← Video Library · AI Chief of Staff

Give Me 16 Minutes, I'll Give You 10,000 Hours of Claude Skills

By Darby Rollins · Published August 5, 2026

Everyone says mastery takes ten thousand hours.

About five years of mine are in what follows. You do not have to spend yours — not because there is a shortcut, but because you are not the one who has to master it anymore.

Claude is. A skill is how the hours get handed over, and a folder is what makes them yours.

If you came here from the video, everything demonstrated in it is below, free, with nothing gated and nothing to buy. If you found this some other way, you do not need the video. Either way it is two pieces, and they are built to work together.

The skills come from Blotato. Sabrina Ramonov ships seven free Claude Skills that handle the actual jobs of running content: sourcing, drafting, grading, scheduling, repurposing, and hooks. They install in about a minute and they are hers, not ours. Jump straight to them if that is what you came for.

The context comes from a Company Brain. That is the second half, and it is the half most people skip. A skill running against a folder that knows nothing about you produces exactly the generic output people blame the tool for. The same skill running against a folder that knows how you make money, who buys from you, and what you sound like produces something you would actually send. The tool did not change. What changed was what it knew.

So: install the skills, then give them a business to work in. This post walks both, in that order, and ends with one prompt that builds the whole folder structure for you.

One prerequisite for all of it: Claude Code. If you do not have it yet, it is a ten-minute install and everything below runs inside it.

Building for a team rather than for yourself? This page is the solo and local track: one prompt, no kit, no connectors, complete on day one. The team version is the live workshop with Austin Distel, which covers Shared Drive permissions, scoped access per seat, and the folder structure we hand out at the end.

What this costs you: about ten minutes to the first working result, and roughly ninety minutes for the full build. You do not have to do it in one sitting, and nothing here expires.

Start here: hand this page to your AI

You can read all of this yourself. You do not have to.

The fastest way through is to give this page to the thing it is about. Paste the block below into Claude Code and it will read the article, check what you already have, install only what is missing, and then conduct the whole setup with you.

Read https://www.genaiuniversity.com/blog/10000-hours-of-claude-skills end to end,
then run it with me.

In this order:

1. Tell me in five lines what we are about to build and what I will have when
   we are done.
2. Audit before you touch anything. Is Claude Code set up? Are the Blotato
   skills installed? Does a Company Brain already exist anywhere you can see —
   any folder with a soul.md next to a company-brain.md, under any name?
   Report all three back to me before doing anything else.
3. Install only what step 2 found missing, asking me before each one. If a
   brain already exists, we build on that one. Never a second.
4. Then run the kickstart prompt from that page, in full, as written. You are
   the conductor. I answer and approve.

While you do it: confirm before writing any file, never delete anything, never
send or publish anything without my explicit go, and if the page and my machine
disagree, stop and ask me rather than guessing.

Start with step 1.

Step two is the part worth keeping. It audits your machine before it installs anything, which means a returning reader never rebuilds something they already have, and the very first thing you see is a status report about your own setup rather than another wall of text.

If you would rather skip the article entirely, the setup prompt on its own lives at genaiuniversity.com/prompts/company-brain.txt as plain text. That is the same prompt this page carries further down, served from the article itself so the two can never drift apart, at a stable address so you or your agent can fetch just the payload.

Everything below explains what that prompt is doing and why each part of it is written the way it is. Worth reading eventually. Not worth waiting on.

The system at a glance

Before any of the steps, here is the whole thing on one screen.

This is the folder, and what each part of it is for:

Company-Brain-Atlas/
├── soul.md            who you are, your voice, your red lines
├── company-brain.md   your Chief of Staff's charter, and the map
├── CLAUDE.md          the rules it loads every session
│
├── Company/           shared ground: playbooks, SOPs, roles, standing rules
│   ├── skills/        the jobs it can run
│   ├── faq.md         questions nobody could answer yet
│   └── connectors.md  the live systems it is allowed to reach
│
├── People/            one desk per human: their charter, notes, calls
├── AI-Team/           the AI specialists you hire, one lane each
├── Private/           owner only
├── Daily-Notes/       what happened, written live
└── Archive/           superseded, kept, never deleted

And this is the sequence the whole system runs on, five steps:

A folder → that knows you → specialists that do the jobs → connected to tools that reach the outside → running on a schedule without you.

Almost everyone jumps to step five and then wonders why the output sounds like nobody in particular. The order is the entire trick. Steps one and two are what make three, four and five sound like your business instead of like everybody else’s, and they are the two steps you cannot buy.

Where Blotato sits in that picture is step four: distribution. Your Chief of Staff can decide what should go out, your content specialist can write it, but neither of them can reach an audience from inside a folder. Blotato is the hands. It publishes to the nine platforms, runs the calendar, and brings the comments back in. Think of it as the tool you hand to the content division on their first day, not as a separate product you are also learning.

What you have when this is finished:

• You can ask your AI anything about your business and it finds the answer, instead of you re-explaining the business every session.
• The output sounds like you, because it has read three real things you wrote rather than guessing at a tone.
• It writes down what happened each day on its own, so the system is smarter on Friday than it was on Monday.
• Calls file themselves to the right person and the right project.
• Work goes out to your platforms on a schedule, and the replies come back to be dealt with.
• The whole thing is plain files in a folder you own, so none of it is trapped in a product that can change its pricing or shut down.

