For those who find ChatGPT's answers shallow: A basic template to stabilize output with JSON prompts

Introduction | Have you ever felt that ChatGPT's answers are shallow?

"I asked ChatGPT, but the response felt thin."

"I want a more concrete answer, but it ends with generalities."

"The output format is different every time, making it hard to use."

Have you ever had an experience like this?

Even though you are asking similar things, one day you get a proper answer, and another day it ends with harmless, generic content.

It was good when I asked Claude, but when I ask the same question to ChatGPT, it feels thin.

When I ask Gemini, it returns an answer with a slightly different vibe.

When this happens, it is easy to think, "Maybe the AI's performance is not that great."

However, it is often not the AI's performance, but the way you give instructions that is the cause.

"Do it nicely"
"Summarize it well"
"Make it easy to read"

With vague instructions like these, the AI has no choice but to answer "vaguely" as well.

That is where the concept of JSON prompts that I am introducing this time comes in handy.

This is not about difficult programming.

The JSON prompt referred to here is a method of organizing and passing instructions to the AI by item.

In this note, "Gorilla AI Lab | AI x JSON Laboratory," I share how to use generative AI to get more stable output while actually experimenting with it.

As a first step, I will introduce a basic JSON prompt template that is easy for beginners to use.

Why do normal prompts fluctuate?

For example, have you ever given an instruction like this to an AI?

At first glance, it seems like a common instruction.

However, this single sentence hides many "undecided things."

• Who is the article for?

• What do you want to convey?

• What tone should it be written in?

• What format do you want it in?

• What should be avoided?

Because these are not decided, the AI fills in the gaps with guesses, thinking, "Maybe it's something like this."

As a result,

• Even with the same instruction, a slightly different answer comes back each time

• The content becomes generic and harmless

• The target audience and tone deviate from your intentions

• The output format becomes inconsistent

These kinds of things are likely to happen.

Rather than the AI being at fault, this is a state wherethere is not enough information for judgment in the instructions itself.

Even between humans, if you are just told to "make it nice," the result will vary depending on the person, right?

Instructions to AI are the same.

The more clearly you separate and convey the conditions, the easier it becomes to predict the output.

What is a JSON prompt?

This is where JSON prompts come in.

When you hear "JSON," it might sound like programming and seem difficult.

However, you don't need to worry about syntax correctness here; just grasping the concept is enough.

In short, a JSON prompt is

a way to pass instructions to AI by dividing them into specific items

.

For example, if you organize the previous instruction, "Create a good structure for a note article," into items, it looks like this.

{
  "目的": "note記事の構成を作る",
  "想定読者": "生成AIを使い始めた初心者",
  "読者の悩み": "ChatGPTの回答が浅くて使いにくい",
  "出力形式": "見出しつきの構成案",
  "トーン": "やさしく、実用的に"
}

It is like taking instructions that were bundled in a paragraph and arranging them into separate cards for each item.

By using this format,

• what the task is for

• who the content is aimed at

• what problems it answers

• what format you want it in

• what tone you want it written in

become easier to see.

It becomes easier for humans to understand "what is being requested," and for the AI, it organizes "what to create and how to create it."

This is the reason why using JSON prompts makes the output more stable.

A basic JSON prompt template you can use right away

This is where the main part of this article begins.

First, I will introduce a basic template that is easy to use in any situation.

{
  "目的": "",
  "想定読者": "",
  "読者の悩み": "",
  "前提条件": "",
  "AIにしてほしいこと": "",
  "出力形式": "",
  "トーン": "",
  "含めてほしい要素": [],
  "避けてほしい要素": [],
  "確認したいこと": ""
}

You can use it simply by filling in the blanks in this template and pasting it directly into ChatGPT or Claude.

I will briefly explain each item.

Objective

Write down what you want the AI to work on and why.

For example,

• creating an article outline

• generating title ideas

• summarizing text

• organizing ideas

and so on.

By stating the objective first, it becomes easier for the AI to grasp the direction of the task.

Target Audience

Write down who the text or ideas are intended for.

For example,

• beginners who have just started using generative AI

• people who want to start posting on note

• people interested in AI image generation

• people who want to learn prompt engineering

Write it like this.

When the reader is clearly defined, the depth of the explanation and the choice of words will change.

Reader's pain points

Write about what the reader is struggling with.

Including this prevents the content from ending up as mere generalities.

For example,

• ChatGPT's answers are shallow

• Articles become cliché even when asking AI

• Cannot keep characters consistent in image generation

• Articles on note are not being read

and so on.

Prerequisites

Write down the conditions you want followed and background information.

For example,

• Do not make it sound like a sales pitch

• Make it suitable for beginners

• Break down technical jargon

• Do not hype up monetization

are examples of such conditions.

What you want the AI to do

Write the task you want the AI to perform specifically.

Instead of just saying "write an article,"

• propose 5 title ideas

• create a heading structure

• create 3 variations of an introduction

• rewrite existing text to be more readable

writing it like this will make the output more stable.

Output Format

Write the format in which you want the response returned.

For example,

• bullet points

• table

• heading structure

• checklist

• JSON format

• title ideas and body text drafts

and so on.

Specifying this makes it easier to keep the output format consistent every time.

Tone

Write the desired tone of the text.

For example,

• gentle

• practical

• not too technical

• not inflammatory

• empathetic to the reader

• flat/neutral

and so on.

Specifying the tone helps reduce the feeling that something is 'off'.

Elements to include

Write down keywords or perspectives that must be included.

Using a list format makes it easier to prevent omissions.

