← Video Library · Workshops

Build Your Company Brain: The Full Workshop

By Darby Rollins · Published August 6, 2026

On 6 August 2026 I spent two hours on a live call with Austin Distel, building a company brain in front of about eighty people. Austin was employee number one at Jasper. I was Jasper’s first paying customer, back when it was a GPT-2 demo and a hundred dollars changed hands between roommates for software that did not exist yet.

This page is that workshop, written down. Every framework, every demo, and the folder structure we handed out at the end. Nothing here is gated and nothing here is for sale.

If you want the short version: your AI does not have a capability problem. It has a context problem. And the fix is not a better prompt.

Everything taught on that call is on this page. You do not need to watch the video to get it, and you do not need to leave this page to finish it.

Resources from the workshop

Everything we handed out, in one place. All of it is free and none of it needs an account.

What is in the workshop

Timestamps match the video. Each one is a self-contained section, so you can jump.

What is a company brain?

Three things at once, and it needs all three:

Plain files, in a folder you already own, that your team’s AI reads from. Markdown, mostly. The simplest text format there is.

Connected to AI, with one named assistant that routes every question to the right file rather than reading everything.

Properly permissioned. Google Drive decides who sees what. Not a prompt.

The third one is where most setups fail.

Why does a Shared Drive matter more than the folder structure?

In My Drive, your security is an instruction asking the AI not to show things.

In a Shared Drive, security lives at the Google Admin level, and the AI cannot see past it.

Those two sentences are the whole argument. My Drive is one person’s personal account with no folder-level permissions, so the only thing standing between your front desk and your payroll records is a line of text asking the model to behave. A Shared Drive is owned by the business, carries folder-level permissions per person or per group, and enforces them somewhere the model has no reach.

Austin owns a chain of spas alongside his AI work. His front desk receptionist cannot read payroll. The Austin team cannot read the Dallas team’s files. He keeps an owner-only private folder for tax records and bank statements that only his own AI can see. All of that is enforced by Drive permissions, not by instructions.

One rule that goes with it, and this one is easy to skip: grant access to Google Groups, never to individuals. Roles outlive people. When someone leaves, you remove them from a group instead of hunting through folders.

The seat-by-seat version looks like this:

The same applies to Obsidian, which came up twice on the call. A local folder plus Obsidian is a working version one. It breaks at version two, when five people need scoped access to different parts of the same brain, because the permissions model is not there. If you are solo and staying solo, use what you have. If a team is coming, start where you can grow.

Why build one at all?

Six reasons.

Your AI gets context. The foundation. An AI that actually knows your business rather than the PDF you pasted into one chat window.

It grows with you. Your AI, then your team’s AI, then your AI agents.

You stop being the bottleneck. Knowledge stops living in one person’s head. One of our students is a CPA with a firm of six or eight people. He goes on vacation, lies down by the pool, and his bookkeeper texts him questions that a well-built brain could have answered without him.

Everyone contributes. Your team adds to the business’s intelligence, not just their own.

It is IP the business owns. Systems and processes stay when people leave. Or as Austin put it on the call: if anybody quits, if anybody goes on vacation, that means your IP is also on vacation.

The entity compounds. Every transcript, meeting and plan makes the business more valuable.

I believe also you’re doing a disservice if you’re playing solo player mode.

Austin Distel

What are rot and drift?

Two named failure modes. If you understand these you can diagnose your own setup in about a minute.

Rot is context that was true and is not anymore. You changed your pricing. You changed your org chart. You switched tools. Nothing in the folder knows, so your AI keeps confidently telling people last year’s answer.

Drift is what happens inside a single long session. The context window fills, compresses, and the nuance goes. The output starts sliding away from the point without anything obviously breaking.

Rot is a maintenance problem, and the fix is a change log per department plus the daily close below. Drift is an architecture problem, and the fix is the index: give the AI a table of contents so it opens one file instead of swallowing everything.

What are V1, V2 and V3?

The roadmap, and the order is not optional.

V1 is your source of truth. One place to ask anything about the business. Just you. This is the entire scope of the workshop and it is where you should stay until it works.

V2 is team self-service. Every employee has their own scoped Chief of Staff. This is a bigger step than it sounds, because it drags in security, compliance and governance as real work rather than as a checkbox.