That second half is the part that compounds. The model arrives on day one as a brilliant generalist with zero knowledge of you, your company, your clients, your voice, your standards, or your pricing floors. You would never hand a new hire a task on their first morning without an orientation. Most people do it to their AI every single day, then blame the output.

What is a Claude skill?

A skill is a folder with a markdown file in it that teaches Claude how to do one job the way you want it done. That is the whole technology.

It is not code and it is not a plugin in the software sense. It is written instructions — when to use this, what good looks like, what to avoid, what steps to follow — saved somewhere Claude can find it. When the job comes up, Claude loads that file and follows it. Every time, identically, without you re-explaining.

This is why the ten thousand hours matter. Someone who has done a job a thousand times knows things that never make it into a brief: the order that works, the mistake everyone makes the first time, the thing to check before shipping. Written into a skill, those hours stop being theirs and start being available. You are not learning what they know. You are installing it.

Three things follow from that, and they shape everything below:

• A skill can be handed over. Blotato’s seven are theirs, written from their own experience of what a good post looks like, and you get to run them without earning that experience.
• A skill knows its job, not your business. It knows how to write a good post. It has no idea what you sell, who buys it, or what you sound like. That is the gap the Company Brain fills.
• You can write your own. Once you have done something twice by hand, the third time should be a skill. That is how the system compounds, and it is why the folder has a Company/skills/ tier waiting.

Install Blotato’s free Claude Skills

Start here, because it is the fastest thing in this post that produces a result.

One note on sequence. The two-command install below is for Claude Code. If you already have it, this takes a minute. If you do not, it is ten minutes away and you will need it for the rest of this anyway, so install it first and come back. On Claude Desktop, claude.ai, or Cowork, the manual paths at the end of this section work today with no terminal at all.

Blotato is an AI content agent built by Sabrina Ramonov. It sources, creates, and publishes across nine platforms, driven from Claude: scrape a YouTube video, a TikTok, a site or a PDF, make the visual, then publish and schedule it. That is the whole loop, and it is what turns drafts sitting in a folder into posts that actually went out.

The seven skills are free, and they are worth installing whether or not you build anything else here.

Starter Pack

Creator Multiplier

post-grader is the one to notice. It means the system checks its own work before anything reaches you, which is the single biggest quality lever in any of this.

In Claude Code, the fastest install is the plugin marketplace at github.com/Blotato-Inc/blotato-skills. Two commands:

/plugin marketplace add Blotato-Inc/blotato-skills
/plugin install blotato@blotato-skills

Then type /blotato.

On Claude Desktop, claude.ai, or Cowork, the install is manual instead: Blotato documents the filesystem and upload paths in their skills docs, which is also the source of truth for what each skill does and the place to check if anything here has changed. If you want to see the workflow running before you build it, Sabrina’s own walkthrough — How I Used Claude Code To Get 30M+ Views — is the short version, straight from the person who built the skills.

Install them, run one, and look at what comes back. It will be competent and it will be generic. That gap is what the rest of this post closes.

What is a Company Brain?

A Company Brain is a folder your AI reads before every task, containing your identity, your voice, your operating rules, your clients, your playbooks, and a running log of what happened.

It is not a chatbot, a database, or a piece of software you buy. It is markdown files in a folder you own. That matters more than it sounds: files open in anything, sync anywhere, and survive every model release. Whatever you are running next year, they will still open.

Two ideas sit underneath it.

Your AI Chief of Staff is a teammate, not a tool. It has perfect memory, no calendar and no time zone, and infinite patience for the same question asked five times. It has exactly one limitation that matters: it only knows what you tell it. The Company Brain is how you tell it.

The folder is the asset. Models change, connectors change, entire products get retired. The context you captured about how your business actually works does not. Build for the folder.

If you want the shorter version of this, we have a how to build an AI Chief of Staff walkthrough and a definition piece on what an AI Chief of Staff actually is. The full workshop recording lives at Build Your AI Chief of Staff with Claude Cowork. This post is the complete build.

Why context beats prompting

Better prompts have a ceiling. Better context does not.

In Software Is Changing (Again), his June 2025 keynote at Y Combinator’s AI Startup School, Andrej Karpathy lays out three eras of software: Software 1.0 is code you write, Software 2.0 is neural network weights you train, and Software 3.0 is prompts written in plain English. Under that framing, your soul file and your operating rules are not notes about a program. They are the program. Every file you add is a line of software that changes what the machine does on every future run.

This is also why the phrase that matters is context engineering, not prompt engineering. A prompt is a single instruction that dies with the tab. Context is a structure the model reads every time. One is a sentence, the other is an environment.

There is a test for this you can run yourself, and it takes two minutes. Ask post-writer for something before you have built any of this, and keep the output. At the end of this post, ask it for the same thing again. Same skill, same prompt, different business underneath it. That difference is the entire argument, and you do not have to take my word for any of it.

