← Video Library · Workshops

What to Gather Before You Build a Company Brain

By Darby Rollins · Published August 6, 2026

We ran a two-hour workshop on building a company brain and never said out loud what you needed to have ready. The chat showed it. About a third of the room typed some version of “starting from zero,” two people were on a Microsoft stack we did not serve, and one person said plainly that he had not set anything up because he was worried about security and had come to find out how to do it safely.

This page is the prerequisite list that should have opened the session.

If you have already built an AI Chief of Staff with us, you have most of this. If you are starting cold, read the whole thing before you download anything. The full workshop is here when you are ready.

Where most people actually are

This is what the room reported when asked what they already had:

• Starting from zero. The largest single group, and three people said so in the first sixty seconds.
• Files in Google Drive with no structure. A folder is not a brain, but it is a head start.
• Everything in GitHub. Common among technical founders. Fine for code, wrong for SOPs and transcripts.
• Microsoft shops. SharePoint and OneDrive. The workshop did not serve them. See the honest note below.
• Partway there already, on the Chief of Staff plugin plus an Atlas folder, sometimes with Obsidian on top.
• Scattered across tools. ClickUp, Obsidian, Motion, Drive, iCloud, all at once.

The best summary of the problem came from one attendee in a single line: “How I ended up with three different brains.”

That is the common starting condition: not nothing, but too much, in too many places, none of it readable by anything.

What do you need before you start?

Ordered so that nothing below can be done until the thing above it is done.

The four hard blockers

1. A Claude plan that can run Claude Code. Claude Cowork is the floor and the second half of the workshop assumed Code. Max starting at $100 is enough. One attendee said the $100 plan handles everything he needs and that matches our experience, so treat the $200 tier as a preference rather than a gate.

2. Google Workspace, with the ability to create a Shared Drive. This one is easy to miss and expensive to discover late. Shared Drives and Google Groups do not exist on a free Gmail account. If you are on personal Gmail, you can still build a version one in a local folder, but the permissions model that makes this work for a team is not available to you.

3. Admin rights, or a named admin who will take your calls. Google Admin is where groups and folder restrictions get set. Without it you are building V1 only.

4. Google Drive for desktop, installed and signed in. Two seconds to install, and it is what lets Claude see the folder as a real local path instead of a streamed stub. Set the brain folder to mirrored, not streamed. This trips people up quietly: without it your AI reads placeholder files and writes get flaky.

Before the harvest is worth running

5. One staging folder with everything dumped in. Downloads, desktop, old drives, exports. We called it a content power hour on the call: one focused hour, gather it all, put it in one place, then begin.

6. Your existing corpus located, not sorted. Transcripts, SOPs, handbooks, brand docs, policies, vendor reports. Located is enough. Sorting is the AI’s job. The live demo ran against roughly 600 files. If you bring six, harvest mode has nothing to harvest and you want interview mode instead.

7. Meeting transcription already capturing calls. Granola, Fathom, Zoom, whatever you use. The strongest argument in the whole workshop is that your transcripts are the most valuable material in the business, and that only pays off if they already exist. If you are not recording calls yet, start today and the brain gets better every week without any further effort from you.

8. A decision about what is sensitive. Payroll, tax records, bank statements, client confidentials. The private folder is a prerequisite, not a later cleanup. If you are in a regulated field, redaction before ingestion is your work, not the model’s.

For the outcomes in the demos

9. Connectors authorized in Claude: Drive, Gmail, Slack.

10. MCPs for whatever you actually run on: ClickUp, Asana, QuickBooks, Stripe, Search Console. The Monday morning brief we showed is impossible without these. It is not a prompt, it is access.

11. Org chart, entity names, department owners, written down. The setup interview asks for all of it.

12. Your operating framework, if you have one. Quarterly rocks, KPIs, goals. If you have done the work of an EOS rollout, hand the whole thing over. That is exactly the material that makes a daily brief useful rather than decorative.

Only for V2

13. A Claude Teams plan with your team on it.

14. Google Groups created per role, not per person.

Four decisions to make before you touch a file

Each one is cheap now and expensive later.

