"Teaching AI, not learning AI" I have said this many times. I have never explained how to teach. This article breaks it down into four actionable steps. First, clearly distinguish between the data repository, knowledge base, and rule base. Identify which one you are missing. Second, use the 3X4 data organization method to turn your documents into a system that AI can understand. Third, learn how to write a truly useful rule. Finally, directly transfer your original method of leading people. The latter part of the article includes a message for bosses who rarely use digital tools. This is the most effective version I have given in person.
- People who feel they "haven't learned AI yet" and are hesitant to start
- People who have used AI for a while but have to start from scratch every time when explaining requirements
- People who have many documents but AI can't find the key points, always answering irrelevantly
- Managers and bosses who naturally know how to lead teams and conduct training, but feel they have no connection with technology
- The distinction between data repository, knowledge base, and rule base, and knowing which one you are missing
- The 3X4 data organization method: a complete explanation of three diary types × four time horizons
- What a good rule looks like, plus four common ways of teaching incorrectly
- A "new employee onboarding guide" that you can directly say to AI
Reverse the order
I have never spent a single day learning AI. When I use AI, I always think about "how to teach AI". I believe it should be more obedient and smarter, and my thinking pattern has always been like this.
This is not a word game. The two approaches will lead to completely different results.
Focusing on learning, you will always be chasing tools: what this model is better than the one from last month, how to open that new feature, and what new terminology has emerged. Tools change every month, and you are always catching up.
Focusing on teaching, you accumulate something else: your judgment criteria, your workflow, and your values. These things won't become obsolete just because you switch to a different model. In fact, when you switch to a stronger model, they will perform even better.
There is also a practical reason. The current models are already very strong. They lack information about you. They don't know who your clients are. They don't know why you rejected a case last time. They don't know what you always check before submitting a file. These things even the strongest model cannot guess. Only you can teach.
So, what exactly should we teach it?
Speaking of this, the next question naturally arises: if we are to teach, then what exactly should we teach?
I will not just talk about the word 'knowledge' anymore, because people tend to confuse it easily. I will divide it into three categories:
More structured things, data, tables, lists, and quotation details.
General documents, meeting minutes, presentations, SOPs, teaching materials, and notes.
How I make rules and judgments. What to choose in what situations, and what to absolutely not do.
Most people stop at the second level. The knowledge base contains a lot of documents. The documents may not contain judgments. Some people write their value judgments and rules in the documents. Most people do not. I give it another name: rule base.
Without the third level, AI's performance will stop at 'being able to find data but not being able to select the right answer.' It can read your three quotation sheets to you. It doesn't know why you chose the more expensive one last time.
That 'why' is exactly tacit knowledge: the feeling of a master teacher, the judgment of a business person, and the unspoken line in a manager's mind. The core work of teaching AI is actually to write these things down.
Where to put things, it can find them
After knowing what to teach, the next question is where to put things and how to put them, so that AI can understand.
The method I use myself is called 3X4. In one sentence, it is: Three types of diary multiplied by four types of timing.
Three types of diary: solving the three things AI does not know
Solving 'AI does not know what you are doing'. What you did today, where you got stuck, and how you solved it.
Solving 'AI does not understand how you think'. Your judgment and reasoning about an issue, this type is closest to the rule base.
Solving 'AI does not know your state'. How your state is this week, what things consume you.
The priority of the three types is different. At the beginning, just write the work diary. The other two types can be left for now. Writing more is more important than writing everything.
Four types of timing: put the most used ones closer
The four types of timing solve 'where AI should look first'. The logic is similar to how the human brain remembers:
- The past few days Put on the workbench or on the board: tasks to be done this week, things in progress
- The past one month Root directory: New diary entries, drafts, and ongoing files. This is the first place AI searches.
- Within six months Place in categorized folders: Content that has come to an end, organized by purpose.
- Over six months Move it to the archive area: move it, do not delete it. Keep the name so it can be found when needed.
Two key habits: New files are initially placed in the root directory, prioritizing easy access. After a month, they are organized. Do not place files in deep folders immediately for aesthetic classification. Move expired files away, never delete them. The archive area is your history.
There is also a small action that provides significant benefits: Each file name includes a timestamp at the beginning. The format is `YYYY-MM-DD-HHMM Title`. This way, files are naturally sorted by time. AI can immediately tell the new and old based on the file name, without opening the content. The saved cost is real.
This is why I say AI doesn't matter which company you use. As long as the files are there, the file names are clear, and the links are complete, the system is there.
What a good rule looks like
After placing the files, the most critical part is how to write the rules. The rule base is the least done but highest value layer among the three layers, so this section will be more specific.
