Why Prompts Fail: Cognitive Load and How to Use Markdown Structuring

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

When giving instructions to AI via prompts,

“I wrote down various conditions, but they were completely ignored...”
“I thought I was being specific, but for some reason, the output is different from my intent...”

Have you ever felt this kind of frustration?

For simple instructions, the latest AI can respond accurately without the need for complex structuring.

The problem arises when giving business-level instructions that involve multiple conditions.

In such situations, if you just write things down without organizing the requirements, you may miss requirements yourself, or the AI may fail to understand your intent correctly.

A gap is created where you think you have communicated, but it hasn't been understood.

This is the true nature of “cognitive load” in prompting.

Structuring with Markdown is not meant to make the AI smarter.

It serves to organize the writer's thoughts in order to eliminate this cognitive gap.

In this article, for those struggling with prompts that are hard to convey, I will explain:

• The mechanism of “cognitive load,” the cause of prompts not being understood

• How to turn instructions into “easy-to-read manuals” using Markdown

• How to distinguish between situations that require structuring and those that are fine as natural text

Regarding the above, I explain this while incorporating practical knowledge gained from verifying prompt design on a daily basis.

Prompt structuring is not a special skill.
It is a communication habit for organizing requirements and conveying them to AI without discrepancies.

Please use this as a reference and try starting by organizing 'what you are looking for' using headings and bullet points.

The reason prompts 'don't get through' is cognitive load

When instructions to an AI are not conveyed well, the cause might not be that 'your writing style is poor'.

In fact, a problem called cognitive load, which is more likely to occur as prompts get longer, is causing the gap between the writer's intent and the output.

Here, let's organize how this cognitive load creates situations where things 'don't get through'.

The mechanism by which AI overlooks conditions in long instructions

'I wrote 8 conditions, but only half were reflected...'

Have you ever had an experience like this?

The background to this phenomenon lies in the 'way of writing' the instructions.

When you pack conditions into one long sentence, the boundaries become ambiguous regarding where one condition starts and ends.

Even in human-to-human conversation, if requirements are rattled off all at once, the listener gets confused, right?

Try thinking that the same thing is happening with AI.

In other words, the reason AI "overlooks" conditions is not due to a lack of AI capability, but rather the cause lies in the ambiguity of the writer's instructions.

This is the true nature of "cognitive load" in prompts.

If your instructions become long, first try questioning whether the boundaries of the conditions have become ambiguous.

Organizing is not "for the AI," but "for yourself."

"If I use Markdown, the AI will become smarter."

Some of you may have that impression.

However, structuring is not "magic that makes AI smarter."

From my experience continuing prompt research, the greatest benefit of structuring lies in "the writer's own thoughts being organized".

This is because the act of separating information with headings and bullet points is, in itself, the process of clarifying "what I am looking for."

Those who have felt, "Maybe the way I write my instructions is bad..." may be standing right at the entrance to this realization.

Once you actually start organizing,

"Oh, this condition was missing..."
"These two were contradictory..."

Discoveries like these should naturally emerge.

Many of the discrepancies where "I thought I communicated it, but it wasn't understood" are caused not by the AI's comprehension, but by the precision of the writer's requirement organization.

Try thinking of structuring not just as instructions for the AI, but also as an act of organizing your own thoughts.

Transforming prompts into "easy-to-read instructions" with Markdown

Now that you understand how cognitive load works, let's move on to concrete solutions.

To reduce the cognitive load of your prompts, you only need these three Markdown notations: "Headings (#)," "Bullet points (-)," and "Bold text (**)".

Basic usage of headings, bullet points, and bold text

"Isn't Markdown something only engineers use...?"

Some of you might be thinking this.

However, the notations used in prompts are mainly the following three types, all of which are intuitively understandable.

• Headings (#): Indicates a break in the topic. By simply writing things like "# Prerequisites" or "# Output Format," you can tell the AI that the theme is changing from that point on.
  

• Bullet points (-): Organizes parallel conditions. Instead of writing "The tone is casual, the length is 800 characters, and the target is people in their 30s" in one sentence, you just list one condition per line.
  

• Bold text (**): Highlights the most important conditions. By enclosing text like "**Rules you must follow**," you can expect the effect of focusing the AI's attention.
  

These three are based on the same concept as "report formats" or "meeting agendas."

You can start using them today without any special technical knowledge.

Assemble them in the order of "Separate -> List -> Highlight"

Even if you know the three types of notation, you might sometimes wonder, "In what order should I use them?"

I recommend thinking in terms of three steps: "Divide, List, and Highlight."

First, divide the entire prompt into meaningful sections using headings.

Think of it as drawing boundaries between topics, such as "Prerequisites," "Output Rules," and "Reference Information."

Next, list parallel conditions within each block using bullet points.

Finally, highlight the most important points you absolutely need the AI to follow by using bold text.

Just by following this order, you should see a significant reduction in missed instructions.

This is because the act of "dividing" first helps you see the big picture of what actually needs to be communicated.

Instead of starting with the details, try organizing the big framework first.

How to distinguish between situations where structuring is necessary and where it is not

So far, I have explained how to utilize Markdown.

However, you don't need to think that "every prompt should be structured."

Structuring is just a tool, and it is important to identify when to use it.

For simple instructions, natural language is fine.

If a task can be handled with this level of instruction, there is almost no need to go out of your way to structure it.

In my own verification results, for simple instructions with about three conditions, there was almost no difference in output quality whether written in a single sentence or structured.

The latest AI models have a very high level of natural language comprehension and can accurately grasp the context of short instructions.

There is no need to brace yourself thinking, "I have to structure this..."

For casual, everyday requests, a conversational style is more than enough.

The important thing is to keep in mind that structuring is not a "mandatory etiquette," but rather a tool to be used when necessary.

When conditions become complex, it is time to use structuring

On the other hand, when it comes to professional-level prompts, the story changes.

"Write in a different tone for each section," "Do not use specific terms more than three times," "Include keywords in the first 100 characters."

Imagine a situation where these conditions pile up to five or ten.

If you continue to write in a single-sentence format at this point, the problem arises that the writer themselves becomes unable to organize the requirements.

"I'm sure I wrote everything correctly, so why isn't it being reflected...?"

If you have had this experience, you might be entering this exact zone.

The key lies in a shift in perspective: "Structure to organize your own thoughts."

Structuring is not "magic to make AI smarter."

It is a technique for organizing communication to prevent misunderstandings between the writer and the AI.

If you are writing prompts that require quality and reproducibility for work, try starting by organizing what you are looking for using headings and bullet points.

Summary: The essence of structuring is an organizational technique for writing clear instructions.

In this article, for those who want to review how they write prompts and improve the accuracy of their instructions to AI, I have covered:

• The mechanism of cognitive load that causes prompts to fail to communicate

• How to turn prompts into readable instruction manuals using Markdown

• How to distinguish between situations where structuring is necessary and where it is not

I have discussed the above while incorporating my own experience as someone who has researched and practiced prompt design.

Structuring is unnecessary for simple instructions, but it is a different story in professional scenarios where conditions become complex.

The ability for the writer themselves to realize, "Oh, I missed this condition," during the process of organizing.

This is the true value of structuring.

Once you acquire this habit, you will be freed from the frustration of "I thought I explained it, but it didn't get through."

The accuracy of your instructions will stabilize, and reproducibility will be created in your collaboration with AI.

First, try starting by organizing "what you are looking for" using headings and bullet points the next time you give instructions to an AI.

That small extra effort will be the first step toward changing the quality of your prompts!

✅ Click here for Poke-Go's content