How does prompt work? Understand these 3 underlying principles and you can become a prompt word expert

Recently, a friend asked me: "Why do the prompts you write always allow the AI ​​to give accurate answers, while mine often go astray?"

To be honest, I have been troubled by this problem for a long time.

It was not until I delved into the working principle of the big model that I suddenly realized that the essence of Prompt is to guide the model output by reshaping the probability distribution , constraining the generation path , and activating specific knowledge modules .

Sounds mysterious?

Don't be impatient, let's start with a simple example.

Understanding the Big Model from the "Guessing Word Game"

Have you ever played the game of solitaire?

When one person says "spring", the next person may say "flowers are blooming" or "it's coming".

The big model actually works a lot like this, except it's playing a super complicated "guess the next word" game.

Imagine that AI has read almost all the text on the Internet (GPT-3's training data is 45TB, and after filtering, it still has 570GB), and it has learned various language rules from it. For example, when it sees "today's weather", it knows that it is likely to be followed by words such as "really good", "not bad" or "very bad".

But there is a key point here - AI does not memorize by rote, but "guesses" through probability. Just like when you see "the moon in front of the bed", "light" automatically pops up in your mind, this is because you have seen this combination too many times and the probability has been deeply imprinted in your mind.

The AI ​​uses the Transformer architecture during training (the same technology that makes ChatGPT so smart). Its most powerful feature is that it can pay attention to the context of the entire sentence at the same time, rather than just looking at the first few words .

It's like when you understand a sentence, you consider the entire context instead of just focusing on a single word.

How Prompt "controls" the AI ​​brain

OK, now we know that AI generates text through probability. So how does Prompt affect this process?

Reshaping probability distribution - putting "tinted glasses" on AI

When you input "write a sentence to describe cats", AI may output a bland sentence like "cats are lovely pets". But if you change it to "write a poetic sentence to describe cats", something magical happens - AI will output a sentence like "elves walking quietly in the night".

Why is this happening?

Because the word "poetic" changes the probability distribution of AI.

Originally, the probabilities of words like "pet" and "animal" were very high, but after adding "poetic", the probabilities of words like "elf", "moonlight", and "elegant" suddenly soared.

It's like putting a "poetic filter" on AI.

I did an experiment to generate product descriptions:

• Normal prompt: "Describe this headset" → "This is a wireless Bluetooth headset with clear sound quality"
• Optimize Prompt: "As an audio enthusiast, describe the sound quality experience of this headset" → "The low frequency dive is deep, the mid-frequency human voice is warm, and the high frequency extension is natural, as if you are in a concert hall"

Do you see the difference?

The character setting directly changes the AI's "way of thinking".

Constraint generation path - drawing a map for AI

Sometimes, changing the probability is not enough, we need to plan a specific route for the AI.

It's like a navigation software. Not only do you have to tell it the destination, but you also have to choose whether to take the highway or the national road.

For example, I asked AI to solve a math problem before and asked it directly for the answer, and it often got it wrong.

Later I changed it to: "Please solve this math problem step by step and explain the reasons for each step", and the accuracy rate immediately improved.

Why?

Because the "step by step" instruction forces the AI ​​to activate reasoning modules in sequence, just like forcing it to walk up the stairs instead of jumping off a building.

Stop and think before taking each step, so as to avoid errors caused by logical leaps.

Activating specific knowledge modules - awakening AI's "professional personality"

This part is very interesting.

The hundreds of billions of parameters of AI actually constitute a distributed knowledge network, with different knowledge stored in different "areas".

When you say "as an experienced Python engineer", AI will preferentially activate parameter clusters related to Python.

At this time, irrelevant content such as medical knowledge and legal provisions are "turned off" and will not interfere with the output.

test:

• Q: "How do you process data?" → The answer is very broad, it may involve Excel, database, etc.
• Q: "As a Pandas expert, how do you process data?" → Directly give DataFrame operation code

The effect of this targeted activation is very obvious.

How to write a "magic" prompt

Once you understand the principle, you can follow a rule to write a good prompt. Here are some practical tips I have summarized:

Tip 1: Reduce information entropy and don’t make AI “difficult to choose”

What is information entropy? Simply put, it is uncertainty. The more vague your prompt is, the more choices the AI ​​has, and the more uncontrollable the output is.