A usable rule usually has three parts: What situation, how to do it, why. The third part is often omitted, but it is the key that allows AI to think beyond what is written. Writing only the first two parts makes AI copy directly; writing all three parts allows it to understand how to handle situations not written.
Reply to the customer in a professional manner.
There is no context, no action, no reason. AI can only guess what professional means, and the guess is often not what you want.
When the customer asks about price, first ask for the usage context before quoting. Because the value of the same service varies greatly for different contexts, quoting a number first turns the conversation into a price negotiation.
Context, action, and reason are all present, and AI can also adapt to a different context.
Other four common ways of teaching incorrectly, and if you fall into any of them, you will feel AI is dumb:
- Only provide examples without giving reasons. Throw three past documents and say 'write like this', and AI will copy the surface format, but not your judgment.
- Repeating the requirements every time. Telling the requirements once and then again next time, which is equivalent to being a permanent customer service, and the rule base never grows.
- Writing too abstractly. Words like 'having warmth' and 'being precise' are hard for you to define, and AI is even harder. Change them into observable actions.
- Write what should be done, but do not write what cannot be done. Red lines are more important than positive rules, because they block the mistakes you cannot accept.
The most effective habit is only one: When AI makes a mistake, do not repeat the requirements once, instead, add that rule into the rule base. This is the only action that will make you get more comfortable with using it. If you repeat the requirements every time, you are always acting as a customer service; if you write it as a rule, you are building a system.
There is a type of person who learns this particularly fast
The above are all methods. But in one-on-one sessions, I have seen something that determines whether a person can get the hang of it more than methods.
To create your own AI employee, you actually need two kinds of abilities.
Knowing what it can do and cannot do, and knowing how to assign tasks to it. Everyone is currently improving in this area.
Knowing how to break down work, provide standards, conduct reviews, and clearly communicate when someone makes a mistake. Many people already have these skills, but few have considered applying them here.
Currently, everyone is mostly improving in the first area, but few pay attention to the second area, which is the real threshold.
Because of this, I often encounter people with an advantage in one-on-one sessions: those who already have experience in organizational leadership, management, or training. Once they understand how to communicate with AI, they progress very quickly, as the logic of managing people can be directly applied.
Conversely, if you have always delegated tasks by saying "just do it," you will also get stuck with AI, because AI also needs to know what constitutes good work.
For bosses who are not very familiar with digital tools
I have met some bosses in their fifties who ask: I'm not very good at using digital tools, and I don't understand AI. Can I really create my own AI employee?
My usual response is: You can. You're really good at teaching people.
Then I ask them to imagine AI as a new graduate assistant. More precisely, a university graduate who is very smart, understands everything you say, but has no practical experience, doesn't know the company's rules, and doesn't know what you value.
He has one advantage that a general new employee doesn't: if you teach him seriously and patiently once, he will remember. He won't forget, won't take leave, and won't think you're nagging because you mentioned it last week. As long as you teach him well once, he can perform well.
Usually, by this point, the boss knows how to start, and the results are usually good. Because he doesn't need to learn new things; he's just doing what he's done for twenty to thirty years: training a new employee.
A directly copyable new employee onboarding guide
After completing the following section, paste it to your AI, and it will transform from a "general assistant" into "your employee." This is the first page of the rule base.
If you only do one thing, start here
After reading the whole article, if you only do one thing, I suggest this order:
Pick the decision you make most often, and write down how you judge it, remember to include the reason. This is the first rule in the rule base.
New files should be named `YYYY-MM-DD-HHMM Title`, and all files should be placed in the root directory. This step takes almost no time, but the cost for AI to find things drops immediately.
When AI makes a mistake, change to adding a rule instead of restating the requirements. This step is the dividing line between whether the entire method will take off.
You don't need to wait until you learn AI to start. You can start teaching it today.
Conclusion
Learning AI is chasing tools, teaching AI is writing down your own judgments. The former's results will become obsolete with version updates, while the latter's results will become more valuable as the model gets stronger.
Teaching AI has a side effect: you will be forced to clarify the judgments you've been making based on intuition. Many people find out halfway through that the biggest gain isn't that AI becomes more useful, but that they finally understand what they're doing.
Your documents are your system.
Further Reading
- Solving AI Knowledge Anxiety: Finding Anchors That Remain Unchanged in a Rapidly Changing Era What to write in the rule base depends on what task you want to complete.
- Why Your AI Adoption Isn't Working: From the Three-Layer Theory to the Organization's Wall Writing down your judgments is exactly the third-level entry ticket mentioned in that article.
- Information Lag and Market Forecasting: How to Be Ahead of the Market Without Being Too Early Tools will always change, but the rule base is the only thing you can take with you when switching tools.
- Common AI questions 37 This article's original source, a collection of one-on-one received questions.