Elements to avoid

Write down expressions you want to prohibit or tones you want to avoid.

For example,

• expressions like 'anyone can easily make money'

• unsubstantiated assertions

• explanations full of jargon

• expressions that incite the reader

and so on.

Including this will reduce the amount of back-and-forth for revisions.

Things to verify

Finally, write down the perspectives you want the AI to check.

For example,

• Is the content understandable even for beginners?

• Is the text too long?

• Does it contain inflammatory language?

• Does it answer the reader's concerns?

and so on.

You don't need to fill in every item every time.

To start, just writing down the 'Purpose,' 'Target Audience,' and 'Output Format' will make it much easier to predict the output.

Practical Example | JSON Prompt for Creating a Note Article Structure

I will now introduce a sample of how to use this template to 'have an AI create a structure for a Note article.'

{
  "目的": "note記事の構成を作る",
  "想定読者": "ChatGPTを使って記事を書きたい初心者",
  "読者の悩み": "AIに頼んでもありきたりな記事になる",
  "前提条件": "売り込み感は出さず、実験記録として書く",
  "AIにしてほしいこと": "記事タイトル案と見出し構成を作る",
  "出力形式": "タイトル案5個、見出し構成、導入文案",
  "トーン": "やさしく、実用的、煽らない",
  "含めてほしい要素": [
    "AI活用",
    "JSONプロンプト",
    "失敗例",
    "改善方法"
  ],
  "避けてほしい要素": [
    "誰でも簡単に稼げる表現",
    "専門用語だらけの説明",
    "根拠のない断言"
  ],
  "確認したいこと": "初心者にもわかる内容になっているか"
}

Using this prompt will lead to several changes.

First, the AI will be less likely to get confused about 'what it needs to create.'

Because the purpose and task are clear, off-target suggestions will decrease.

Next, the output format will stabilize.

Since you specified '5 title ideas, heading structure, and introduction draft,' it will be easier to get the same format returned every time.

Also, because you have written down the target audience and their concerns, the content is less likely to be generic and more likely to be tailored to the reader.

By writing down 'elements to avoid,' you can also clearly communicate prohibited items.

Also, if the output is not what you expected, you can see which items need to be changed, making it easier to request revisions.

Asking vaguely and correcting vaguely.

From there, you can make adjustments while looking at the items.

This difference is a major benefit of JSON prompts.

Notes on using it with ChatGPT, Claude, and Gemini

Even if you use the same JSON prompt, the tone of the output may vary slightly depending on the AI.

Roughly speaking, there are times when you might feel the following differences:

• ChatGPT: Versatile and easy to use for brainstorming and general writing tasks

• Claude: Sometimes feels better suited for structuring long texts or creating natural flow in writing

• Gemini: Sometimes feels worth trying for organizing information, making comparisons, or combining with Google services

This is not a discussion about which one is the best.

It is simply that the tone of the output and the areas of expertise differ slightly depending on the AI, even with the same instructions.

This article is not intended to compare the performance of AIs.

However, the concept of the JSON prompt introduced here can be used commonly across any AI.

In future posts on this note, I plan to test the differences in output by feeding the same JSON prompts to ChatGPT, Claude, and Gemini.

How much will the results change with the same instructions?

I plan to compile those findings as experimental records as well.

Points to keep in mind when using JSON prompts

While JSON prompts are a useful concept, they are not a panacea.

This is a point I want to convey honestly.

First of all, just because you use JSON format does not mean you will always get a good answer.

If the quality of the information you put in is shallow, the quality of the output will also be shallow.

If you write without thinking deeply about the "intended reader" or "purpose," the content itself may not change much even if the format is organized.

Also, if you include too many "elements to avoid," the AI may become overly cautious, leading to stiff writing or harmless, generic expressions.

It is recommended to limit prohibited items to only those that are truly important.

And the most important thing is not to use the AI's response as is.

Using JSON prompts makes it easier to organize output.

However, before finally publishing, you must always check it with human eyes and edit it in your own words.

JSON prompts are not tools for dumping everything onto the AI.

They are tools for organizing your interaction with the AI.

I think it becomes much easier to handle when you use it with this mindset.

Summary | Prompts are blueprints, not requests

So far, I have introduced the basics of JSON prompts.

Finally, I will summarize the contents of this article.

The trick to using generative AI well is not to memorize special "magic words."

What is important is organizing your instructions to the AI.

"Make it feel good"
"Summarize it well"
"Make it easy to read"

From these kinds of vague requests,

• purpose

• intended reader

• Reader's pain points

• Output format

• Tone

• Elements to avoid

Change the format to one that conveys these by separating them into individual items.

Just by doing this, output fluctuations will decrease, and it will become easier to handle revisions.

A JSON prompt is a blueprint for achieving that.

At first, it is okay if you do not fill in every single item.

Start by trying the basic template introduced here just once, instead of your usual prompt.

The overall theme of this note and the experimental content I will be handling in the future are summarized in this pinned article.

▶ [For first-time visitors] This note is a laboratory for "Generative AI x JSON x Content Creation"

In this note, I record what I have actually tested under the theme of "Generative AI x JSON x Content Creation."

Next time, I plan to compare how the output changes across ChatGPT, Claude, and Gemini using this template.

Articles on Generative AI and JSON prompts are summarized in this magazine.
▶ AI x JSON Prompt Laboratory

If you are interested in what comes next, I would be happy if you followed me and read the next article.

I also compared the differences in article structure by feeding the same JSON prompt to ChatGPT, Claude, and Gemini.
If you want to know how to use these three AIs differently, please check this out as well.