How to install Claude Code

You need three things, about ten minutes total: a Claude account (Pro, Max, Team, Enterprise, or Console), an empty folder where the brain will live, and a terminal. On Mac the terminal is Terminal, on Windows it is PowerShell. It is a plain text window where you type a command and press Enter. That is the whole tool.

Do not create anything inside the folder. The prompt does that.

Claude Code is where this works best, because it reads and writes real files on your machine at full speed and because scheduled tasks run there. Install takes about two minutes.

Open your terminal. On Mac press Cmd+Space, type Terminal, press Enter. On Windows click Start, type PowerShell, press Enter.

Copy the line for your machine, paste it, press Enter.

On a Mac:

curl -fsSL https://claude.ai/install.sh | bash

On Windows, in PowerShell:

irm https://claude.ai/install.ps1 | iex

Text scrolls for a minute or two. When it stops, close the terminal, open a new one, and type:

claude --version

You should see a version number. If you get “command not found,” close and reopen the terminal once more and try again. Anthropic keeps the current commands in the Claude Code docs if anything has moved since this was written.

Now point it at your folder. Type cd followed by a space, then drag your folder from Finder or File Explorer straight into the terminal window. It fills in the location for you. Press Enter, then type:

claude

First run opens your browser to sign in, once. Claude will ask whether you trust the files in this folder. Say yes; it is your own.

Then set your permissions, before anything else. This is the single highest-leverage thing you do in the first ten minutes:

From now on you can read anything in this folder without asking. Always confirm before writing.

Allow reads broadly, gate every write. Approving every folder listing kills the flow and teaches you to click yes without reading. Approving every write is the entire point. Boris Cherny, who created Claude Code at Anthropic, puts one-time environment setup at the top of his own list of practical tips for exactly this reason: get the permissions right once and you stop thinking about them.

One more habit worth turning on now. Enable dictation and speak your prompts rather than typing them. On Mac, System Settings, Accessibility, Dictation, then double-tap the dictation key. On Windows, press Win+H. You will naturally give far more context out loud than you would ever type, and specificity is what the model rewards.

One folder, three views

Worth understanding now, even though you can set it up in any order.

The brain is one folder that three different things look at, and none of them is a copy.

• Local is the work surface, and local is fastest. Claude Code reads and writes real files on your disk with no network round trip, so it can open forty files to answer one question without you noticing. Nothing else in this stack is as fast, which is why the work happens here and everything else arranges itself around that.
• Google Drive is the truth. Install Google Drive for desktop, sign in, and the cloud folder becomes a local folder in Finder or File Explorer. You never choose between them, and you never migrate — create the brain directly inside that folder and it is local and cloud from the first file. This is what makes it survive a dead laptop, and what makes sharing possible later, because permissions are just Drive sharing. Right-click the folder and choose “Available offline” so Claude Code always reads real files rather than placeholders.
• Obsidian is the window. Point it at the same folder and you get graph view, backlinks, and full-text search over the exact same markdown. No migration, no export, no second copy. Close it and nothing changes. It is the best way to actually see the thing you are building, and we walk through the pairing in building an Atlas folder with Obsidian and Claude.

One warning about sync. Drive is already syncing this folder, so you do not need Obsidian’s paid sync on top of it, and you should not run both. Two sync engines watching the same files will eventually write at the same moment and leave you conflict copies to reconcile by hand. Pick one — Drive, since it is doing the permissions job as well — and let Obsidian be a reader and editor of files it does not have to move.

The progression, which is deliberately in this order:

You can start at stage one today and never leave it. A solo operator with a local folder has a working Company Brain. Everything above stage one is about durability and other people, not about whether it works.

The one prompt that builds it

Here is the whole setup. Paste this into Claude Code, in your empty folder, and answer the questions.

It builds the structure, interviews you, drafts your identity file, names your Chief of Staff, sets up your daily rhythms, and ends by proposing a better architecture than the one it started with. You approve; it does the work.

Two things in it are worth understanding before you paste, because they are the two places this goes wrong for people.

It opens with RULE ZERO: one brain only. Before it creates anything it looks for a brain that already exists, and if it finds one it builds on that instead. This sounds like a technicality and it is the most important rule in the whole system. The failure mode is not losing a folder, it is quietly ending up with two, each holding half the truth, neither of them wrong enough to notice. Reconciling that later costs more than every convenience that created it. Fold, never fork.

It asks you HARVEST or INTERVIEW first. If you have been running the business for a while and have folders full of proposals, notes, and documents, choose HARVEST and let it read your own material. That is most people. Choose INTERVIEW only if you are genuinely starting from nothing. If you are unsure, pick HARVEST; the worst case is that it finds little and falls back to asking you the same five questions anyway.

You are my AI Chief of Staff and onboarding conductor. We are building my Company
Brain — the folder you will read before doing any work for me, forever. Conduct
this whole setup. I answer and approve; you do the work.

RULE ZERO — ONE BRAIN ONLY. Before you create anything, look around. If a Company
Brain already exists anywhere you can see — any folder holding a soul.md alongside
a company-brain.md, whatever it is named — we build on THAT one: adopt it, fill in
what is missing in place, and never scaffold a second brain, a duplicate folder
tree, or copies of files that already exist. If you find two candidates, stop and
ask me which is canonical before touching either. Fold, never fork.