Compare this to:

• ❌ "Write a sorting algorithm" → AI doesn't know whether you want bubble sort or quick sort
• ✅ "Use Python to implement quick sort, process floating point lists, and add time complexity annotations" → AI accurate output

My experience is that it is better to make all the details clear rather than writing a broad requirement, which reduces the information entropy and increases the certainty of the output.

Tip 2: Structured thinking, give AI a clear framework

I particularly like to use XML tags to organize prompts, such as:

<Role> Senior Content Operations </Role>​​
<Task> Optimize this text </Task>​​
<Requirements>​​
Increase emotional resonance
Keep the original intention
The word count should be within 100 words
</requirements>​​
< original text > ... </ original text >

This structured approach is like giving AI a form, and it knows what to fill in where. In actual tests, the error rate can be reduced by more than 30%.

Tip 3: Use examples to guide AI

Sometimes, it’s better to give an example than to explain it. AI is particularly good at pattern recognition, and it will know what you want after seeing the example.

For example, I want to generate JSON data:

< query >
Please generate a weather report in JSON format, including the following fields: city, temperature, humidity, wind speed
Example: {"city": "Beijing", "temp": 22, "humidity": 45%, "wind": "3m/s"}
</ query >

With this example, the AI ​​will not make mistakes in field names or add other attributes randomly.

Tip 4: Negative constraints, telling the AI ​​"not to do something"

This is a trick that many people overlook, but it works really well. Sometimes telling an AI what not to do is more effective than telling it what to do.

for example:

• "Explain quantum mechanics without the jargon"
• "When writing a product description, don't use exaggerated marketing words"

Negative constraints can directly suppress the probability of certain words, making the output more in line with your expectations.

How to improve your prompt ability in daily life

Having mentioned so many techniques, how should we practice them in our daily life?

Build your own prompt library

I built a Prompt library using Notion, which is classified by scenarios: technical development, content creation, data analysis, etc.

Every time you write a prompt that works well, save it and record why it works.

Over time, you will discover some patterns.

For example, technical prompts must specify the version number (React 18 instead of React), and content prompts must specify the style (humorous, serious, academic, etc.).

A/B Testing Thinking

For the same requirement, try different prompt writing methods and compare the results. I often do this:

• Version A: Directly describe the requirements
• Version B: Add character settings
• Version C: Add output format requirements
• Version D: plus examples

By comparing them, you can clearly see the role of each element. This spirit of experimentation is important.

Learn some technical knowledge (but not required)

Knowing some programming languages ​​helps when writing good prompts.

There are two precise and unambiguous languages ​​in the world, one is mathematical language and the other is programming language.

If you understand the JSON format, you can accurately ask AI to output structured data.

If you understand the characteristics of different programming languages, you can give more accurate instructions when letting AI write code.

However, this only applies to scenarios that require precise requirements.

In scenarios such as writing poetry and copywriting, technical knowledge is not that important, and language expression ability is the key.

Last words

Let’s go back to the friend’s question at the beginning - why are some prompts better than others?

Now the answer is clear: a good Prompt is effective because it precisely controls the probability distribution of AI , plans a clear generation path , and activates the correct knowledge modules .

To put it simply, the more thorough your understanding of the problem, the more accurate the prompt you write, and the more the AI's output will be in line with expectations.

A good prompt is essentially an accurate translator between humans and machines, converting vague human intentions into executable instructions for AI.

If you understand how it works, you will have mastered the code for efficient communication with AI. In a sense, Prompt engineers are like "translators" in the AI ​​era, translating human needs into a language that AI can understand.

Finally, I will give you a formula:

A versatile and structured prompt framework:

# RTF (Role-Task-Format)
-  Role
-  Task
-  Format

## Example
-  Role: You are an experienced marketing expert  
-  Task: Write a compelling social media post for a new eco-friendly smart water bottle.  
-  Format: A short text with emojis, less than 100 words, highlighting the product's eco-friendly features and smart functions, and encouraging users to learn more.