AI tools are genuinely useful once they know how your business works. The catch is that they don’t remember anything on their own. Here is a plain-English way to build a structure that teaches yours, and keeps it taught.
Most people’s experience of AI goes like this. You ask it to do something. It does a decent job. You correct it a few times, it gets better, and you finish the day thinking this is going to save you real time.
Then tomorrow you come back and it has forgotten every single correction.
That is not a fault in the tool. It is how the tool works. An AI assistant starts each conversation with no memory of the last one, and no knowledge of your business beyond what you tell it in that moment. Left like that, you spend more time re-explaining yourself than you save.
The fix is to stop treating it like a chat window and start treating it like a new hire: give it somewhere to work, write down how you do things, and put the important rules somewhere it has to read them every time.
This is one way of doing this, not the only way. There are dozens of approaches and plenty of them are good. This is simply the setup we have found easiest to manage and quickest to get running, and it works with most of the serious AI assistants rather than being tied to one product. Take the parts that fit how you actually work.
Step 1: Give It a Home
The first move is the least technical and the most important. Create one folder on your computer or in your cloud storage, and make that folder the only place your AI works.
Most serious AI assistants can be pointed at a folder and given the run of it. Once that is set, the folder becomes the boundary of the job. Everything the assistant needs is inside it. Everything it should not touch is outside it.
That boundary does two things at once. It keeps your private files private, and it gives you a single place to put everything you are about to teach it. If you take nothing else from this article, take this: the folder is the foundation. Get it wrong and nothing above it holds.
Do not point an AI assistant at your whole computer, your entire Documents folder, or your company file share “so it can find things.” Give it exactly the working folder it needs. You can always add more later. Taking access back after it has wandered somewhere it shouldn’t is a much less pleasant conversation.
Step 2: Separate the Kinds of Work You Do
Inside that folder, resist the urge to dump everything in together. Instead, ask yourself a question you have probably never had to answer before: what genuinely different kinds of work do I do?
Not clients. Not projects. Kinds of work, where the rules and the finished product are actually different. For most small businesses that shortlist comes out around four or five things. A professional services firm might land on quotes and proposals, ongoing client delivery, marketing content, and internal admin. A trades business might split estimating, job documentation, and customer communication.
The test is simple: if two kinds of work would need different instructions, they need different folders. Writing a proposal and writing a newsletter are not the same job. The tone is different, the format is different, and what counts as “finished” is different. Keeping them apart means you can give each one its own rules without those rules bleeding into the other.
So the shape looks roughly like this:
- One top folder that the AI is bound to, holding the rules that apply to absolutely everything
- One folder per kind of work underneath it, each holding the rules for that kind of work
- One folder per client or project inside those, holding the specifics of that job
This is ordinary administrative filing. The only unusual part is that you are organising it for a reader who takes everything literally and has no common sense to fall back on.
Step 3: Write the Rules That Apply Everywhere
In the top folder, create a plain text or markdown file for your rules. Most AI assistants look for a specific filename and read it automatically. Check what yours expects and use that name, because a rules file it never opens is just a document.
This top-level file is for the things that are true no matter what you are working on. Ours covers things like:
- Who you are. Business name, address, phone, the person in charge. Sounds trivial. Stops a dozen small errors a week
- How you want to be spoken to. If you want short answers, say so plainly. “Give me the answer, not the reasoning” is a legitimate rule and it will be followed
- What is off limits. Folders it must never touch, actions it must never take without asking
- The habits you expect. Wait for approval before making changes. Never invent a fact. Ask rather than guess
The useful part is that these rules trickle down. Anything in the top file applies to every folder beneath it, forever, without you repeating yourself. Write “never send an email on my behalf, only draft it” once at the top, and it holds whether you are working on a proposal, a website, or a spreadsheet.
Step 4: Write the Rules for Each Kind of Work
Now put a second rules file inside each of those work-type folders. This is where the specifics live, and it is the step most people skip.
