Everyone has a brain design. That doesn’t mean it fits yours.
Scroll for five minutes and someone will show you their AI brain.
Markdown notes in a code repository. A database that searches by meaning. Thirty agents with job titles. A diagram with more arrows than a battle plan. Every builder is sure theirs is the one to copy.
Most of them are good designs. That’s not the problem.
The problem is that every one of them was built for a job – the builder’s job. A developer’s brain is shaped around a developer’s day. A creator’s brain is shaped around publishing. Copy it and you inherit their priorities, their tools and their trade-offs, whether or not you share any of them.
So, is it important to tailor your brain to its objective? For a few decisions, it’s the whole game. For most of the rest, it isn’t. Here’s how to tell the difference, and how we made those calls on our own brain.
What doesn’t need tailoring
Some things are true for every brain, whatever it’s for. Don’t reinvent these:
-
- One source of truth: every fact lives in exactly one document. Everything else points to it. Copies drift, and a brain with two versions of your pricing will quote the wrong one with total confidence.
-
- One owner per document: someone (or some AI employee) is responsible for keeping it current. A document nobody owns is a document nobody updates.
-
- A log of every change: what changed, when and who approved it. When an answer goes wrong, the log tells you why.
-
- Archive, don’t delete: you will want the old version back eventually.
That’s the floor. Above it, the right design depends entirely on what you’re asking the brain to do.
Start with what the brain is for
Ours has two jobs.
First, keep one narrative consistent across every AI employee. When a content writer, an outbound rep and a proposal writer all draw from the brain, they have to tell the same story, with the same proof, at the same price. That’s what controlling the narrative means in practice.
Second, make go-to-market teams faster without making them sloppier.
Both jobs fail the same way: the AI says something that isn’t true. A wrong price in a proposal. A customer result nobody gave you permission to use. An old positioning line that you killed a month ago, back from the dead.
So we ranked two characteristics above everything else. Accuracy first. Security second. Every design decision after that had to serve one or both.
That ranking ruled things out, too. More on that in a minute.
Decision 1: where the documents live
A lot of the brains you see live in a code repository, usually on GitHub. Plain text files, version history, fast for anyone who writes code.
We put ours in Google Drive. Three reasons.
-
- Security you can read is security you’ll get right: GitHub isn’t insecure. A private repository locked down properly is perfectly safe. But most founders we work with have never configured repository permissions, and a setting you don’t understand is a setting you’ll get wrong. They already run their company on Google Workspace. They know who’s in a shared drive and how to remove someone.
-
- Walls, not labels: sensitive material – contracts, pricing models, raw intake answers – sits in a separate restricted drive that the everyday brain can’t reach. Marking a document “confidential” is a label. A drive the AI has no access to is a wall. Only the wall works.
-
- Portability: the documents are ordinary Google Docs in the client’s own account. If a client stops working with us, they keep every document, and so does anyone they hire next. Nothing is locked in a format only we can read.
The trade-off is real: some things a developer gets for free in a repository take more work in Drive. For a founder with a technical team focused on priorities other than maintaining a brain,, we decided that was the right price to pay.
Decision 2: accuracy is designed in, not hoped for
AI will fill a gap with something plausible. The design has to make gaps visible and keep the AI from guessing. Here’s what that looks like in ours:
-
- Proof carries a status: every claim in the brain is marked verified, needs permission, or missing. The AI may only use what’s verified. Missing proof stays visibly missing – it never gets invented.
-
- Prices live in one place: every other document points to that one source. Change it once and every AI employee quotes the new number.
-
- Each AI employee reads a short list: three to eight core documents, by name, then only what the task needs. Reading less means fewer chances to pick up something stale, and it costs less to run.
-
- Accuracy gets tested, not assumed: a separate AI employee writes questions with known answers and scores the replies.
That last one taught us something. Our first test scored a perfect 20 out of 20. Then we realized the test was stored somewhere the AI being tested could read it. That isn’t an accuracy score. That’s an open-book exam. We threw the result out, moved the tests to a drive the tested employee can’t open, and had the tester run from its own separate login. The retest also scored 20 out of 20 – and this time it counts.
The tests are conducted routinely with dashboard reporting on it.
Take-away: If an attribute of your brain is important, you should routinely test and report on it.
Decision 3: Security is about who can change things
Most security talk is about who can see the brain. Just as important is who can change it.
-
- One writer: only one role files and edits documents in the brain. Anything important – a live document, a price, a claim – changes only with the owner’s yes.
-
- Outside material is untrusted: anything pasted or pulled in from the web lands in a holding folder until a person approves it. A web page can contain hidden instructions written for an AI. Treat outside text as information, never as orders.
-
- The checker isn’t the maker: the AI employee that tests the brain is separate from the one that maintains it, and it never fixes anything itself. Nobody grades their own homework.
What our objectives ruled out
This is where tailoring pays off. These ideas show up in a lot of popular designs. We said no to each one because it worked against accuracy or security:
-
- Copying customer data into the brain: contacts, deals and pipeline stay in the CRM, where they’re already current. A copy in the brain is stale by Friday.
-
- API keys stored in files: handy for developers. For us, it’s a password lying in a folder.
-
- Unlimited growth: we cap the number of master documents. A brain that keeps growing gets harder for the AI to search and harder for people to trust. When we hit the cap, something gets merged or retired.
None of these is wrong in general. They’re wrong for a brain whose job is to never say something untrue.
Five questions before you copy anyone’s brain
1. What’s the one job it must never fail at? That’s your first priority. Ours is accuracy
2. Who will look after it? A developer, or a founder with ten other jobs?
Pick tools that person already understands.
3. What must never leak? Put it behind a wall, not a label.
4. What happens when you leave the tool, or the vendor? If you can’t take it with you, it isn’t yours.
5. How will you know it’s right? If you can’t name the test, you’re trusting, not knowing.
Final thought
The best brain design isn’t the cleverest one in your feed this week.
It’s the one built for the job you’re giving it – and tested until you can prove it does that job.