V3 is agents that act. Drafting, publishing, scheduling. Safe because V1 and V2 came first.

Connecting an autonomous agent to a folder nobody has checked for accuracy produces a confident, tireless, wrong employee.

If you just jump ahead and you start building AI employees that connect into this, it’s like hiring a real employee, a human employee, without putting them through training or giving them any rounds of introduction to the tools they’re gonna be using on the job, to the org chart.

Austin Distel

What actually goes in the brain?

Evan Brady asked the sharpest question of the session: how do you think about evergreen context versus real-time context?

Evergreen is truth that holds. Policies, handbooks, SOPs, org charts, brand guidelines. Start here, because it is the smallest set that makes the AI useful, and because it is the part you already have written down somewhere.

Ever-growing is the heartbeat. Call transcripts from your one-on-ones, team meetings, project kickoffs, sales calls, customer success calls. Vendor reports. Daily notes.

On the second category:

I believe the transcript is actually one of the most valuable parts of your business because it’s the communication with different people.

If every one-on-one with your head of marketing is in the folder, you can ask how long you have been discussing the same problem. If thirty sales calls a month land there, you can ask which ones closed and what the closers did differently, and then train the thing to coach you. If your SEO agency’s monthly report goes in alongside access to Search Console, you can audit your own vendor.

None of it works on a folder of PDFs nobody indexed.

What does the folder actually look like?

Six pieces. This is the stack, and each one has exactly one job.

Two mechanics hold it together.

Every folder teaches the AI how to use it. A README.md sits inside each one carrying directions, SOPs and formatting rules. Claude reads it first when it opens the folder, then follows it. Changes get logged back into the file, so the rules stay current instead of rotting.

When you do not know where something goes, inbox it. One folder, called File Inbox - Ready to Organize. Anyone drops anything there with no filing decision. Vendor reports emailed in land there too. Your Chief of Staff sorts and logs all of it.

Removing the filing decision is what gets the material actually filed.

What does a soul file have in it?

My Chief of Staff is called Riff and I have been building it for about eight months. Its file is detailed about how I want it to operate, who I am, what my priorities are, and where I want it to hold a line.

Some of what is in there is not business at all:

Hey, every Friday is date night. If I skip date night this week, I need you to call me out on Friday.

That is the kind of instruction that makes an assistant into a chief of staff. It knows what matters to me, not only what I am working on.

One point of confusion worth heading off, because it will bite anyone who watches the workshop and then opens the kit. In the starter kit, soul.md is the company’s identity: what you sell, how you decide, your red lines. It is owner-gated, and nothing else in the setup proceeds until you confirm it. In my live demo, the file I was describing is my Chief of Staff’s persona. Both are called soul.md. Put the charter material in the wrong one and you will trip the setup gate. The kit’s version is the company. Your assistant’s personality lives in its own skill file.

How do you keep it from eating your token budget?

Julie Hutchinson asked this during the build, watching the harvest run: how did you get it to not use so many tokens?

Two answers.

Vectorizing, which is a heavy word for a simple habit: signal, not noise. A client call includes the weather and the small talk. Before it gets written down, have Claude keep the meaningful part and drop the rest.

The index, which does the real work. You do not hand the AI the whole folder. You hand it a table of contents, and it flips to the entry it needs. Think of a librarian: you do not ask them to read every book in the building to find one fact, you ask them where it lives.

The AI, when you have it set up and organized in a way that even a human can go find this stuff, it’s like, oh, you want to talk to client, this thing tells me clients folder is right there.

There is a third, smaller lever: ask Claude to audit its own token efficiency and confirm prompt caching is on. It will not always do that by default.

What is the daily habit that makes it work?

At the end of the day, ask for a wrap-up and a summary written into the daily notes. Tomorrow you open the folder and the context is already there.

Three things make that loop compound rather than just accumulate:

Give feedback like you would to a person. When the output is off, say what you got, say what you wanted, and have it write the correction down. This is training, not prompting.

Keep a mistakes log. I have a standing rule: if you ever make a mistake more than once, document it here so we never make it twice. The list grows for a while and then stops growing, because it stops making the mistakes.

Ask “even better if.” A tiny reviewing habit. Hand it a process and ask what would make it better. Sometimes the answer is nothing. Sometimes it surfaces something you would never have thought of.

