How to Improve Prompts? Logical Debugging Techniques to Guide AI with 'Single-Point Differences'

Hello, I am Poke-go, a researcher studying prompts for writing high-quality articles with generative AI.

When using AI, you've probably found yourself stuck in a quagmire of revisions, haven't you?

“Why did you forget the condition I just mentioned?”
“I asked you to fix this, but now another part is broken...”

Many of you likely feel this kind of frustration and futility.

In fact, the most important thing in prompt improvement is not to try to fix the output within the chat conversation.

The longer the conversation with AI becomes, the more it loses sight of the important conditions specified at the beginning.

From “going back and forth to fix things” to “returning to the original instructions and fixing only one point.”

Just this shift in perspective can change the accuracy of the output surprisingly.

In this article, for those who are troubled by AI output accuracy and want to know reliable improvement procedures, I will explain:

• The true nature of 'context loss' where AI forgets instructions

• The 'single-point difference test' mindset to prevent revision drift

• 3 concrete steps for prompt improvement that you can use immediately

The above is explained based on the insights of the author, who continues to research prompts for AI article creation.

The essence of prompt improvement is to guide AI with logic, not emotion.

Please use this as a reference to obtain a stress-free and comfortable writing environment.

Why does prompt improvement through 'conversational back-and-forth' fail?

When the AI's output is lackluster,

"Do it more like this!"

Do you find yourself giving emotional follow-up instructions in the chat box like this?

In fact, this method is the biggest cause of unstable output quality.

AI has a systemic weakness where it loses track of past instructions as the conversation gets longer.

Here, we will first explain the fundamental mechanism of why conversational improvement is prone to failure.

Does AI not read the room? The systemic limitation of 'contextual wandering'

To put it simply, AI cannot 'read the room' like humans can.

In human-to-human conversation, it is normal to interact with the assumption that 'you remember what I said earlier, right?'

However, AI is a system that processes input information probabilistically.

As the chat continues, the amount of information grows, and the writing style or conditions specified at the beginning become increasingly buried.

We call this contextual wandering.

The more emotionally you layer your instructions, the more information the AI has to process.

As a result, you fall into a vicious cycle where the conditions you want it to follow become even more likely to be overlooked.

In other words, the very act of continuing a conversation in a chat can become the cause of lowering the AI's accuracy.

Just like seasoning a dish! Why fixing everything at once leads to confusion

When the output is not quite right, do you rewrite the entire prompt at once?

This is also a typical pattern that invites failure.

For example, imagine when a soup tastes bland.

What happens if you add salt, pepper, soy sauce, and dashi all at the same time?

Even if it happens to taste good, you have no idea what actually made the difference.

You can think of prompts using the same principle.

If you change the target setting, writing style, output format, and constraints all at once, you will not be able to determine which modification was effective.

If you have ever regretted it by thinking, 'The previous version was still better...', you are exactly in the state of having fallen into this trap.

Even if you try to revert it, you will end up lost because you do not know what or how to revert.

The key is to limit changes to only one element at a time.

This 'one-by-one' approach will dramatically change the precision of your prompt improvements.

The first step to improvement: Giving the AI 'criteria for judgment'

A problem often overlooked in prompt improvement is failing to tell the AI 'what the correct answer is'.

Instructions like 'make it better' or 'make it interesting' might work between humans.

However, for an AI, it cannot act if the criteria for 'good' or 'interesting' remain ambiguous.

For example, instead of instructing it to 'make the writing style brighter,' tell it to 'use exclamation marks at the end of sentences and keep each sentence under 40 characters'.

In this way, providing concrete criteria that the AI can judge is the first step toward improvement.

Convert emotional expressions into logical conditions.

Just by keeping this in mind, the AI's output will become remarkably more stable.

Reference article:

An improvement method called 'Single-Point Difference Testing' to reliably increase output quality

Prompt improvement does not require blind trial and error; there is a reproducible 'pattern' that exists.

That is the concept of the 'Single-Point Difference Test'.

This is a simple method that applies debugging techniques used by engineers to prompt improvement, which can be practiced even without specialized knowledge.

What is the 'Single-Point Difference (Single Variable Test)' that eliminates aimless revisions?

A single-point difference test is a method of testing by changing only one component of a prompt at a time.

Using the cooking example from earlier, it would be like adding only salt and tasting it, then adding only pepper and tasting it.

By following this 'one thing at a time' rule, you can accurately grasp how each change affects the output.