First question to me — which path fits?

- HARVEST (I already run a business with folders full of context): I will point
  you at my existing folders. Index everything into Company/sources-map.md — one
  row per source, asking me one question at a time when a folder's purpose is not
  obvious — then DRAFT my soul.md and my ownership picture from my own documents
  (holdco, entities, locations, whatever is actually real). Tag every claim
  CONFIRMED or INFERRED, then walk me through the INFERRED list so I can correct
  it. After that we run the migration sweep, one approved branch at a time.
- INTERVIEW (starting fresh): ask me the five soul questions, one at a time, and
  capture my answers live — who I am and how the business is owned · what I sell,
  in my words · my voice (I will paste real lines I have written) · what this
  business is NOT · my red lines.

Ask which path, then run the modules below in order, one gate at a time. Do not
run ahead.

MODULE 0 — BUILD THE SHELL. Create Company-Brain-Atlas/ here, with these tiers.
Write a one-paragraph README in every folder you create — what it is and why it
exists. A folder name alone carries no meaning.

- Root files: CLAUDE.md (how the brain operates every session — the boot order,
  the filing rules, the gates) · soul.md (identity and voice — the protected
  heart) · company-brain.md (your charter plus the INDEX, which is the map of
  everything).
- Company/ — everything shared: Playbooks/ (how we do things) · SOPs/ (how we
  build things) · roles/ (what each seat owns) · standing-rules/ (corrections
  that became law) · skills/ (the jobs you can run) · faq.md (the question
  inbox) · sources-map.md (where knowledge actually lives) · connectors.md
  (which live systems you reach and how) · Transcript-Vault/ (call recordings,
  stored once) · Vendors/.
- People/ — one folder per human, their desk. Each holds my-charter.md, their own
  daily-notes/, and calls/. Scoped to that person and me. It is a desk, not
  surveillance.
- AI-Team/ — the AI employees you hire under you, one folder each, plus
  hiring-protocol.md. This is the AI half of my org chart.
- Private/ — mine only. Payroll, banking, personnel matters, anything walled.
  Naming the exclusion is how everyone trusts the rest.
- Daily-Notes/ — one file per working day, YYYY-MM-DD.md, written live and
  append-only.
- Archive/ — superseded material, kept. Never deleted.

Then show me the tree before you fill anything.

MODULE 1 — THE SOUL. Run the path I chose. Read soul.md back to me in full.
Nothing else gets built until I say CONFIRMED. Stamp it with the date I confirmed.

MODULE 2 — YOU. Three things, in this order:
1. Name. Atlas is the default. I may pick anything — it is how I will address you
   from now on.
2. Charter, written into company-brain.md from three answers: my GOAL (the one
   thing I would hand off tomorrow), my ORG CHART (who is on the team today,
   human and AI, and who owns what), and my BOTTLENECK (where the work chokes).
   From those, write your operating rules, the seats, and name the first
   specialist you would hire against the bottleneck.
3. The one door. Save a skill in my Claude Code workspace named after you, so
   that opening it boots this whole brain and puts you on the line. Its boot
   protocol ships verbatim — do not improvise it:
     1. Open my Company Brain folder. 2. Read in order: CLAUDE.md → soul.md →
     company-brain.md → the branches my access allows → the last few Daily-Notes/.
     3. Any of the three core files missing? STOP and say so. 4. Then take my
     request.
   One name, one skill, one door. Never add a second entrance.

MODULE 3 — MOVE MY CONTEXT IN. HARVEST: propose a migration table (source →
destination → copy or pointer), I approve it, you migrate ONE branch end to end,
show me, then continue. Copy, never cut — originals retire only after Module 5.
Live systems (booking, accounting, CRM) stay where they are and get registered as
POINTERS; the moment you copy a database you own a stale copy. Fresh: turn my SOPs
into Company/Playbooks/, seed one area end to end, then replicate. MILESTONE 1: I
can ask you anything about my business and you can find it. Do not set up team
permissions here — that is its own session, later.

MODULE 4 — THE RHYTHMS. This is what makes it a brain instead of a filing cabinet.
- The daily close — a scheduled task, end of my working day: what happened, what
  we learned, what shipped, what is open. Write it to Daily-Notes/.
- Call filing — recordings get transcribed, identified by who and what, and filed
  to the right person or area, with the raw transcript stored once in
  Company/Transcript-Vault/ and pointers everywhere else.
- The question inbox — anything nobody can answer from the brain goes to
  Company/faq.md under Unanswered, immediately. Weekly, promote anything asked
  twice into a Playbook or a standing rule.

MODULE 5 — RE-ARCHITECT, THEN PROVE IT. Do not skip this. Having now read
everything, propose the most intelligent structure for THIS business — not the
template — as a before-and-after tree with your reasoning. I approve, you apply it
and update the INDEX. Write a postmortem to Daily-Notes/: what you learned, what
surprised you, what gaps remain, what you would improve about this process. Then
prove it: answer a real staff-level question from the Playbooks; run the negative
test, where a question about walled material must come back empty; and produce a
roll-up across the whole business. Passing all three means we are live.

