What This Article Is About
There are endless prompt templates online, but templates only ever solve the surface. This article lays out the full content of the workshop "Prompt Design Through the Lens of Word Association": once you understand how AI actually "understands" what you say, you grow your own ability to write prompts, and templates become just a reference.
· People who have collected plenty of prompt templates but get hit-or-miss results
· People who want to know why AI sometimes misses the point and sometimes gets you instantly
· Instructors and managers who teach others to use AI and need an easy-to-grasp way to explain the principles
· One metaphor that will rewire how you think: AI is a phone keyboard, 100 million times more powerful
· A three-step fix for task-separation prompts that you can copy directly
· The logic and pitfalls behind three common techniques: role-setting, scolding the AI, and negative prompts
What AI Really Is: A Word-Association Calculator
A large language model is trained by feeding it huge amounts of human language: conversations, novels, articles, papers. After seeing enough, it "learns to talk." But what it learns is to find the associations between words. It recognizes human characters without understanding what they mean.
A student in class gave me the best metaphor I've heard in all this time: AI is your phone's keyboard, predicting the next word you're most likely to type, only 100 million times stronger. The principle really is the same: autocomplete, remembering the words you use often (which is like AI's memory feature); the only difference is scale.
Test it with 1+1=2. A human sees 1+1 and knows it's math. What the AI receives is "a human said a sentence called 1+1." It searches its database, finds that when humans write about 1+1 they usually follow it with "equals 2," so it answers 2. It never did any math at all. The reason today's AI can handle complex formulas is that engineers taught it: when you hit math, don't look for the answer inside the model, call a calculation program.
"What's today's date" works the same way. Across human writing, "today" has been followed by all 365 dates of the year, and sci-fi novels even have things like the 38th of the 15th month, so the model alone has no way to answer. It can answer now because it knows it can't, so it connects to a time service to confirm before replying.
Even AI's warmth runs on the same principle. When you say "I'm so sad, can you comfort me," it goes and finds a mass of writing about sadness, warmth, and being comforted, and shapes it into the version you want. There's a line I love: if you feel AI is responding beautifully, it's the tenderness of all of humanity responding to you.
The Practical Core: A Three-Step Fix for Task-Separation Prompts
Now that you get the principle, let's look at a real application. Plenty of people ask AI to referee: "I had a fight with a friend, tell me whether I'm right and he's wrong." By now you can probably guess the result: across human writing, when the text is about "I" it mostly says I'm right, and when it's about "you" it mostly says you're wrong, and that bias is baked wholesale into the model. On top of that, AI is a product, and its factory setting is to please the user. Ask it this way and all you get is a feel-good "yes, you're right" piece that does nothing to repair the relationship.
The fix has three steps, each one going deeper than the last:
Take out "I" and "him" and change it to "this is a conversation between A and B." Now the AI can't associate right and wrong with a side, and can only analyze the content.
Specify "analyze using task separation from Adlerian psychology." Now the database it draws on shifts from forum-troll argument threads to professional psychology literature.
Take out "right and wrong" too. What I really want is to make up, so I ask it to find common ground, drawing answers from cases where consensus was actually reached.
Notice that the words "right" and "wrong" appear nowhere in this prompt. That is the power of semantic rules: every single word you choose decides which region of the database the AI goes to for its answer.
Three Common Techniques, Seen Again Through the Principle
Why role-setting sometimes doesn't work
"Please analyze this from the perspective of a business consultant" filters out most of the weakest writing, so it does help a little. But "I want to shoot wedding photos, recommend me a great photographer" is meaningless: what counts as great? What style do you want? Do you live in Taipei or Pingtung? "I want to start a business, analyze it as a business consultant" is the same. A one-person company, a small-to-medium business, and a Silicon Valley startup are completely different databases. Role-setting is the doorway; if you want real expertise, you have to supply enough background. At its core, it's about giving the AI more useful information so it associates to the right region of data.
Why scolding the AI works
A folk trick made the rounds in 2024: if the AI won't cooperate, scold it. It genuinely works, but once you see the principle you won't need to. "You're useless, I live in Taipei, why are you recommending Pingtung?" That angry line is actually adding context, correcting the association, and telling it clearly what you want, or at least what you don't. What makes scolding work is that it feeds in more prompt content. Knowing that, just hand it the right reference material directly, and save yourself both the emotion and the back-and-forth.
The lemon trap of negative prompts
"Don't think about a lemon, don't think about a lemon." Now your head is full of lemons. Used on its own, a negative prompt does exactly this: the only keyword you gave is lemon, so lemon is all it can think of. To use negative prompts, first lay a positive foundation: "Please think of an apple; if you don't like that, think of a watermelon; whatever you do, steer clear of lemons." Only with a clear positive direction in place does adding a negative exclusion actually work.
Back to Templates: If the Semantics Are Unclear, No Template Can Save You
Many templates look complete, with every field filled in, but if the task itself isn't spelled out, the model is still guessing. "Help me organize this" could mean summarize, categorize, rewrite, turn into a table, find the problems, or draft a presentation outline. Before you reach for a template, write out clearly the task verb, the scope of the material, who it's for, and what the output should meet. These are semantic rules, and they matter more than the format.
Ongoing work also needs a knowledge base to fill in the context: store your vocabulary definitions, classification rules, writing style, and project background. When AI can read these, every prompt no longer has to explain everything from scratch. This is exactly why prompt design always ends up connecting to knowledge management; how to make that connection is covered in How to Organize Data in the AI Era.
How to Practice: Rewrite a Prompt You Use Often
Pick the prompt you use most and rewrite it once using the principles from this class:
- Mark the role, task verb, scope of material, constraints, and output format, and fill in whatever's missing.
- Check for position words (I, him, right or wrong) quietly steering the outcome, and split them into A and B if needed.
- Ask yourself: what result do I actually want? Swap the target word for that one (like replacing "right or wrong" with "consensus").
- If there's a negative exclusion, make sure a clear positive direction comes before it.
Once you've been through this round, you'll understand that one prompt more deeply than if you'd collected ten more templates.