Teaching an AI your way of working

Two businesses use the same AI tool. One gets generic, vaguely-corporate output it always has to rewrite. The other gets drafts that sound like the owner wrote them on a good day. The difference is almost never the tool — it’s that the second business wrote down how it works, and the first one is improvising every request.

The insight: AI runs on the same thing employees do

A new hire doesn’t become useful through talent alone; they become useful when someone explains how this business does things — what we say, what we never say, what good looks like, what to do when X happens. AI is identical, with one twist: it only knows what you tell it in the moment. It has no tenure, no osmosis, no memory of last month’s correction unless that correction lives somewhere it gets re-read.

So the businesses that get consistent AI results all converge on the same artifact: the written playbook. Small documents that capture, once, the things you’d otherwise re-explain forever.

What a playbook looks like

A good playbook for one task fits on a page or two:

• The job: “Draft replies to new customer inquiries.”
• The voice: “Warm, direct, no exclamation marks, never call anything ‘amazing.’ Sign off with first name only.”
• The rules: “Never quote prices — say a person will confirm pricing. Always offer the booking link. If the customer sounds upset, don’t draft — flag for a human.”
• Examples: two or three real before/afters, which teach tone better than any adjective list.

Write one per task, not one giant manual — an inquiry-reply playbook, a proposal-summary playbook, a report-commentary playbook. Small documents stay accurate; big ones rot.

Why this compounds

The first payoff is better AI output. The bigger one sneaks up on you: you’ve documented your business. The voice rules, the pricing policy, the escalation triggers — that knowledge used to live in your head and walk out the door with departing staff. Playbooks written for the AI turn out to be the onboarding manual, the QA checklist, and the “what would the owner say” reference, all at once. Some of the most valuable process documentation we’ve seen was written for a machine and ended up training humans.

Where to start

Pick the one task where you correct AI (or staff) output most often — that correction pattern is the first playbook, dictated in twenty minutes. Test it, fix what it gets wrong, and move to the next task. If you’re deciding which tasks deserve this treatment at all, How to scope an automation project before you buy applies just as well at this small scale — and LLMs in plain English for business owners explains why the instructions matter more than the model.