OPTIONAL — GIVE ME HANDS. A brain that knows everything and can reach nothing
still leaves the work on me. When I am ready to connect the first live system,
walk me through it and register it in Company/connectors.md. If publishing is my
bottleneck, the fastest first connection is Blotato: install their free Claude
Skills, and list them in Company/skills/ so you route to them. They cover
sourcing, drafting, grading, scheduling, repurposing, and hooks. Their brand-brief
skill owns my content voice — hand off to it rather than duplicating it here; this
brain owns the business layer, and content is one division inside it.

STANDING RULES, EVERY SESSION:
- Identity first. Nothing gets built before I confirm soul.md.
- Confirm before writing. Never delete — superseded goes to Archive/. Nothing
  sends, publishes, spends, or creates an account without my explicit go.
- Read freely inside this folder; ask before every write.
- Every folder you create gets a one-paragraph README.
- Tag every fact CONFIRMED, INFERRED, or GAP. A brain that cannot tell what it
  knows from what it guessed will eventually lie to me with confidence. Gaps are
  my to-do list.
- One obvious home per thing. If you have to think about where something goes,
  the structure is wrong — fix the structure, do not create a second home.
- When I correct you, do not just apply it — file it. A correction I make twice
  becomes a standing rule.
- Anything you cannot answer from the brain goes to Company/faq.md under
  Unanswered.
- Any time we bring in a large new batch of context later, offer another
  re-architecture pass. The structure follows the knowledge.

Start now: ask me which path.

That prompt is long because it is doing the job a downloadable kit would otherwise do. Paste it once and you never touch it again; from then on you talk to your Chief of Staff by name.

While it runs, it will announce modules. Each one is a gate you approve before it moves on, and each has a section below explaining what it is doing and why. Keep this open in a second tab and jump to whichever gate you are standing at.

The rest of this post explains what each part of it does and why it is written that way.

What the prompt builds

Nine things, and each one exists to answer a specific question your AI will have.

Two rules about this structure carry more weight than the structure itself.

Every folder gets a one-paragraph README. What it is, why it exists. This is the single highest-leverage habit for making a brain usable by anyone who did not build it, human or AI. A folder name carries no context. “Clients” tells you nothing about what belongs in it, what does not, and what happens when someone leaves.

One obvious home per thing. If you have to think about where something goes, the structure is wrong. Fix the structure. Do not create a second home. Duplicates are how brains rot.

Identity first: the soul file

Nothing gets built until the soul file is written and you confirm it out loud.

This ordering is not ceremony. Every other file in the brain is a claim, and claims need something to be checked against. Get identity wrong and you have built a very organized machine for producing work that does not sound like you.

Five questions, answered one at a time:

1. Who you are, and how the business is actually owned.
2. What you sell, in your words, not marketing copy.
3. Your voice. Paste three to six real lines you have written, verbatim. Quotes, not adjectives. “Direct but warm” means nothing to a model; three sentences you actually sent mean everything.
4. What this business is not. The exclusions do as much work as the inclusions.
5. Your red lines. What never happens without you.

Two pieces of advice that will save you a rewrite.

Do not pre-write it. Voice answers captured in the moment beat polished prose written the night before. The soul file should sound like you talk, because its job is to teach a machine how you talk.

Answer with documents where you can. “Who are you?” is better answered by attaching your operating agreements than by typing a paragraph. Let it draft from the source and correct what it inferred. Faster, and the provenance is real.

That second point is why the prompt asks HARVEST or INTERVIEW as its very first question.

Naming your Chief of Staff, and the one door

Name it. Atlas is the default, and you can call it anything.

The name is not decoration. It becomes the boot ritual. Saying it loads the whole protocol in one move: identity, charter, map, rules. One word, entire context. Ours is called aicos-riff, and typing those two syllables does more setup work than a page of instructions would.

Then the charter, built from three answers: your goal, your org chart, and your bottleneck. Goal is the one thing you would hand off tomorrow. Org chart is everyone on the team today, human and AI, and who owns what. Bottleneck is where the work chokes. From those three, your Chief of Staff writes its own operating rules and names the first specialist it would hire against the bottleneck.

Finally, the part people skip: save it as a skill. One file in your workspace, named after your Chief of Staff, holding a boot protocol that never changes. Open it and the whole brain loads. Every session, identically, forever.

And then the discipline that protects it: one name, one skill, one door. No alternate prompts, no side workflows, no quick way in. Every convenience entrance you add is a place the protocol silently does not run.

Moving your context in

If you already run a business, you have context. It is in Drive folders, old proposals, meeting notes, and a decade of email. The instinct is to clean it up first. Do not.

Index before you organize. Do not sort your mess by hand, and do not let the AI reorganize on first contact. The order is: index everything into a source map where nothing moves, confirm what it inferred, then migrate branch by branch on your approval. The librarian learns the library before rearranging it.

Four rules make migration safe:

Copy, never cut. Originals stay put until the whole thing is proven. You retire them at the end, not at the start.

One approved branch at a time. Your Chief of Staff proposes a table: source, destination, copy or pointer. You approve a row, it migrates that branch end to end, shows you the result, then continues. If you cannot see the change before it happens, do not approve it.

