Why Are You Teaching AI the Same Job Every Week?

If you have explained the same task to AI three times, you probably don't need a better prompt. You need to save the way you want the work done. That sounds obvious once you see it, but it took me some time to change the way I thought about working with AI.

Like many people, I started by trying to improve the prompt: add more context, explain the tone, give it a role, tell it what not to do, add an example, ask it to check the answer. The results improved. Then I would come back a few days later and do much of it again. That is the part I think we need to rethink.

A good prompt can improve one answer. If the work happens regularly, the bigger opportunity is to make the way the work is done reusable. That is what I mean by an AI skill.

What is a reusable AI skill?

I am using the word "skill" in a simple, tool-independent way. It is a reusable set of instructions that teaches AI how you want a particular type of work done. It can live in a Markdown file, an AI workspace, an agent, a project — whatever system you use. The technology is less important than the idea.

Imagine you regularly ask AI to research a potential customer. Every time, you may explain: look at the company's current business, find recent developments, identify likely decision-makers, do not rely on old information, separate facts from assumptions, look for evidence of the problem we solve, do not make up revenue numbers, give me the sources.

After doing this several times, why should those instructions still live inside individual prompts? They are not really part of today's research request — they are part of how you do research. That difference matters.

Separate what changes from what stays the same

This is probably the simplest way to decide what belongs in a skill. Every repeated AI task contains two kinds of information: the part that changes — the customer, document, meeting, product, market or problem you are working on today — and the part that stays the same — your standards, process, rules, preferred output and checks.

Most of us put both into the prompt. I think we should separate them: keep the changing information in the request, and put the repeated information into the skill. Today's request might simply become "Research Company X before my meeting tomorrow. Use my customer research skill." The skill already knows what good research means to you. That is a much cleaner way to work.

What should go into a skill?

I have found six things particularly useful:

• Purpose — what job is this skill supposed to do?
• Context — what does the AI need to understand about you, the business or the situation?
• Rules — what must it always do, and what should it avoid?
• Process — what steps should it follow?
• Output — how should the answer be structured?
• Quality check — what should it verify before giving you the result?

You do not need pages of instructions. In fact, a narrow skill is often easier to improve than a huge one trying to handle everything.

Let's build one

Suppose you frequently send follow-up emails after customer meetings. Instead of starting from scratch each time, create a simple file — customer-follow-up-skill.md — and put something like this inside:

PURPOSE
Prepare follow-up emails after customer meetings.

CONTEXT
I work with business and technology leaders.
The email should confirm what matters and make the next step clear.

RULES
Use simple English.
Keep the email concise.
Do not use exaggerated sales language.
Never invent commitments, dates or deadlines.
Do not use generic phrases such as "just checking in."
If important information is missing, say so.

PROCESS
Read the meeting notes. Identify: the customer's main requirement,
important concerns, commitments from either side, unresolved
questions, and the logical next action. Then draft the email.

OUTPUT
Subject
Email
Next action

QUALITY CHECK
Is everything based on the meeting notes? Is anything invented?
Is the next action clear? Does the email sound natural?
Can anything unnecessary be removed?

Now give the AI your latest meeting notes and tell it to use the skill. The meeting changes. Your way of handling the meeting does not.

The first version will probably not be very good

That is fine. The real value starts when you use the skill. Suppose the AI writes: "As agreed, we will send the proposal by Friday." But nobody agreed on Friday. You correct the email. Normally, that correction disappears inside the conversation. With a reusable skill, you can improve the system instead: "Never create a deadline unless it appears clearly in the meeting notes." Now the next task benefits from today's mistake.

A few days later, perhaps the AI writes an email that is too pushy. Add another instruction: "Do not ask for another meeting unless there is a clear reason for one." Then perhaps it misses an important concern the customer raised. Improve the process: "Identify unresolved concerns before drafting the email." This is where reusable skills become much more interesting than saved prompts — you are capturing what you learn.

Your corrections become an asset

Think about how much judgement disappears every day inside AI conversations. You tell AI: "That's too formal." "Don't make that assumption." "Check the source first." "That isn't how we qualify an opportunity." "Don't recommend something until you have compared the alternatives." "Keep the management summary to one page."

Those corrections contain knowledge. But most of them are lost when the conversation ends. A reusable skill gives you somewhere to put them. After twenty uses, the skill may contain something valuable — a small piece of your working judgement. After a hundred uses, it may become quite difficult for someone else to reproduce simply by asking AI to "do a good job."

This matters even more when you use several AI tools

We are already moving towards a world where people use different AI systems for different jobs — one for research, another for coding, another for writing, another for data or testing. The best model will keep changing. If your way of working exists only inside the history of one AI tool, changing tools means rebuilding part of it. That is unnecessary. Your method should belong to you. The AI should be able to change without forcing you to teach the work from the beginning. Reusable skills are one way to move in that direction.

Now imagine this inside a company

One skill may not look revolutionary. Twenty good skills start to look different. A sales team could have a prospect research skill, an opportunity review skill and a proposal review skill. An operations team could have skills for analysing exceptions, reviewing weekly performance and preparing management updates. A product team could capture how requirements should be reviewed. A leadership team could capture the questions that should be asked before a major investment decision.

The point is not to automate every decision — some decisions should remain firmly with people. The opportunity is to stop repeatedly explaining the parts of the work that are already understood. Over time, a company can start converting scattered working knowledge into something both people and AI can reuse. That is much more valuable than collecting a folder full of clever prompts.

How I would start

Do not begin by trying to build an AI operating system for your company. Look at your own work from the last seven days. Find one task you did more than once, and ask: what did I have to explain to AI each time? Write those repeated instructions down. Add the steps you normally follow. Add the mistakes you want it to avoid. Add a final check. Use it three times. Each time you correct the AI, decide whether that correction should improve the skill. That is enough to start.

There is also a limit

Not everything should become a skill. A badly designed process does not improve because AI can repeat it faster. A skill also should not quietly make important decisions that require human judgement. And a forty-page instruction file is not automatically better than a two-page one. Keep the skill focused. Let AI handle the repeatable parts. Keep people responsible for judgement, exceptions and accountability.

The interesting part comes later

Today, most of the attention is still on models and prompts — which model is better, which prompt gets the best answer. Those questions matter, but capable AI models are becoming widely available. Many of us will eventually have access to the same underlying intelligence. What will differ is what we build around it: our context, our working methods, our accumulated corrections, our examples, our quality standards, our reusable skills.

A better prompt may improve what AI gives you today. A good skill can improve how you work with AI next month as well. So the next time you find yourself writing the same long prompt again, don't immediately try to make it better. Ask a different question: have I already explained this job before? If the answer is yes, perhaps it is time to stop rewriting the prompt. Build the skill.