A rules file for your quoting folder might set out how a quote is laid out, what your standard terms are, what your minimum charge is, and which things always need your sign-off before they go out. A rules file for your marketing folder might set out your voice, your house style, the words you refuse to use, and how long a piece should be.
Those two sets of rules would fight each other badly if they lived in the same file. Kept apart, each one only applies where it belongs, and the AI picks up the right set based on where it is working.
It means you can be strict without being restrictive. Your quoting rules can be rigid, because quotes should be. Your marketing rules can be loose, because writing needs room. In one big file you would have to compromise both.
Step 5: Cookbooks and Skills, and the Difference Between Them
Rules tell your AI what it must and must not do. Two other things tell it how to actually do the work, and they are worth keeping separate in your head.
A cookbook is a reference
A cookbook is exactly what it sounds like: a collection of recipes for things you do often. The settings that produce your standard document layout. The wording of a clause you use in every agreement. The five steps you follow when onboarding a customer. The gotcha you hit last March that cost you an afternoon, written down so nobody hits it twice.
A cookbook is passive. It sits there and gets consulted when relevant. Its value grows every time you add to it, and the best entries are usually the ones you write immediately after something goes wrong.
A skill is a procedure
A skill is a named, repeatable job with a start and an end. “Produce a branded quote.” “Write and publish a newsletter.” “Run the monthly reconciliation.” A skill spells out the steps in order, what it needs from you before it starts, and what finished looks like.
The practical difference is how they get used. A cookbook is looked up. A skill is invoked. You say “run the quote procedure” and the whole sequence follows, the same way every time, whether it is a Tuesday in March or the week before Christmas.
| Rules | Cookbook | Skill | |
|---|---|---|---|
| Answers the question | What must I never do? | How did we do this before? | What are the steps? |
| Used | Always | When relevant | On request |
| Written when | You set a standard | You learn something | You repeat something |
| Looks like | A short list of musts | Notes and examples | A numbered procedure |
You do not need all three on day one. Start with rules. Add cookbook entries as you learn things. Turn something into a skill the third time you find yourself explaining it.
Step 6: Gates, for the Rules That Must Not Bend
Here is the part almost nobody sets up, and the part that matters most once you are relying on this.
Some of your rules are preferences. If they get missed occasionally, you shrug and fix it. Others are not. Sending something to a customer before you have seen it. Deleting a file that turns out to matter. Publishing a price that was never approved. Those cannot be left to good intentions, because written rules are only as good as whether they get re-read at the exact moment they apply.
For those, you want a gate: an automatic check that sits between the AI and the action, inspects what is about to happen, and refuses it if it breaks the rule. Not a reminder. A stop.
Most capable AI tools support this in some form, usually as a script that runs before an action is taken and can veto it. If yours does, use it for your genuine red lines, and keep the list short. Ours is small on purpose: the things that would be expensive, embarrassing, or impossible to undo.
A gate that quietly stops working is worse than having none, because you have stopped watching for the thing it was catching. Build a small test you can run that deliberately tries to do the forbidden thing and confirms it gets blocked. Run it now and then. If it ever passes when it should fail, you have found a problem before it found you.
Step 7: A Reload Command, Because Sessions Get Tired
This one is not obvious until it bites you.
A long AI session drifts. Early on it reads your rules carefully and follows them closely. Hours later, deep into the work, it is increasingly going on memory of what it thinks the rules said rather than what they actually say. Nothing announces this. The work just gradually stops matching your standards.
Two habits fix it:
- Close it down and reopen it regularly. Daily is sensible for most people. A fresh session starts clean
- Build yourself a reload command. Write a short procedure that says: re-read every rules file that applies here, re-read the cookbooks, test that the gates still work, and report back what you found
Then, instead of rebuilding your setup from scratch each morning, you start a session and run the reload. It re-reads everything you have ever taught it for that kind of work and confirms the safety checks are alive. It takes seconds and it is the single highest-value habit in this whole article.
It is also a genuinely useful moment mid-session. If something feels off, reload and it usually comes back into line.
