidian folders directly. I don’t click through sidebar trees looking for files. I do almost everything through Claude Code — the desktop app. I just talk to it.

“What did we decide about pricing last week?” Claude finds the decision log entry.
“Draft an email to the Accenture contact.” Claude pulls context from the CRM and the last meeting note.
“What’s my investor pipeline looking like?” Claude queries the Dataview dashboard.

The vault is the brain. Claude is the interface. I rarely touch the files myself — I just talk, and the system responds with full context.

What changed

Before: 15 minutes of context-setting every session. Strategy docs buried in Google Drive. 5,000 files at one flat depth. One tool for everything. Meetings, decisions, and insights lost in conversation.

After: AI orients itself in 10 seconds. Every domain has a summary. Manus handles research, Claude handles judgment, scripts handle bulk. Meetings auto-import, decisions auto-log. The vault fills itself and gets smarter over time.

And the most important change: I went from organizing as procrastination to organizing as infrastructure. The vault isn’t a project I maintain — it’s the foundation everything else runs on.

Set up yours in an afternoon (free)

I open-sourced the whole system.

You can install the ai-brain-starter skill from https://myceliumai.co/install

Claude will interview you about your life and build everything from your answers. The memory file, the folder structure, the context layer, the CRM, the journal routine, the accountability rules, the power tools. All of it.

What took me a week of figuring things out, making every mistake, and optimizing until it ran smoothly — you can have running in one conversation. The CRM standardization that took me a full day? Claude does it in 30 minutes now. The monthly journal summaries from 2,325 entries? Batched and automated. The graph cleanup, the domain summaries, the dashboards — all built into /optimize-brain so you skip the trial and error.

You don’t need to be technical. You don’t need to read a guide. You paste one line, have a conversation, and by the end Claude knows who you are.

What the repo doesn’t do

The skill builds the architecture. It can’t do the interior design. That part is specific to your life, your patterns, your mess.

I ran Claude on 12 years of my journals and it pulled out 10 questions my therapist, business coach, and best friend would all ask. That took thousands of entries processed with judgment. The skill can’t replicate that.

I built a panel of 60+ thinkers (Naval, Brene Brown, Esther Perel, Peter Attia, Rick Rubin, more) that Claude consults daily based on what comes up in my journal. That needs to be calibrated to your world, not mine.

I trained Claude to write in my voice by feeding it my actual writing. I built a shared vault for my startup team with 128 files and live dashboards. I connected my vault to HubSpot, email, calendar, meeting notes auto-import, investor pipeline tracking. That’s where a personal vault becomes a business operating system.

The repo gives you the architecture. The call gives you the interior design.

The meta-lesson

The tools don’t matter as much as the architecture. Obsidian is just markdown files. Claude is just a language model. Manus is just a research agent.

What matters is how they talk to each other. The memory file is the bridge. The summaries are the compression layer. The routing table is the traffic system. Build the architecture once. Then everything you do compounds.

No one can sell you a second brain. They can only give you the tools to build one.

Install it. Use it. Make it yours. If you build something cool with it, tell me.

You can install the ai-brain-starter skill from https://myceliumai.co/install

I write about consciousness, AI, and building a life that makes sense at Regard on Substack.

Founding members ($240/yr) get a 1-hour call. We build your vault live, run the deep optimization together, set up your team vault, or talk about whatever matters most. The repo gives you the architecture. The call gives you the interior design.

→ Become a founding member

I’m Adelaida. I run Onde, I write about consciousness and AI, and I built all of this with Claude Code and too much coffee in Bogota.

• Generate images that actually look like you

• Use them for LinkedIn, press, founder profiles, and editorial features

• Refresh your visuals quickly whenever you need them

• Avoid AI distortion, beautification, or “modelization”

This is not about fantasy images.
It is about identity-accurate representation. Can you tell which of these (if any) are AI or real?

What This Is and What It Is Not

This works because you treat AI like a photographer, not like an artist.
You are not asking it to invent you.
You are giving it enough reality that it has no room to guess.

The goal is simple:
Preserve identity first. Style comes second.

Step 1: Create Strong Identity References

Everything depends on this step.
If you skip or rush it, the results will drift.

Your reference photos are not content.
They are technical inputs.
They should be boring, clear, and honest.

Use the Identity Reference Photo Guide at the end of this article and take the photos exactly as instructed.

Do not style.
Do not pose.
Do not filter.
Do not try to look good.

These photos exist so the system does not change you.

Step 2: Add Body Reality and Measurements

AI often changes bodies when it does not have constraints.

If you have body measurements from tailored clothing, include them as an image or text reference. If you don’t, it’s easy to do at home with a tape measurer.

