Manage AI Like an Intern From a Top College

The most useful way I know to think about an AI model at work is as an intern from a top college. Smart, fast, eager, and well read. Also brand new to your business, with no idea that your Tuesday crew starts at seven, that you never discount the first visit, or that a certain longtime client hates being called by his first name. Manage it the way you'd manage that intern and most of the frustration goes away.

The frustration usually sounds like this: "I told it, and it did the thing anyway." That complaint is a management problem wearing a technology costume, and the fix is a habit you already have.

What the intern is good at, and what it can't know

A good intern can draft a letter, summarize a stack of notes, rewrite a messy paragraph, and do it all before you finish your coffee. The ability is real. What the intern lacks is context, the hundred small facts about your place that never got written down because you never needed to write them down.

AI models have the same shape. They've read more than any person alive and know nothing about your shop. So when a dental office asks one to draft a recall reminder, it writes a perfectly fine generic reminder. It can't know the practice stopped mentioning the cleaning price two years ago, or that the office closes early on Fridays in summer. How the model actually produces an answer explains why. It predicts from what it has seen, and it hasn't seen your office.

Every new hire starts there, and the gap closes the same way for both.

Watch, correct, and carry the lesson forward

You wouldn't hand an intern the client list on day one and leave for the week. You'd watch the first few pieces of work, point out what's wrong, and check that the correction stuck.

Do the same here, in three steps. Read the first several outputs closely, even the ones that look fine. Say what's wrong in plain words. Then carry the lesson somewhere it will survive.

That third step is where most people stop, and it's why the same mistake keeps coming back. A correction typed into a chat lives in that conversation and nowhere else. Open a new chat and the intern has amnesia. The rule I'd hold to is simple. The second time you fix the same thing, the lesson belongs in a file.

A file here means a short plain-text document the model reads every time it starts work, usually loaded into a project or a folder the tool can see. The post on ChatGPT projects and folders covers where those files live. What goes in them is ordinary business knowledge. A lawn crew's file might say that quotes always list mulch by the yard, never by the bag. A law office's might say that retainer letters go out on letterhead and never mention a fee range by email. One line each, written once.

You can be blunter with a model than with a person

With a human intern you soften feedback, because people have feelings and you want them to come back tomorrow. A model has nothing to bruise. You can say, "That's wrong. We never call it a sale, we call it a season close-out," and move on.

Bluntness gets better when you make the model do some of the work. After you correct it, ask it to say what it would change and why. "You keep putting the price in the first sentence. What rule would stop that?" Its answer is often a cleaner version of the rule than the one you'd have written. Read it, fix anything it got wrong, and paste it into the file.

That small move turns every irritation into a permanent improvement. The first week of using AI on real work feels slower than doing it yourself. By the third week the file carries a dozen rules, and the drafts start arriving close to right.

Lean context is better context

Here's the part that surprises people. Giving the intern more material does not always help.

Hand a new hire three versions of the price list and they'll ask which one is current. A model won't ask. It'll blend them, or pick one, and you won't know which until a customer quotes a price you retired in March. Old drafts, abandoned templates, and duplicate files all compete for attention with the one document you meant.

So keep the folder clean. When a price list changes, replace the file rather than adding a new one beside it. Delete the drafts once the final exists. A retail shop with one current return policy file gets better answers than a shop with five policies from five years, even though the second shop gave the model "more."

The same discipline helps people. A new employee given one clear page learns faster than one handed a binder.

Why this works where better prompts don't

A lot of advice about AI is about writing the perfect instruction. Instructions matter, and defining the problem before you hand it over is the single best habit I know. But you can't write the perfect instruction on day one, because you don't yet know what the model will get wrong.

The intern approach accepts that. It assumes mistakes, catches them, and stores the fix. This site works the same way. Its rule files, read by the people and the AI models who work on it, carry dated entries, and nearly every one was written right after something went wrong. Nobody sat down and imagined those rules in advance. Each one is a mistake that only had to happen once.

Some tells won't yield to a written rule at all, because they're habits of sentence shape the model falls into by default. Why AI writing still sounds like AI covers that case, where the answer is a checking step that reads the work after it's written.

Where the files end up

A few months in, the folder of rules is something you didn't set out to build, a written description of how your business does things. It covers how you price, how you talk to customers, and what you never do. That document helps the model. It also helps the next person you hire, and anyone who has to cover for you, which is the quiet overlap with continuity work.

How far to take it is your call. Some owners stop at one file of tone and pricing rules. Others end up with the model drafting most of their routine correspondence and spend the recovered hours with customers. The levels of AI use lay out that range. Either way, the intern is only as good as the manager who writes things down, and I haven't found a shortcut around that part.