Why 'Make it good' Doesn't Work in Prompts | Converting to Specific Instructions in Requirement Definition

Hello, I am Pokego, and I research prompts for writing high-quality articles with generative AI.

If you have been using AI to write text, you have likely experienced this:

"I told it to 'write it in a good way,' but the text that came back was just... off..." "No matter how many times I try, it doesn't turn out the way I imagined..."

You have probably felt this kind of frustration before.

Actually, this problem is not due to the AI's performance.

Adjectives like "in a good way" or "easy to understand" are ambiguous words that are interpreted differently by each person.

Because the AI interprets and fills in that ambiguity in its own way, the output you get ends up deviating from your image.

In other words, this problem can be solved if you break down intuitive words into specific conditions and convey them.

This is a concept known in the engineering world as "requirement definition," but no specialized knowledge is required.

In this article, for those who feel that their instructions to AI are not getting through, I will cover:

• The structural reasons why "make it good" doesn't get through to AI

• Steps for "prompt requirement definition" to turn adjectives into specific instructions

• Self-check methods to eliminate ambiguous expressions before sending

I discuss the above based on the insights of the author, who has repeatedly researched and practiced giving instructions to AI.

Just by replacing words you used to convey intuitively with conditions, AI output will become stable.

Please feel free to use this as a reference.

How AI interprets 'make it look good' in prompts

'Please write it in a good way'

Have you ever sent an instruction like this to an AI?

I understand that you have written many other things and want the AI to read them and finish it in a good way.

However, in reality, the meaning the AI receives from that may be quite different from the meaning you intended.

Here, let's look at the true nature of that structural gap.

The meaning of 'good' is not aligned between you and the AI

When you instruct it to 'write it in a good way,' the AI will output some kind of text.

But have you ever had the experience of reading the returned text and feeling, 'Hmm, it's just not quite right...'?

The cause is very simple.

There are as many correct answers for 'good' as there are peopleis why.

For you, 'good' might mean a 'friendly, casual tone'.

On the other hand, someone else might be imagining a 'logical and persuasive writing style'.

AI cannot clearly define which one is the 'best' from among this ambiguity.

Therefore, it generates the text that it thinks is the most 'plausible'.

In other words, the 'good' that the AI returns is not your personal 'good', but an output that averages out various 'goods'.

The reason you get text that differs from your expectations may actually be caused by this 'mismatch in meaning', not by the AI itself.

Output inconsistency happens because you haven't written specifically about 'what is good'.

Another issue many people struggle with is the phenomenon where the output changes every time.

Even if you generate text several times with the same prompt, you get slightly different results each time.

Have you ever experienced this?

This happens because the AI is filling in the 'blanks' in your instructions in a different way each time.

The instruction 'make it good' does not contain the criteria for 'what is good'.

The AI tries to fill in those blanks with its own guesses, but since there is more than one pattern for those guesses, the results fluctuate every time.

This structure is exactly the same when working with humans, isn't it?

For example, when I commission a manuscript from a freelance writer for media operations.

If I only say, 'Please write in easy-to-understand language,' each writer will submit a completely different manuscript.

Or rather, they would probably get angry and say, 'Please give me more specific instructions!'...😅

However, when giving instructions,

'Use vocabulary that an average office worker can understand at first glance,'
'Keep each sentence under 40 characters'

If you communicate conditions like these specifically, the quality of the submitted manuscripts will stabilize.

In other words, even with instructions for AI (prompts), if you write the conditions specifically just like you would when asking a person, the output will stabilize.

The quality of a prompt is determined by whether or not you can write the conditions specifically .

'Prompt requirements definition' that turns adjectives into specific instructions

I myself started to see the output clearly stabilize once I became conscious of breaking down adjectives into specific conditions before handing them over.

This way of thinking, 'breaking down intuitive words into conditions,' is called 'requirements definition' in the IT engineering industry.

However, this does not mean doing a full-scale system development requirements definition.

It is just applying the essence of that to prompt creation.

Here, let's look at the procedure divided into two steps.

Find adjectives and ask yourself, 'What does that mean specifically?'

The first step in conversion is to consciously identify the adjectives in your prompts.

You might think, 'Isn't that obvious?', but this is actually the biggest blind spot.

This is because adjectives slip into prompts unconsciously.