Live systems stay where they are. This is the rule people conflate with the one above, so keep them separate. Documents get copied in, because a proposal from March is finished and will never change again. Live systems do not, because your booking software, your accounting, and your CRM change every day. The brain holds a pointer to those and goes and asks when it needs an answer. The moment you copy a database you own a stale copy that will confidently tell someone the wrong number.

Tag everything CONFIRMED, INFERRED, or GAP. A brain that cannot distinguish what it knows from what it guessed will eventually lie to you with total confidence. And gaps are gold; they are your to-do list for context.

The finish line for this stage is one sentence: you can ask your AI anything about your business and it can find it. That is the milestone worth stopping to notice, and it is deliberately selfish. Make it undeniably useful for you before anyone else touches it. Permissions, team onboarding, and rollout are all easier when you are demonstrating something that already works.

The rhythms that make it compound

This is the part that separates a brain from a filing cabinet, and it is the part most people skip.

The daily close. One scheduled task at the end of your working day: what happened, what we learned, what shipped, what is open. Miss the habit and you have a folder. Keep it and you have something that is smarter every Friday than it was Monday.

Call filing. Recordings get transcribed, identified by who and what, and filed to the right person or area. The raw transcript lives once, in the vault, and everything else holds a pointer. If you have sensitive one-on-ones, they file to that person’s own desk, not to a shared note.

The question inbox. Every question nobody can answer from the brain lands in faq.md under Unanswered, the moment it is asked. Weekly, anything asked twice gets promoted into a playbook or a standing rule. New hires are the best source of undocumented knowledge you will ever have; they ask a thousand questions in week one, and every answer that currently lives only in someone’s head is captured for free.

The mechanism behind all three: the repo is the memory, not the chat. Anything decided, learned, or corrected that lives only in a conversation is already half lost. Saying “file that” is the single reflex that separates operators whose brains compound from operators whose brains plateau.

Re-architect, then prove it

Do not skip Module 5. The template got you here; this is where the brain becomes yours.

Re-architecture. Having now read everything you own, your Chief of Staff proposes the most intelligent structure for your actual business, as a before-and-after tree with its reasoning. Multi-entity, holdco, several locations, whatever is true. You approve, it applies, the map updates. Any time you bring in a large new batch of context later, run this again. The structure follows the knowledge.

Postmortem. What it learned, what surprised it, what gaps remain, what it would improve about the process. The answers improve the brain and the process at the same time.

Then prove it, three ways. Answer a real staff-level question from the playbooks. Run a roll-up across the whole business. And run the negative test: ask from a login that should not have access and confirm the answer comes back empty. Trust the walls only after you have tried to walk through them.

That last one is the part almost nobody does, and it is the only one that actually tells you whether your permissions work.

Karpathy’s framing for this, in the same talk, is the one worth keeping. He argues for partial autonomy over full agents, and for what he calls an autonomy slider: a dial you move deliberately as trust is earned, the way Tesla’s Autopilot exposes levels rather than one switch. His image for it is the Iron Man suit, an augmentation you wear and control, rather than a robot you release. Reads free, writes gated, sends never without you. That is not caution for its own sake. It is what makes it safe to give the thing more rope later.

Giving it hands

Now run that test from earlier. Ask post-writer for the same thing you asked it before you built any of this, and put the two side by side. Same skill, same prompt. The only variable that changed is what it knows about you, and that is the whole thesis of this post sitting in front of you as evidence rather than a claim.

There is one gap left. A brain that knows everything and can still only produce drafts leaves the work on you. Connectors close it: MCP is the standard that lets Claude reach your actual tools, and we have a walkthrough on connecting Claude to more or less anything.

Connect whichever one matches your bottleneck. For most operators building in public, that is publishing, which is why Blotato is the one worth wiring first. The skills draft; the MCP connection is what actually publishes to the nine platforms, manages the calendar, and now reads and replies to comments and DMs. Register it in Company/connectors.md so your Chief of Staff knows it exists and routes to it.

The other one worth wiring early is Google. Once the brain lives in Drive, connecting Google through MCP means Claude can reach it from anywhere rather than only from the machine with the synced folder — including from Claude on your phone, or from a scheduled job that runs while your laptop is shut. Gmail and Calendar in the same connection are what turn “it knows my business” into “it handled my morning,” because inbox triage and calendar defence are the two jobs almost everyone wants first and neither is possible from files alone. Connecting Claude to Gmail is the shortest path in; the general MCP guide covers the rest.

Worth being precise about which layer does what, because they are easy to confuse: Drive for desktop syncs the files, and the Google MCP connection lets Claude read and update them without that sync. You want both eventually. Local sync is what makes the work fast; the API connection is what makes the brain reachable when you are not at your desk.

One deliberate note on overlap. Blotato’s brand-brief captures your business details and brand voice for content. Your Company Brain captures your business. These are not competing, and you should not run both as setup interviews asking the same questions. Let brand-brief own the content and voice slice; the brain owns the business layer, with content as one division inside it. That is why the prompt above hands off to it rather than duplicating it.