This locks scale and proportions and prevents:

• Height changes

• Torso slimming

• Hip or shoulder widening

• Fashion model distortion

Example of what to include:

• Height

• Bust, waist, hips

• Shoulder width if you have it

• Any tailoring notes you already use

This is what I gave it:

You are not asking the system to improve your body.
You are telling it what already exists.

Step 3: Define Your Color Profile

Color is one of the biggest sources of drift.
If you already know your color season, include it.
If not, you can feed ChatGPT multiple photos and ask it to analyze your undertone and best palette.

You can also use an existing color analysis you trust.
In my case, I used a Deep Autumn palette and explicitly restricted wardrobe colors.
This matters because:

• It keeps skin tone accurate

• It prevents washed-out or overly cool results

• It keeps wardrobe realistic and repeatable

You can also reference an existing color profile as an example and say you want something similar.

Here’s the one I gave it for me:

Step 4: Use a High-Constraint Prompt

Below is the core prompt I used.
This is not poetic on purpose.
The structure matters more than the words.

You are establishing hierarchy:
Identity first.
Body second.
Style last.

You can adapt it, but do not loosen it too much.

Subscribe now

Sample Prompt

Create a high-end editorial photograph of the SAME WOMAN shown in the reference images.

This is a re-photograph of the same real person. Not a casting. Not an interpretation.

IDENTITY LOCK

Preserve her exact facial structure, proportions, and scale.
No beautification. No reshaping. No facial feature changes.

FACIAL AND PHYSICAL REALITY

Preserve these traits exactly as shown:
Dark brown straight hair
Hazel-green eyes with olive and gold flecks
Neutral-warm skin with peach olive undertone
Natural smile and real teeth shape
Petite height 5’1” with real-world proportions

BODY ACCURACY

Respect real measurements exactly.
No elongation. No resizing. No fashion model proportions.

STYLE AND PRESENCE

Editorial, modern, grounded.
Founder energy. Calm authority. Intelligence and warmth.
Real photograph, not styled concept.

WARDROBE

Timeless tailored pieces only.
Deep Autumn palette only. Natural fabrics. Clean lines.

HAIR AND MAKEUP

Natural, polished, understated. Skin texture visible. No smoothing or contouring.

LIGHTING

Soft editorial daylight or diffused indoor light. Warm and true skin tone.

COMPOSITION

Medium or close portraits. Neutral architecture or lived-in interiors. No staged luxury.

CAMERA

Natural full-frame look. 50 to 85mm perspective. Shallow depth of field. No wide-angle distortion.

EMOTION

Calm confidence. Approachable. Self-possessed.

DO NOT

Change facial features. Change body size. Over-smooth skin. Replace her with a similar woman.

RESULT

A refined editorial photograph suitable for founder profiles and press.

Step 5: Upload References Correctly

Upload:

• 5 to 10 identity reference photos

• Measurements image or text

• Color profile

• Nothing else

Do not mix style references at first.
Do not add inspiration images yet.

Get one result that makes you say: that is me.

When that happens:
Lock the seed.

Then iterate.

Step 6: Iterate With Discipline

Only change one variable at a time:

• Location

• Outfit

• Expression

• Environment

Do not rewrite the entire prompt every time.
That introduces drift.
If an image does not immediately look like you to someone who knows you, discard it.
Do not fix with more words.

Why This Works

This works because:

• Identity is locked before style

• Body reality is respected

• Color is constrained

• Camera and lighting are specified

• The system is not asked to invent

You are not generating a version of yourself.
You are re-photographing yourself.

Identity Reference Photo Guide

These photos are not for Instagram.
They are identity anchors.
Follow these rules exactly.

1. Front Neutral Expression

Camera at eye level.
Even daylight.
Hair back.
No smile.
Minimal or no makeup.
This is the most important photo.

Subscribe now

2. Front Natural Smile

Same setup.
Teeth visible.
Relaxed and real.

3. 3/4 Angle Left

45 degrees.
Neutral expression.
Hair tucked on camera side.

4. 3/4 Angle Right

Same as above, mirrored.

5. Full Side Profile

True 90 degree profile.
Hair fully off face.

6. Full Body Front

Camera at mid-torso height.
Use 2x zoom.
Simple fitted outfit.
Arms relaxed.

7. Full Body Side

Same setup, side view.

8. Walking Candid

Mid-step.
Natural movement.

9. Seated 3/4 Body

No objects.
Hands visible.
Neutral expression.

10. Close Face Daylight

Minimal makeup.
Visible skin texture.
Optional but helpful:

• Warm indoor light face crop

• Best color outfit

• Clean professional look

General rules:

No filters.
No beauty mode.
No dramatic lighting.
Consistency matters more than quantity.

Minimum set if short on time:

Front neutral.
Front smile.
3/4 angle.
Side profile.
Full body front.
Seated 3/4.