How to Train AI to Write Like You: The Rules Doc (20 Minutes, Out Loud)

You would never expect a new hire to be ready to go out of the box. Even a CMO with twenty years of experience still gets trained on your business: your goals, your inner workings, how you operate. Your AI deserves the same onboarding, and this guide is how you write it.

The deliverable is a rules doc: the written standard your AI reads every time it works for you. One doc, two levels, and you can talk most of it into existence.

Level one: the company-wide rules

These are the rules that apply no matter what the AI is doing for you. Answer these, in writing:

• What's the business? One paragraph. What you sell, who buys it, what you're building toward.
• How do we write? Your tone, described specifically. Skip "professional but friendly" (every brand on earth says that) and describe it like you'd describe a person.
• What do we never do? Words you can't stand, formats you hate, topics that are off-limits, claims you'd never make.
• How do we operate? The defaults: how you like information delivered, how decisions get framed, what "done" means to you.

Level two: the per-role rules

The same way your team has SOPs per job, your AI needs guidelines per job. For each role you want it to play (copywriter, researcher, inbox assistant), answer:

• What does this role produce, and why?
• What does success look like? What does poor look like?
• The do's and don'ts specific to this work
• Templates it should follow
• Past examples of yours that worked, that it can model
• External examples you admire and want it to learn from

Some people set these up as agent files, some as skill files, some as one long doc with sections. The structure matters less than the existence. What matters is that the rules live somewhere the tool reads every time, not in a one-off chat it forgets.

The 20-minute method: talk it, don't type it

You don't have to draft any of this from scratch. Start a new chat, turn on voice, and say something like:

We're going to create a set of instructions for [copywriting] for my
business. I'm going to talk through how I like this done, give you
examples I love and examples I hate, and I want you to find the
patterns and turn them into a clean, organized rules doc I can reuse.
Ask me questions where I'm vague.

Then just talk. Describe the style like you'd describe it to a friend. Name accounts that do it well. Paste in two or three pieces of your own that worked. When you run out of things to say, have it show you the doc, and fix what it got wrong.

Here's the register I'm going for with mine, as an example of how specific to be: colloquial, like you're talking to your best friend at a bar, but the smartest friend you've ever had. Comfortable and smart, never pandering, never condescending. Aspirational, like the person you want to be. That's a rule an AI can actually follow. "Be engaging" is not.

Expect 70 to 80%, then train the rest

The first work your AI does against the rules doc will be 70 to 80% right. That's not failure. That's a new hire's first week.

The move that separates people who get great AI output from everyone else happens here. When a draft comes back off, don't silently fix it and move on. Give feedback: this part is good, this needs to change, here's what I wanted instead. Then say the sentence that makes it all compound:

Save this feedback to your rules so I don't have to give you this
feedback again.

Silent fixes teach the tool nothing, and you'll be making the same correction forever. Saved feedback means you never give the same note twice. That's the entire difference between correcting your AI and training it.

FAQ

How is a rules doc different from a prompt? A prompt is instructions for one task. A rules doc is the standard that applies to every task: your voice, your non-negotiables, your definitions of done. You write it once and every prompt inherits it.

What should I put in my AI's custom instructions? The two levels above: company-wide rules (what the business is, how you write, what you never do, how you operate) and per-role guidelines (what success looks like for each job, with real examples).

How long until my AI actually sounds like me? The rules doc gets you to roughly 70 to 80% on day one. The feedback loop closes the rest: each saved correction compounds, and within a few weeks of normal use the drafts stop needing the same notes.

Does this work for a team, not just me? Yes, and it's better for a team. One shared rules doc means every person and every tool works to the same standard, instead of five people prompting five different voices.