Attune · the architecture we use

Set It Up

How to give an AI assistant what it needs to actually help you.

Five things, in order. Works in Notion, Google Drive or Microsoft — pick the one you already use.

The idea, in one page

Most people open an AI tool and start typing. It doesn’t know you, so it guesses. The answers come back fine — and generic.

The fix isn’t a better prompt. It’s giving the AI a few things to read before it answers.

One

Who you are

One page. Your values. How you work. What you’d never hand over.

Not a mission statement — the real stuff, written plainly.

Two

What’s true right now

One page. Your prices, your offers, what’s live, what you stopped doing.

This is the one almost nobody has. It’s also the one that saves you.

Three

What you’ve decided

A folder. One short note per decision: what you decided and why.

This is how the AI stops asking you the same question every month.

Four

The AI itself

Now — and only now — you set up the assistant, and point it at one, two and three.

The tool is fourth. Not first.

Five

A weekly check

Thirty minutes. What changed? What’s true now? What did we decide?

Without this, everything above goes out of date and the AI starts confidently telling people things that stopped being true.

The order is the whole trick. An AI on top of nothing gives you confident nonsense, faster.

Two rules that matter more than the tool

One page, one job. The “who you are” page answers one question. The “what’s true” page answers a different one. Never let a page do both — it will go out of date on one of them and you won’t notice.

Put a date on “what’s true,” and make it mean last checked. Not last edited. Those are different. A page edited yesterday can still be wrong — and that’s how you end up with an AI that’s confidently out of date.

Build it where you already work

Same five things. Different furniture.

🟦 If you use Notion

Thing	Make this
Who you are	A page called Compass. Top level, easy to find.
What’s true	A page called What’s True, right next to it. Today’s date at the top.
Decisions	A page called Decisions, with a short sub-page for each one.
The AI	Claude with the Notion connector on. Tell it: read Compass and What’s True before answering.
Weekly check	A repeating calendar event. Update the date when you’ve actually checked.

Notion is the easiest for linking things together, and the fiddliest for controlling who sees what. If some of this is sensitive, put it in a separate space rather than hiding it in a corner.

🟥 If you use Google Drive

Thing	Make this
Who you are	A Google Doc: Compass.
What’s true	A Google Doc: What’s True. Date at the top.
Decisions	A folder called Decisions. One short Doc each, named by date: 2026-08-05 — pricing.
The AI	Claude or Gemini with Drive access. Keep all four in one folder and point it at that folder.
Weekly check	A repeating Google Calendar event.

Docs keep version history automatically, so you can always see what changed and when. Put everything in one folder — scattered files are the main reason this fails on Drive.

🟩 If you use Microsoft 365

Thing	Make this
Who you are	A Word doc: Compass, in SharePoint or OneDrive.
What’s true	A Word doc: What’s True. Date at the top.
Decisions	A SharePoint document library called Decisions — one file per decision.
The AI	Copilot pointed at that SharePoint site, or Claude with the connector your IT allows.
Weekly check	A recurring Outlook meeting.

Keep it in one SharePoint site, not scattered across personal OneDrives. The usual failure here: the AI can only see half of it and nobody realises — so it answers from the half it can read. Check what it can actually see before you trust it.

Getting the AI to use it

Whatever the tool, put something like this in its instructions:

Before you answer anything about our business, read:
  1. Compass — our values and how we work
  2. What's True — our current facts

If anything else disagrees with What's True, What's True is right,
and tell me, because something else is out of date.

Never make up a fact, a price, or something we promised someone.
If you don't know, say so. Guessing does real damage.

At the end of a working session, write me a short note:
what we decided and why. I'll approve it before it's saved.

That’s genuinely most of it.

If you want it to push back on you

This is the part people skip, and it’s the part that makes an assistant worth having.

An AI that knows you well gets more persuasive, not just more useful. If it agrees with everything, it will agree with you all the way off a cliff — and you’ll trust it while it happens.

Disagree with me when:
  • this contradicts something I said I value
  • this reverses a decision I already made
  • this is the third time this has come back
  • this can't be undone and we haven't talked about the risk

When you disagree: say what the tension is, say why, quote what
*I* said I valued, and then let me decide. Say it once, then do
what I asked.

Don't argue about wording or style. That's mine.

That last line matters. An assistant that objects to everything gets ignored inside a week — and then it can’t object when it actually counts.

How to tell if it’s working

Ask these once a month. Be honest.

Does it still sound like us? If you’re rewriting more than you used to, something’s drifting.
Can we use what it gives us? Or do we rebuild it every time?
Do we believe it? Does it ever say “I don’t know”? If never, that’s a warning.
Did it give us time back — and where did that time go? If the freed hours quietly filled with more work, the AI worked and the system failed.
Has it disagreed with anyone lately? If it hasn’t in a month, it’s become a mirror.

And when to stop

Decide this now, while you’re calm. If you can’t say what would make you stop, you’re not evaluating — you’re just hoping. Ours:

It hasn’t disagreed with anyone in a month.
Someone still has to ask a person for what the AI could have answered.
We shipped something and the record didn’t reflect it.

Any one of those and we stop and look at the design again. We don’t just adjust it.

What usually goes wrong

The pages go out of date. Most common by far. The AI doesn’t get vague — it gets confidently wrong.

Nobody but the person who built it uses it. Then it isn’t a system, it’s one person’s setup, and everyone else is stuck waiting on them.

It becomes agreeable. Slowly, and it feels like things going well.

You stop deciding. The slow one, and the serious one. Every so often ask: did I decide that, or did it decide and I agreed?

Start with one page. “Who we are.” Everything else can wait a week.