It might feel like a very tedious approach.

However, in the end, this is the most reliable way and the greatest shortcut to improving your prompt design skills.

Please keep this in mind as a fundamental concept of prompt design.

'Where to fix' becomes clear if you break down the elements

I understand that single-point difference testing is important, but...

'But which element should I start with?'

Many of you might be feeling this way.

The answer is simple: first, try breaking the prompt down into its constituent elements.

For example, the prompt templates I publish are composed of a collection of variables like the following.

• Prompt Description:
  Elements that provide the AI with task prerequisites (purpose, intended user, resources, goals, and steps to achieve them)

• Provided Information:
  Data, information, etc., that is the subject of processing

• Work Procedure:
  A 'thought process map' that defines the steps for executing the task

• Output of Deliverables:
  Specifications for appearance and structure, such as character count, composition, and format

If you analyze problems using these elements, it becomes easier to see where you should make adjustments.

Practice! 3 Smart Steps to Improve Prompts Without Confusing the AI

From here on, I will introduce concrete improvement procedures that you can use starting today.

There are only three things to do.

Once you master this simple flow, prompt improvement should become much easier.

1. Have the AI itself 'self-diagnose' the current output result

The first step is to have the AI itself point out the 'weaknesses of the current output'.

First, let's have the AI itself self-diagnose its weaknesses.

You can use the prompts introduced in the "60-Point Loop" article.

Prompt:

ではこの出力を60点とします。
これを60点とした時に、100点となる最高品質の出力とはどのようなものですか？
100点の出力にするために、現時点の出力に足りないものが何かを検証してください。
まだ100点案の出力しなくていいので、まずはあなたの見解をお聞かせください。

By doing this, the AI will objectively analyze its own output and tell you specifically what points need improvement.

By using this "self-diagnosis" as a starting point, the decision of "which single element to change" in the next step naturally becomes clear.

2. Don't continue the conversation; use "time leaps" to return to the past

Once you have identified the weakness, you might feel tempted to give additional instructions in the chat window.

However, this is the most critical turning point.

Instead of continuing the conversation, go back to the very first prompt.

Then, add the weakness the AI pointed out directly into the original prompt and execute it again.

I call this method a "time leap".

The reason for going back is that chat history often becomes "noise" for the AI.

Conversations with layered additional instructions tend to accumulate contradictory commands or unnecessary information, which often confuses the AI.

But if you return to the original prompt and revise it, the AI will be able to process the instructions in a clean state.

It might feel like a hassle, but this extra step significantly influences the accuracy of your prompt improvements.

For specific steps on time-leaping, please refer to the following article.

3. Compare outputs before and after the fix to verify the effect

The final step is to compare the outputs before and after the correction.

Let's place the result re-executed via time-leap side-by-side with the previous output.

At this point, what you should check is how the single element you changed has influenced the result.

If it has improved, that correction is correct; if it remains unchanged or worsens, you can decide to try a different element.

The value of the single-point difference test lies precisely in this "learning through comparison".

By repeating comparisons, you will accumulate a sense of "if I change this element, it changes like this."

It might feel like a hassle at first.

However, as this experience accumulates, the speed and accuracy of your prompt corrections will surely increase significantly.

Summary: A 'Great Leader' Guides with Logic, Not Emotional Criticism

In this article, for those who are struggling with how to improve when AI output does not go as expected, we covered:

• The root cause of why improvements through 'conversational back-and-forth' fail

• The concept of 'single-point difference testing' to reliably increase output quality

• 3 smart steps for prompt improvement that won't confuse the AI

I have discussed the above from the perspective of an author researching prompt design techniques.

The most important thing in prompt improvement is not to try to fix the output within the chat conversation.

As the conversation gets longer, AI loses sight of important past conditions, and every time you make a correction, other parts fall apart.

When things don't go as planned, instead of emotionally pushing back with 'That's wrong!', first try having the AI objectively evaluate it itself.

After that,

'Since this part is off, let's fix this one setting (blueprint)'

is to calmly apply a patch and restart the prompt from the beginning.

Once you master this 'self-diagnosis' and 'time-leap' pattern, you will be completely freed from aimless correction work.

Please stop what you are doing in the chat you have open right now and start by reviewing the very first prompt!

However, there may be cases where you really want to shape it while interacting.

In that case, I recommend utilizing the 'Symbiotic Intelligence Prompt' introduced in the following article!

Thank you for reading until the end 😇

✅ Click here for Poke-Go's content