What is this article talking about?
There are endless prompt word templates on the Internet, but the templates only solve superficial problems. This article summarizes the complete content of the "Prompt word design based on word relevance" workshop: first understand how AI "understands" what you say, and you will develop the ability to write prompt words yourself, and the template will become a reference. Who is
· People who have collected many prompt word templates, but the results are sometimes good or bad.
· People who want to know why AI sometimes answers wrong questions and sometimes understands you in a second
· To teach others how to use AI, you need lecturers and supervisors who can explain the principles easily.
· A metaphor to change your cognition: AI is a mobile phone input method that is 100 million times more powerful
· Three-step correction of topic separation prompt words, which can be copied directly
· The principles and pitfalls of three common techniques: character setting, scolding AI, and negative prompt words
The essence of AI: a computer that associates words
The way a large language model is trained is to let it see a large amount of human language: conversations, novels, articles, papers. The more you watch it, the more it "learns to speak". But what it learned was to find the connection between words. It recognized human words, but it did not understand the meaning of the words.
A student in the class gave me the best metaphor I have ever heard for so long: AI is the mobile phone input method. It can predict your next most likely word, but it is only 100 million times better. The principle is really the same: it automatically selects words and remembers your common words (this is like the memory function of AI). The only difference is the scale.
Check once with 1+1=2. Humans see 1+1 and know this is mathematics. What the AI received was "Humans said something called 1+1". It looked through the database and found that when humans said 1+1 in articles, it was usually followed by "equal to 2", so it answered 2. It doesn't do math from beginning to end. Today's AI can calculate complex formulas because engineers teach it: when encountering mathematics, don't look for answers in the model, call the calculation program. The same goes for
"What is today's date?" In human articles, "today" is followed by 365 dates in a year. In science fiction novels, there is even 15 and 38. It is impossible to answer it by relying on models alone. It can answer now because it knows that it cannot answer and will first connect to the time service for confirmation before replying to you.
Even the gentleness of AI is based on the same principle. If you say "I'm so sad, can you comfort me?", it will find a large number of articles about sadness, warmth, and being comforted, and organize them into the version you want. There is a saying that goes well: If you feel that AI responds well, it is the gentleness of humans around the world that is responding to you.
Practical core: three-step correction of topic separation prompt words
Understand the principle and let’s look at a real application. Many people will ask AI to comment: "Please help me see if I am right and he is wrong when I quarrel with my friend." Now you can guess the result: in human articles, when talking about "I", most of them say that I am right, and when talking about "you", they usually say that you are wrong. This inertia is completely embedded in the model. In addition, AI is a product, and the factory setting is to serve users well. If you keep asking like this, you will only get a "yes, you are right" cool article, which will not help repair the relationship.
The correction is divided into three steps, each step is deeper than the last:
Remove "I" and "him" and change it to "This is a conversation between A and B." AI cannot correlate right and wrong from a standpoint and can only analyze from content.
Specify "subject separation analysis using Adlerian psychology". The relevant database was changed from the quarreling articles of forum users to professional psychology literature.
Remove “right and wrong” as well. What I really want is reconciliation, so I ask it to find consensus and let it find answers from the information that has successfully reached consensus.
Please note that there is no word "right or wrong" in this prompt. This is the power of semantic rules: every word you choose determines which database the AI goes to find the answer.
Three common techniques, re-read using the principles
Why character setting sometimes doesn’t work?
"Please analyze from the perspective of a business consultant" will filter out most of the poor articles, so it is somewhat useful. But "I want to take wedding photos, please recommend me a great photographer" is meaningless: What does it mean to be great? What style do you want? Do you live in Taipei or Pingtung? "I want to start a business, please analyze it from the perspective of a business consultant." The same is true. One-person companies, small and medium-sized enterprises, and Silicon Valley startups are completely different databases. Character setting is the entrance. It must be really professional and provide sufficient background information. The essence is to give the AI more useful information and connect it to the right data area.
Why scolding AI is effective
A folk remedy is circulating in 2024: scold the AI if it disobeys. This trick is indeed effective, but if you understand the principle, you don’t need to curse. "You suck, I live in Taipei, why do you recommend Pingtung?" This curse is actually adding context, correcting the relationship, and clearly telling it what you want or at least what you don't want. The effective nature of scolding is that it gives more prompt words. If you know this, just give it the reference material directly, which will save you emotions and time.
The Lemon Trap of Negative Prompt Words
"Don't think about lemons, don't think about lemons." Your mind is full of lemons now. Using negative cue words alone has this effect: among all the keywords you give, there are only lemons, and it can only think of lemons. To use negative prompt words, first lay out the positive foundation: "Please think of apples. If you don't like it, you can think of watermelon. In short, avoid lemons." Only when there is a clear positive direction, and then the negative direction is eliminated, will it be effective.
Return to the template: The semantics are unclear and the template cannot save it.
Many templates seem to have complete fields, but the task itself is not clearly explained, and the model is still guessing. "Help me organize" can be summarizing, classifying, rewriting, converting tables, finding problems, or making a presentation outline. Before setting up the template, clearly write down the task verbs, data scope, usage objects and output standards. These are semantic rules and exceed the format requirements.
Long-term work also requires a knowledge base to supplement the context: save vocabulary definitions, classification rules, writing styles, and project backgrounds. AI can read these, so there is no need to explain the prompt words from scratch every time. This is also the reason why prompt word design will definitely be connected to knowledge management in the end. The connection method isHow to organize data in the AI eraThat one. Correct usage of
How you can practice: Rewrite a common prompt word
Pick one of your most commonly used prompt words and rewrite it using the principles of this lesson:
- Mark the roles, task verbs, data scope, restrictions and output formats, and fill in the missing ones.
- Check whether there are position words (I, him, right or wrong) secretly guiding the direction, and separate them into A and B if necessary.
- Ask yourself: What result do I really want? Replace the target word with that one (like replacing "right or wrong" with "consensus").
- If there is a negative exclusion, confirm that there is a clear positive direction before.
After this round, your understanding of this prompt word will be deeper than if you collected ten more templates.