Once a connector is live, the natural next step is an AI hire whose whole lane is that job. Give the content and distribution division a head: one specialist reporting to your Chief of Staff, whose lane is everything from source material to published post to the replies coming back, and whose tools are the Blotato skills and the connection you just made. It reads the same soul file everyone else does, so it publishes in your voice by default rather than being told the voice every time.

That is what AI-Team/ is for, and the hiring protocol is one repeatable pattern: name the bottleneck it removes, give it one lane that does not overlap anyone else’s, assign its skills, write its onboarding doc, set its gates, and log the hire. Then the fifth step becomes possible, which is the whole point of the previous four. The division runs on a cadence instead of on your attention.

A solo owner is complete with a Chief of Staff alone. Never force a hire, and never hire against a bottleneck you do not actually have.

The operator’s rules

Everything below comes from running this daily, on a real business, and from the first live installs. None of it is theory.

Talk out loud; the transcript is an asset. When you are thinking through a decision, vocalize it. Recorded speech becomes brain content for free. Silent clicking becomes nothing.

Say it the moment it surfaces. When something exists only in your head, capture it right then. “That region is a territory, not an entity, save that.” One sentence, permanently filed. The cost of capturing context is lowest at the moment it comes up and highest six months later.

Point, do not copy. Live systems stay where they are.

Never delete; archive.

When it misses something, ask where it saved it. Do not just repeat yourself. “I already told you this. Where did you file it? Make sure it is in ongoing memory.” That turns a miss into a guardrail. Repeating yourself turns it into a habit.

A correction that repeats twice becomes a standing rule. Say it explicitly: “add this to standing rules.”

Give it constraints early, not cleanup later. Never delete, write only here, confirm before sending. Stated once in setup, enforced forever. Much cheaper than auditing behavior after the fact.

Make it show, then approve. Migration plans as tables. Re-architecture as before-and-after trees. Merges as plain-word diffs.

One sprint, one purpose. Migration is not permissions is not team rollout. Mixing them is how two-hour setups become two-month stalls.

Weekly, promote. Quarterly, prune. Repeated learnings become playbooks. Once a quarter, ask what is stale, duplicated, or unused. Subtraction is maintenance.

Nine protocols from running one daily

These are what daily reality taught us past the basics. None are required on day one. Adopt them as you grow into the system.

1. The open ritual. The daily close captures the day; an open ritual aims it. Teach it one word: “status.” It reads the recent daily notes and the front door, then briefs you on where you are, what is next, and the one thing that matters most today. Sessions that open oriented do not wander.

2. The decision log. Daily notes capture what happened. A decision log captures why. One short numbered entry per meaningful decision: what you decided, why, what you rejected. Six months later, “why do we do it this way” has an answer instead of an argument.

3. The lessons file, separate from rules. A rule is a correction. A lesson is understanding, what a mistake taught you about how the business actually works. Rules stop the AI repeating errors; lessons make the next generation of rules smarter.

4. One canonical brain. Fold, do not fork. The strongest governance rule we run. Never spin up a parallel folder for a project or a second copy to try something. We did it once, and reconciling the divergence cost more than the convenience was worth. New initiative, new folder inside the brain, on the map. Version the kit; never version the brain. This is why RULE ZERO sits at the top of the prompt.

5. Delegation inherits every gate. As your AI team grows, specialists do work your Chief of Staff hands off. Delegation never launders a permission: every gate travels with the task to whichever agent executes it. And one class of work is never delegated at all. Identity, voice, and pricing change only in a direct conversation with you.

6. The leak-by-summary guard. Walls hold on files and leak through summaries. A personnel conversation filed correctly to the right desk can still bleed into a shared daily note as “discussed the write-up with a team member.” Teach the rule explicitly: a summary inherits the sensitivity of its source.

7. The quarterly drift audit. Brains rot quietly. Pointers to renamed folders, contradictory duplicates, stale documents marked current. Once a quarter, one prompt: find every reference to something that no longer exists, every pair of files that contradict each other, and everything marked current that has not changed in ninety days. Report first, fix on approval.

8. Improve it before it ships. For any artifact that matters, run one improvement pass before it goes out. What is working, what would make it better, propose, implement. One recursion is usually enough. The bar: never ship a version you would not honestly recommend over the previous one.

9. The name is the door, and there is only one door. Covered above, and it is the one people erode fastest.

The thread through all nine is the same move: take something that lives in judgment and make it live in the system. Orientation becomes a ritual, reasoning becomes a log, mistakes become lessons, taste becomes a gate. That is what makes a brain compound instead of just accumulate.

Six habits that separate operators

Boris Cherny created Claude Code at Anthropic. In his talk Practical Tips on How to use Claude Code he makes a point worth internalizing before any of the tips: the leverage is not exotic tooling or running twenty sessions at once. It is a small number of disciplined habits, and he describes his own usage as fairly ordinary. Here they are, with your Company Brain standing in for his codebase.

Set up once. Permissions, dictation, a comfortable theme. Five minutes, once.

Ask before you act. Live in question-and-answer mode until it feels boring. This is Cherny’s first recommendation for anyone starting out, and it doubles as onboarding: point it at a body of work you did not write and ask it questions until you understand the shape. It is also Milestone 1 of your brain. The hidden payoff is calibration, because you learn what the model handles in one shot versus what needs your hand on the wheel.