When do you stop building?

There is a point where the system is good enough and your attention is worth more somewhere else. You do not need to spend two months building the perfect daily brief automation. Get the daily notes and the brief running, then leave them alone and check in weekly.

The question I have my AI ask itself every day is where the small hinge is: the one thing that swings the biggest door, the eighty-twenty we might be missing. That question is more valuable than another week of tuning.

The same instinct applies to learning. When somebody who builds these tools publishes how to use them, take the transcript rather than the afternoon. Boris Cherny put out twenty minutes on getting more out of Claude Code. Andrej Karpathy has been ahead of most people on why plain markdown files work this way at all. You can hand either transcript to Claude and ask what you are not using. You can do that with this page too.

Common questions

Twenty-eight questions came in during the session, and a fair number never got a proper answer live. The full ledger, including the ones we got wrong or skipped, is on the prep page. The short version of the ones asked most:

Can I use Obsidian instead of Google Drive? Yes for V1. A local folder plus Obsidian is a real setup. It limits you at V2 when your team needs scoped access.

Can I use OneDrive, SharePoint or Microsoft 365? Probably, with edits, and we have not tested it. The method needs plain files, a desktop sync client, and folder-level permissions the AI cannot argue with. Microsoft has all three. What we cannot promise is that this folder structure drops in unchanged, or how the permissions argument maps onto SharePoint groups. Four people asked Microsoft questions on the call and none of them got a real answer, which is our gap to close rather than yours.

Do I need the $200 plan? No. Claude Cowork is the floor. Claude Code is the recommendation, and Max starting at $100 is plenty to do everything on this page. One attendee said the $100 plan handles everything he needs and he is right.

Harvest or interview? Harvest if you already have a corpus. Interview if you are starting cold. If you have five hundred files, do not sit through five hundred questions.

Should I version control the markdown files? Google Drive’s own version history is thin. Keep a change log as a markdown file inside each department folder and instruct Claude to append to it on every change. That log is portable and does not depend on the platform.

How does Gen AI University tie to Agentic Society?

This got asked twice on the call and evidently did not land the first time, so here it is cleanly.

They are a public strategic partnership, not a merger. Austin Distel hired me several months ago as chief agent officer for Agentic Society, where I built the Agentic OS along with the trainings and workshops its members use. Gen AI University stays exactly where it is: YouTube, monthly workshops, and the community. Agentic Society is the deeper tier you step up into from anywhere inside Gen AI University, in person in Austin or virtually.

If you want the in-person version, that runs out of Austin and there is a conversation involved before anyone joins. If you are not in Austin, message me and I will point you at the virtual track.

What Agentic Society is

A membership for operators who want this built with them rather than explained to them. A year includes:

• The AI Agent Boardroom in Austin, twice a month, first and third Friday, 9am to noon on East 7th Street
• Weekly AI Breakthrough calls, ninety minutes every Tuesday
• Hackathons every 90 days. The next Agentic Hackathon is 15 and 16 October
• Online training sessions like this one, and the playbooks we build for private clients

Founding membership closed on 8 August 2026 and rates went up. Current pricing, for both the in-person membership and the online track, is at agenticsociety.com.

Austin screens for fit, so there is a conversation before anyone joins: agenticsociety.com/meet.

Three results from members. One owns an HVAC company here in Austin and built an outbound agent that finds apartment management companies and writes the cold email; that brought in $400,000 of new revenue in ninety days. One is an Austin journalist who grew to 40,000 Instagram followers on the back of a daily content research agent. One runs an SEO agency and went from thirty employees to nine, delivering two to three times faster, with more margin than before.

Everything else on this page is free and stays free, including the kit.

Where to go next

Everything above is the workshop. These three go deeper on one part each, and none of them is required.

1. What to gather before you start. The prerequisites this session assumed but never stated. Read it if you are starting from zero, which about a third of the room was.
2. The starter kit, documented. All 46 files, what the prompt does, and what to expect when you run it.
3. The playbooks. The five demos written out so you can run them without the video.

Two other trainings sit either side of this one. The Agentic CEO workshop is the prerequisite this session assumed you had done, and it is where the AI Chief of Staff gets built in the first place. 10,000 Hours of Claude Skills is the solo and local track: one prompt, no kit, no Drive, complete on day one.