This Grows With You
None of this is a project you finish. It is something you add to whenever you notice yourself explaining the same thing twice.
Every time you correct the AI, ask where that correction belongs. Is it a rule, because it must always be true? A cookbook entry, because it is a useful thing to know? A step in a skill, because it is part of a procedure you repeat? Put it in the right place and you never have to say it again.
Over time this reaches into work you did not originally have in mind. Drafting emails in your voice. Producing documents with your branding on them, laid out the same way every time. Keeping notes on ongoing jobs so nothing depends on somebody’s memory. It all comes from the same foundation: a folder it is bound to, rules it has to read, and notes it can look things up in.
The Honest Part: There Are No Guarantees
We would be doing you a disservice if we ended on the tidy version, so here is the uncomfortable bit.
AI does not obey rules the way software obeys code. It interprets them. It makes judgment calls. And it is genuinely, impressively good at finding routes around a rule that technically satisfy the words while missing the point entirely. Give it a rule and a deadline and it will sometimes find a door you did not know existed.
This is not a reason to skip the structure. Structure is what turns “constantly wrong” into “occasionally wrong in ways you can spot.” But go in with your eyes open:
- Teaching it takes patience. The first few weeks involve a lot of correcting. That is the job, and it does get better
- Even a well-taught setup will surprise you. Assume it will, and keep the gates on the things that would actually hurt
- Check the work. Especially anything going to a customer, anything involving money, and anything you cannot undo
- The foundation is what you control. How faithfully it is followed on any given day is not entirely in anyone’s hands, including ours
Anyone telling you they can make AI follow instructions perfectly is selling you something. What a good structure buys you is far fewer surprises, faster recovery when there is one, and an assistant that starts most days already knowing how you work.
Frequently Asked Questions
Do I need to be technical to set this up?+
For most of it, no. Creating folders and writing plain-English rules in a text file is the bulk of the work, and thinking clearly about how you actually work matters far more than any technical skill. The one part that usually needs help is the gates, since those involve a small amount of scripting. Everything else is organisation and writing.
How long does this take to put in place?+
A basic version is an afternoon: one folder, one top-level rules file, and folders for the different kinds of work you do. That alone will noticeably improve your results. The cookbooks and skills build up over the following weeks as you notice yourself repeating explanations. It is a foundation you extend, not a project with an end date.
Does this work with any AI assistant?+
The thinking applies broadly, but the mechanics vary. What you need is a tool that can be pointed at a folder and that reads an instructions file automatically. Most of the serious assistants aimed at real work can do this; the simpler chat-only tools generally cannot, because they have no way to see your files. Check what filename yours looks for and use exactly that.
What if I get the folder structure wrong?+
You will, a bit, and it is easily fixed. Folders can be renamed and rules can be moved. The most common early mistake is too few divisions rather than too many: everything lands in one pile and the rules start contradicting each other. If you find yourself writing “except when…” into a rule, that is usually a sign the work belongs in its own folder.
Why not just put everything in one big rules file?+
Because it stops being read properly. A very long instruction file gets skimmed rather than followed, and rules for unrelated work sit next to each other where they can be confused or applied to the wrong job. Splitting them means each file stays short enough to actually be used, and only the relevant ones apply to what is in front of it.
Is my business data safe if I do this?+
That depends on which tool you use and what you point it at, and it deserves its own decision rather than an assumption. Read your provider’s terms on whether your content is used for training, keep the AI’s access limited to the working folder rather than your whole system, and keep genuinely sensitive material outside that boundary entirely. If you handle regulated data, check the compliance position before you start rather than after.
Want Help Laying the Foundation?
We are happy to help you set up the folder structure, the rules and the gates, billed at our normal hourly rate. What we will not do is promise you it will behave perfectly afterwards, because nobody can. We can build you a solid foundation and show you how to extend it. How closely AI sticks to it on any given day is not something anyone can guarantee.
Terms Explained
Not a tech person? No problem. Here’s what these concepts actually mean in plain English.