Shared Drive or My Drive?

My Drive is one person’s personal account with no folder-level permissions, which means your only security is a written instruction asking the AI not to look. A Shared Drive is owned by the business, has permissions per person or group, and enforces them in Google Admin where the model has no reach.

The tradeoff is small but real: Shared Drive requires Workspace and an admin. My Drive does not. If you choose My Drive to move faster, understand you are not adding a team later without a migration.

Local folder or Drive-synced folder?

Claude points at a path either way. Local is faster and private, and it is what Obsidian users already have. Drive-synced is the only version your team can share.

Starting local is fine for V1 and costs you a reorganization at V2. That is a legitimate trade if you want to learn the system before you commit an org to it.

Google, Obsidian, or Microsoft?

Obsidian works. A local folder plus Obsidian is a simple, real setup. It gets limiting for teams and organizations, which is why the workshop used Drive. If you are solo and staying solo, use what you have.

Microsoft has no verified answer, and pretending otherwise would waste your afternoon. The method needs three things: plain files, a desktop sync client, and folder-level permissions the AI cannot talk past. OneDrive and SharePoint have all three, so it should work. What we have not done is test the folder structure against it, or map the permissions argument onto SharePoint groups. Four people asked about Microsoft during the session and none of them got a real answer. If you try it, tell us what broke.

Harvest mode or interview mode?

The starter prompt asks you this mid-run, so decide before you paste it.

Harvest points the AI at what you already have. Fast, and it inherits whatever mess is in the source. Right if you have a real corpus.

Interview has the AI ask you questions and build from your answers. Slow, and it produces exactly what you told it. Right if you are starting fresh or the business is new.

Asked directly during the workshop which to pick, the answer was one word: harvest. That assumes you have something to harvest.

The one that is not a decision

Copy, never move. Originals stay where they are, untouched. Copies land sorted, with sensitive files going to the private folder and the rest to team folders. Retiring the old drive is a deliberate step you take later, once you trust the system. There is no version of this where relocating your originals first is the right call.

What does it actually cost?

Numbers from the session itself, not from optimism.

The pre-work: one hour minimum if you are organized. A few hours if your files are scattered across a desktop, a downloads folder and two old drives.

The first real build: several hours, and it works better as an hour a day across a few days than as one long push. Austin’s number for the migration specifically was up to four hours of dedicated time.

The ingest run itself: this is the number nobody publishes. In the live demo the harvest started against about 600 files, and forty minutes later it was still going. It had not finished when we moved on. A larger corpus was reported at a couple of hours. You cannot use the machine for much else while it runs.

The part that does not end: it takes time to train this, the same way it takes time to train a person. The difference is that the training compounds and does not leave.

One calibration note. On the call the estimate was “the next hour and a half.” The session ran two hours. The time estimates on this material have historically run about a third light, including ours.

The lower-intensity path

You do not have to do a big-bang build.

Let your normal business cycles load the brain. When you sit down to write next week’s email campaign, teach it your email voice then. When you do the books, teach it your filing. When you onboard a client, build the client folder as part of onboarding. The brain gets built as a byproduct of work you were doing anyway, and the parts that get used most get built first.

That is slower to a complete system and faster to a useful one.

What the workshop assumed you already had

For completeness, since this is where the gap came from. The session assumed Claude Code, a Shared Drive, Google Admin access, Drive for desktop, existing connectors, meeting transcription already running, markdown literacy, and prior familiarity with the Atlas folder from our AI Chief of Staff training. It also ran its live demo against a folder that already held 600 files.

None of that was stated as a requirement, which is why one attendee typed “Overwhelmed” about an hour in and another asked whether he should be building along or waiting for a link.

If you have gone through the Agentic CEO workshop, you have the foundation this one builds on. Start there if any of the above is unfamiliar.

Next

Once you have the four blockers cleared and your files in one place, download the starter kit and read what is in it before you run anything. Then the full workshop will make sense end to end, and the playbooks are what you run once the folder exists.