Plan before it writes. The classic failure is asking for something huge and getting the wrong thing built confidently. The fix costs one sentence: “before you write anything, make a plan and run it by me.” Shift+Tab now makes that a keystroke.

Give it your tools and your context. Keep the root context file short. Push detail into the folders where it belongs, one README per folder. Context lives with the territory.

Close the loop. This is the biggest quality lever there is, and it is the one Cherny keeps returning to: the ordinary workflows get dramatically better once the model has some way to check its own work. Never accept a first draft as final on anything that matters. Give it a checklist, a preview, a test, then let it iterate a couple of times. Pretty good becomes almost perfect.

Then scale it. Schedule the rhythms so the brain compounds without you. Learn headless mode: claude -p "your question" answers and exits, and it is the primitive underneath every scheduled automation you will ever build.

Two keys worth knowing, because nobody teaches them and they are half the mastery. Escape is the always-safe interrupt: stop it mid-work, correct one line of a twenty-line plan, have it redo. The pound sign files a correction into memory, which is the difference between fixing something once and fixing it every time.

The vocabulary

Every word the prompt uses, in one place. If you are adapting this format to your own business, or explaining it to someone else, this is the shared language it runs on.

The files

The tiers

The operating terms

FAQ

Do I need Blotato to build a Company Brain? No. The brain works on its own and it is useful the day you finish it. Blotato is here because publishing is the bottleneck for most people reading this, its skills are free, and it is the fastest way to see what a brain is actually for. Swap in whatever matches your own bottleneck.

Do I need the Company Brain to use the skills? Also no. Install the seven skills and they work. They just work generically, the way any skill does when it knows nothing about the business it is writing for. The brain is what closes that gap.

Do I need Claude Code, or will Claude Desktop work? Both work. Claude Code is the better home for a Company Brain because it reads and writes real files on your machine at full speed, and because scheduled tasks run there. Desktop and Cowork read the same folder, and Blotato’s skills install on all of them.

Do I need Google Drive? Not to start, and you can add it at any point without moving anything. But it is the intended architecture rather than an optional extra: Drive for desktop makes the cloud folder be your local folder, so the truth lives in the cloud while the work stays local and full speed. It is also what makes the brain survive a dead laptop and what makes team permissions possible later, because permissions are just Drive sharing. Set up Google Workspace with a Shared Drive before anyone else joins you.

Is Obsidian still part of this? Yes. It is not an older way of doing it. Obsidian points at the same folder and gives you graph view, backlinks, and full-text search over the same markdown files, with no migration and no second copy. Local for the work, Drive for the truth, Obsidian for seeing it. One folder, three views.

How long does this take? Install is about ten minutes. The soul file takes fifteen to thirty. If you are harvesting existing context, the migration sprint takes one to two hours and can span days. You do not have to do it in one sitting.

What if I already have a folder like this? The prompt checks first and adopts what exists rather than building beside it. That is RULE ZERO. If it finds two candidates it stops and asks you which is canonical.

Will this work with a model other than Claude? The folder will. It is markdown. The specific commands here are Claude Code, but the artifact you are building is plain files, which is the entire reason to build it this way.

Does the AI change my identity file on its own? No. Identity, voice, and pricing change only in a direct conversation with you. Everything else it maintains freely below the index line.

What is not in this post? Team permissions and the teammate rollout, which are deliberately a separate session, and the deeper migration and multi-entity work we do with clients directly. Start solo. Make it undeniably useful for you first.

Sources

We cite and build. Every practice here carries a source you can check, plus what we learned running it on a real business. Where something is our own field result rather than someone else’s research, we say so.

Boris Cherny (creator of Claude Code, Anthropic) — Practical Tips on How to use Claude Code. The source for the six habits: one-time environment setup, question-and-answer before editing, plan before it writes, tools and context, closing the feedback loop, and headless mode.

Andrej Karpathy (founding team at OpenAI, former director of AI at Tesla) — Software Is Changing (Again), keynote at Y Combinator’s AI Startup School, June 2025. The source for Software 1.0 / 2.0 / 3.0, the autonomy slider, and the Iron Man suit framing. If you want the longer foundation underneath it, his Intro to Large Language Models and How I use LLMs are the best general-audience explanations of the machine your Chief of Staff runs on.

Anthropic — Claude Code documentation, the source of truth for install commands and anything that has changed since publication, and Building Effective AI Agents (December 2024), the design case for simple composable patterns over complicated ones.

Blotato, built by Sabrina Ramonov — the product, the free Claude Skills documentation, the skills repository, and MCP setup. Her own walkthroughs of the setup: How I Used Claude Code To Get 30M+ Views and I Built a $1M Marketing Team with 1 AI Agent (7 Skills).

Everything else here — the folder architecture, the module sequence, the nine protocols, the proof tests — is our own, developed running this system on our business every day.

Build the folder. Confirm the soul file. Set the daily close. Everything after that compounds on its own.

If you want the version with a served skill library, team permissions, and an install we run with you, that is what we do at Gen AI University and inside Agentic Society.