'Make it look good',
'Make it easy to understand',
'Make it feel natural'

These words work fine in daily work and casual conversation, don't they?

That is why we tend to use them unconsciously in prompts as well.

Once you finish writing your instructions, try reading them over once.

Then, look for adjectives like 'good', 'easy to understand', or 'natural'.

When you find an adjective, ask yourself, 'What does that mean specifically?'.

This self-questioning is the starting point for organizing your conditions in the next step.

Organizing conditions using the four elements: 'Purpose, Material, Constraints, and Format'

Even after asking 'What does that mean specifically?', you might not find an answer right away.

In such cases, the four elements of 'Purpose, Material, Constraints, and Format' introduced by the author in the following article are helpful.

Reference article:

For example, let's break down the adjective 'write in a warm tone'.

• Objective: Create text that allows the reader to proceed with peace of mind

• Content: Include two personal anecdotes from the author

• Constraints: Do not use technical jargon, use more hiragana than kanji

• Format: Use polite 'desu/masu' style, include one question for the reader in each paragraph

By organizing it this way, the subjective word 'warmth' changes into specific conditions that AI can process.

The key point is that you don't need to fill in all four elements.

Even filling in just one or two elements will significantly improve the stability of the output compared to instructions using only adjectives.

Pre-send check to eliminate ambiguous expressions in prompts

Make it a habit to perform a 'zero-adjective check' as a self-check before sending a prompt.

If there are any adjectives left in your prompt, that is a sign that 'there is still room for further specification'.

Here, I will share some ready-to-use replacement examples and tips.

Examples of commonly used adjectives and how to replace them with specific instructions

There are standard replacement patterns for adjectives frequently used in prompts.

Let's look at a few examples.

• 'Easy to understand' → 'Use vocabulary that a middle school student can understand'

• 'Simple' → 'Limit to 2 sentences per paragraph and organize into 3 bullet points'

• 'Natural feel' → 'Use polite 'desu/masu' style, do not omit conjunctions, and keep each sentence under 40 characters'

• 'Polite' → 'Include at least one piece of evidence, avoid definitive statements, and use phrases like 'it is considered that...''

The common trick is to 'turn the adjective into a condition that a third party can verify as being met'.

Whether something is 'easy to understand' is subjective, but if you specify 'vocabulary a middle school student can understand,' anyone can judge the level of clarity.

This 'verifiability' is the boundary between an adjective and a specific instruction.

Start by picking one adjective you use often and try replacing it.

Tips for replacing elements that are hard to quantify with 'observable conditions'

'Warmth'
'Professionalism'
'Emotional feel'

Many people feel that these adjectives are difficult to turn into numbers.

But in reality, you don't need to force them into numbers.

The important thing is to convert them into 'observable conditions'.

It is the same concept as breaking down 'warmth' into four elements earlier.

Visualize specifically what a "sentence that meets that adjective" looks like, and articulate exactly what makes it different.

This is the foundation of the conversion process.

For example, try to visualize a "professional-sounding sentence."

"It contains data and numbers,"
"It has no vague expressions,"
"It states the conclusion first."

Don't these characteristics come to mind?

These become the conditions for your prompt as they are.

Instead of instructing it to "write professionally,"

By writing it this way, the AI can now judge the substance of what makes something professional.

You don't need to write perfect conditions.

Even just removing one adjective and adding one condition in its place will change the AI's output.

A step from "communicating by feeling" to "communicating by conditions".

Please try this out in your next prompt.

Summary: The quality of a prompt is determined by your "ability to communicate through conditions."

In this article, for those struggling with the ambiguity of their prompts, I have discussed:

• The mechanism behind why adjectives are not understood by AI and their structural causes

• Prompt requirement definition: The process of converting adjectives into specific instructions

• Examples of replacements you can use before sending and tips for converting to 'observable conditions'

I have covered the above while incorporating insights from my own research and practice in giving instructions to AI.

What I wanted to convey in this article is to rewrite adjectives written vaguely into conditions that mean the same thing to anyone who reads them.

Whether it's prompts, requests to people, or reports, simply replacing ambiguous expressions with specific conditions reduces discrepancies in interpretation.

The same change I experienced when giving instructions to writers should happen with your prompts as well.

From today's prompt, try to be conscious of finding adjectives and rewriting them into conditions!

✅ Click here for Poke-go's content