The AI Foundations Checklist

Every chat starts with four sentences about your business

You know the opening. Something like: “So I run a six-person studio, we do brand and web for allied health practices, our clients are usually…” And that’s all before you can even ask for the thing you came for.

Then you close the chat, and tomorrow you type it all again, slightly differently. Which means you gave it a slightly different brief, so of course you got a slightly different answer.

I think this is one of the most fixable problems in AI, and it’s fixed once, not every day. There are nine pieces. Set them up and every prompt after them gets shorter, and the output gets closer to something you’d actually send.

Do them in order. The first three do most of the work, and honestly, plenty of people could stop there and lose very little.

The first three are the ones that matter

1. The business context file

One page that describes your business the way you’d explain it to a sharp new hire on their first day.

Without it: everything comes back written about a generic version of your industry, and you edit your own specifics back in by hand, every time.

What goes in it: what you sell and what it costs, who buys it and who doesn’t, the words your clients use for their own problem, how you deliver the work, what makes you different in one plain sentence, and the things you never say.

BUSINESS CONTEXT
We are: [name, size, location, what we do in one line]
We sell: [offers, prices, typical engagement length]
Our clients: [who they are, what they were doing before us]
They say things like: "[three real phrases from real clients]"
We are different because: [one sentence, no adjectives]
We never: [three things: claims, tactics, phrasings]

2. The voice file

This one isn’t a list of adjectives. It’s rules and examples.

Without it: the writing comes back competent and bland, so you rewrite the first paragraph every time, and eventually you stop using AI for anything a client will see.

What goes in it: rules a stranger could follow, plus before-and-after examples from your own writing. The examples teach it more than the rules do. The Voice File walks through every part, and has a prompt that builds the whole file from writing you’ve already published.

3. The standard

What “good” looks like in your business, written down.

Without it: nobody can explain why a draft is wrong, only that it is. Feedback stays at “make it punchier”, and the next draft is a coin toss.

What goes in it: a short pass/fail list for each type of thing you make often, like proposals, client emails or social posts. The easiest way to write one is to take your best recent example and describe what makes it good.

STANDARD: client proposal
Pass requires all five:
1. Names their problem in their words in the first two lines
2. Scope is specific enough to argue with
3. One price, no menu of options over three
4. No claim we can't evidence
5. Next step is a single action with a date

The next three tell it who it’s talking to and where things live

4. The audience file

One page per type of client you sell to. Start with two at most.

Without it: the same message goes to a first-time buyer and a long-term client, and neither of them feels like it was written for them.

What goes in it: what they already believe, what they’ve tried that didn’t work, the objection that comes up when you talk about price, and what success looks like to them (and to whoever they answer to).

5. Named, reusable prompts

A small library of prompts with actual names, instead of scrolling back through chat history hoping to find the good one.

Without it: you rebuild the good prompt from memory each time, a bit worse, and nobody else on the team can use the version that worked.

What goes in it: eight to twelve prompts, each with a name, what it’s for, what to paste in, and what good output looks like. Keep them in one place, and build each one on a structure that holds. Date them, and delete the ones nobody opens.

6. A file structure AI can read

Flat, clearly named and a bit boring. That’s the point.

Without it: you upload the wrong version of the proposal, or you spend the first few minutes of every task hunting for the source document.

What goes in it: one folder per client, dated file names, the current version at the top, and older work in an archive subfolder. 2026-09-12-clientname-proposal-v3.md beats Proposal FINAL (2) copy.docx every time.

The last three keep it running once the novelty wears off

7. The connector layer

This is where your AI can see your real work: your documents, calendar, CRM or inbox.

Without it: you become the courier. You copy out of one system, paste into a chat, and paste the result back. That copying and pasting is a lot of the time you thought you were saving.

What goes in it: start with the two places your work actually lives, and use read-only access wherever the tool gives you that option. Grant access on purpose, review it every few months, and keep a written list of what can see what. When you’re ready for it to do things as well as read them, build automations that tell you when they stop.

8. The recurring jobs list

The work that happens every week whether you plan for it or not.

Without it: AI stays something you use for one-off tasks, while the repeated work, which is where the hours really are, stays manual.

What goes in it: every recurring task, how often it happens, who does it and how long it takes. If you’re not sure what repeats, log it for a week first. Then sort by hours per month. The top three are your build list, and I’d bet none of them is the exciting one.

RECURRING JOBS
Task | Frequency | Owner | Minutes each | Minutes/month
Weekly client update emails | 6x weekly | Sam | 20 | 480

9. The review loop

Who checks the output, against what, before it goes out the door.

Without it: either you end up checking everything, which is exactly the bottleneck you were trying to remove, or nobody does until a client spots something.

What goes in it: a named checker for each type of output, the standard from piece three as their checklist, and a rule for what happens when something fails twice. Twice means fix the prompt, not the draft.

Once the foundations exist, your brief shrinks to a third

Here’s a brief written from scratch, the way most of us do it:

Write a follow-up email to a client who hasn’t replied to a proposal. We’re a six-person brand and web studio, we work mostly with allied health practices, our tone is warm but direct, not salesy, don’t use exclamation marks, the proposal was for a website rebuild at $18,000 over ten weeks, they went quiet after two weeks, they’d mentioned their practice manager needed to sign off, please don’t be pushy, and keep it under 150 words, and don’t start with “just checking in”…

And here’s the same brief once the foundations exist:

Follow-up email. Client: Harding Physio. Proposal sent 26 Aug, quiet since. Blocker they mentioned: practice manager sign-off. Use voice file and the client email standard.

The output is at least as good. Often it’s better, because the standard is doing work your tired 4pm brief wasn’t.

When not to bother

Don’t try to build all nine in a weekend. Three done properly beat nine done vaguely, and the last three only really pay off once the first three are being used.

Hold off on the connector layer if you handle health records, legal files, or anything under a confidentiality agreement you haven’t reread lately. Read it first, then decide.

Skip the audience file if you really only have one kind of client. Writing two near-identical files just gives the model something to get confused about.

And if you’re the only person who’ll ever use any of this, you can probably keep the standard and the review loop in your head and write down the other seven. A lot of process for one person is just overhead.

What you’ve got now

A business context file, a voice file and a written standard, so any AI tool you open already knows who you are. A prompt library with names. A file structure tidy enough to find things in. Your recurring work sorted by hours. And one person responsible for checking the output against something specific.

Write the business context file today. It’s one page, and the other eight all lean on it.