Stop Prompting, Start Briefing: How to Set Up AI So It Actually Knows Your Business

Most owners I talk to have tried AI for marketing exactly once. They typed "write a blog post about water heater repair," got back something that read like a brochure for nobody, and decided AI content is junk.

The tool wasn't the problem. The setup was. There's a name for fixing it: context engineering. It sounds technical. It isn't. It's the difference between yelling an instruction at a stranger and briefing a new hire who already knows your business.

The prompt is the least important part

Everybody online is selling prompts. "75 ChatGPT prompts for small business." Copy, paste, done. I understand the appeal. It feels like a shortcut.

But a prompt is just the question. The best question in the world still gets a mediocre answer from someone who knows nothing about you.

Think about hiring a new office manager. You wouldn't walk in on day one, say "handle the customers," and leave. You'd tell them what you sell, what you charge, who calls, what those callers always ask, and what you never promise on the phone. The instruction is one sentence. The briefing is the actual job.

What is context engineering, in plain English?

Two words to keep straight:

• Prompt. The thing you type. The question or the task.
• Context. Everything the AI can see when it answers: your standing instructions, the files you gave it, the earlier messages in the conversation, and any examples of what good looks like.

Context engineering is deliberately deciding what goes into that second bucket. Anthropic, the company that makes Claude, describes it in its engineering guide as curating the right set of information the model sees, as opposed to just wording the instructions well.

That's the AI labs' version. For a small business, it boils down to one thing: a folder. A place where the AI keeps everything about your business, so you never have to explain it twice.

Here's why it matters so much. The AI only knows what's in front of it. When the folder is empty, it fills the gaps with the most typical answer it can produce, which is why generic prompts get generic output. I go deeper on that in why all AI content sounds the same.

What goes in the folder

Before I ask AI to write a single word for a client, I load their setup with these:

• Voice. How you talk, with examples of what sounds like you and what doesn't.
• Services and prices. What you do, what you don't, rough price ranges, and the towns you serve.
• Customer questions. The things people ask on the phone, in their words, plus the objections that kill deals.
• Proof. Your experience, your process, and what actually makes you different. Real things only.
• Rules. Words you never use, claims you can't make, formatting you hate.
• Samples. Three to five pieces of writing that sound like you: emails you sent, review replies, a page you wrote yourself.

I walk through each one with a fill-in checklist in the 5 documents your AI needs before it writes anything. Most owners can get a solid first version done in one afternoon.

How to set up an AI project in Claude or ChatGPT

Both of the big tools now have this folder built in, and both call it a Project. A project holds your files and a set of standing instructions, and every chat you start inside it can see them. Here's how I'd set one up.

1. Step 1

   Pick one tool

   Don't build the same folder in three places. Pick the one you already pay for or like using. I use both, but an owner only needs one.

2. Step 2

   Create the project

   In Claude, Projects lives in the left sidebar. Anthropic's help article on projects says they're available on free accounts too, with free users capped at five projects. In ChatGPT, you create a project from the sidebar as well. OpenAI's projects help page lists the file limits by plan: 5 files on Free, 25 on Plus.

3. Step 3

   Upload your documents

   In Claude this is called project knowledge. In ChatGPT they're project files. Plain documents work fine. Keep each one focused on a single topic so you can update it later without hunting.

4. Step 4

   Write short instructions

   In Claude, use "Set project instructions." In ChatGPT, open the project's menu and choose Project settings. Keep it to a few lines: who you are, who the reader is, always follow the voice document, never invent prices, facts, or reviews, and ask when something is missing.

5. Step 5

   Test it with a real idea

   Hand it one topic a customer asked about last week. Read the result like a skeptical customer. When something's off, fix the document, not just the chat. That's how the folder gets better every week.

A note on custom GPTs, since a lot of older advice points there. OpenAI's help page on creating a GPT now says personal plans (Free, Go, Plus, and Pro) can't create new GPTs, though existing ones still work. For most owners, a project does the same job with less fuss.

Before and after: generic prompt vs. briefed project

Say you run a made-up three-truck HVAC company called Gateway Comfort Heating and Air. You ask for a post about a furnace making a banging noise. Same request both times. The only difference is whether the AI has the folder.

The second version isn't better because of a clever prompt. It's better because the AI walked in already knowing the business. That's context engineering doing its job.

Why this beats one-shotting a blog

"One-shotting" means asking for a finished blog post in a single request and publishing whatever comes back. It's the most common way people use AI for content, and it's why so much of it reads the same.

A briefed project fixes the knowledge problem. The other half is process: researching, outlining, drafting in sections, then editing against your voice doc. I break that sequence down in why "write me a blog post" gets you garbage.

The real payoff is that you only build the folder once. The same project writes your blog posts, your Google Business Profile posts, your review replies, your ad variations, and your follow-up emails. Every one of them starts from what the AI already knows about you.

That's what changes the economics. Once the setup exists, your job shrinks to handing it ideas and checking the results. That's a big part of why one experienced person can now produce what used to take a small agency team, which I covered in the agency model is dead.

Mistakes I see when owners set this up

• Dumping everything in. A 60-page employee handbook buries the stuff that matters. Give it focused documents, not your whole hard drive.
• Never updating it. Prices change, services change, you add a town. If the folder is stale, the content is wrong.
• Rules that only live in chat. If you told it "no exclamation points" in one conversation, it won't remember in the next. Put rules in the documents or the instructions.
• Publishing the first draft. A briefed AI is a great first-draft writer. It's still not you. Someone who knows the business reads everything before it goes live.
• Letting it guess. If your docs don't include a price, the AI may make one up. Tell it plainly to leave a placeholder instead.

Where to start this week

Don't start with prompts. Start with the folder. Open a project, write a one-page description of your services and service area, paste in three emails you've written to customers, and add five questions people ask you every week.

Then hand it one idea and see what comes back. I think you'll be surprised how much better it gets when it actually knows who it's working for.

If you'd rather have someone build and run the whole setup for you, that's exactly the kind of thing I do in workflows and tools.