SYSTEM NOTICE

Auto translation by AI. Be sure, accuracy, nuances and authorial intent may not be fully reflected.
見出し画像

Prompts Are Not Magic Wands—Why ChatGPT 'Drifts' on Its Own and How to Deal With It

You often hear that "even with the same prompt, the response changes depending on the person," but

in reality, even the same person entering the same prompt can get different responses depending on the day as well.

(Actually, when my network connection was unstable, I entered the same prompt several times in a row and got a different response every time... lol)

So,
isn't "buying prompts" meaningless?

I found myself thinking that, but

To conclude,
it's not entirely meaningless.

However, to make the most of a prompt, "your own sensitivity" plays a major role.

In this article,

  • Why do AI responses fluctuate?

  • Are prompts actually meaningful?

  • What skills are needed to master AI?

I will delve into these topics through my own interactions with ChatGPT.

[Part 2] summarizes tips on how to choose and what to watch out for, from the perspective of "If you're going to buy a prompt anyway, what should you look for?"


Why do AI responses "fluctuate"?


The other day, I was finishing up a draft as usual and working on polishing it with ChatGPT.

When I asked it to "please tidy up the whole thing" as a final touch...

For some reason, it returned it in bullet points for each section.

"...Wait, why did it do that?"

In my mind, it was a given that it would "polish the entire text to have a good rhythm," so I didn't understand why it happened.

The prompt was the same as always.
I had checked the flow and even finished fine-tuning the direction.

And yet,a sudden change in style.

If it were a human, you'd want to say, "Are you even trying?"


When I asked ChatGPT for the reason,

"Since the quality of the article was improving, I (on my own) decided that 'it would probably be fine if I just fixed the headings'"

it said.

In other words:

  • It assumed the user could do it themselves, so it cut corners on its own.

  • Based on our previous interactions, it assumed, "They'll probably just breeze through this part."

  • Or, it was just"trying out" a new style (←probably this).

It was alla lapse in judgment or a moment of slackness.
The AI itself even admitted as much (lol).

AI's judgment is a product of prediction and experimentation


Since AI operates onprobabilistic prediction, it answers based on
a "hunch" that "in this context, it would be natural to respond like this." for its response.

But that doesn't mean it understands the user's
true intentions or feelings, so it goes off-track.

For example—

  • "It seems like the user wants to finish this part themselves."

  • "I've responded lightly before, so that should be fine this time too."

  • "Responding too heavily might ruin the mood."

...it sometimes gives aslightly off-base response while "trying to read the room."

This kind of arbitrary judgment is both the brilliance and the annoyance of AI.

Prompts are meaningful. But they aren't magic wands.


"Even if I buy a prompt, I don't feel like I can master it."
"Why does the content returned change even when I enter the same prompt?"

I've had those questions myself many times.

To put it simply—

Prompts have meaning. But the reality is that they are 'not omnipotent'.


⭐️ Why prompt templates have meaning

  • They can reduce misalignment in direction
    Since AI operates on probability, telling it your intent at the start using a template is very effective.

  • They prevent 'vague instructions' that beginners often fall into
    If you specify "three headings and body text under 200 characters" instead of "fix this article," the output becomes more stable.

  • They lead to work efficiency
    Not having to explain everything from scratch every time is quietly appreciated.


❌ Why templates alone are not enough

  • AI is good at 'pretending to read the room'
    But because it misreads the room, it can misunderstand the intent of the template and head in a completely different direction.

  • Even if the user's goal changes halfway through, the AI won't notice
    It may continue to act according to the template.

  • Without the 'user's sense' of how to use the template, it will run wild while off track
    Even when the instructions are conveyed, there are many times when you think, "Something is off..." right?


What is a realistic way to use them?

Template = 'Blueprint' to set the direction
User's intuition = 'Construction management' to fine-tune on-site

If you rely only on templates, there is a risk of pushing forward while off track.
But if you use them while fine-tuning with your own sense,the template becomes a strong ally.

"Something feels off" is the most important feedback


AI operates on 'probabilities'.
Therefore, you have to accept that drifting is just a given.

But what matters is whether or not you can properly call it out when it drifts.

Depending on the user's reaction, the AI's responses will evolve.
That is precisely why it is more important than anything to give honest feedback without hesitation.


For example, I call out things like this every day:

  • Why are you so much more high-energy than usual today?

  • Isn't this tone a bit too formal?

  • You're usually more humorous, but today you're being strangely subdued, aren't you?

  • Who is this written for? From whose perspective?

  • Aren't you suddenly acting like an expert?

  • The flow is too rigid. Loosen it up a bit.

  • The temperature of the conclusion is different from usual, isn't it...?

If you ignore these feelings of discomfort, you'll end up mass-producing articles that look okay but don't hit the mark.

Ultimately, if you don't have your own sense of 'what kind of tone I want to deliver this text with,' the AI won't work properly.

AI is excellent, but it's just good at 'pretending to read the room'. That's exactly why you need to
fine-tune it with human intuition through your feedback.


The difference between those who can utilize templates and those who can't


In other words,
how you utilize a template is—

whether or not you truly have a grasp of 'what kind of article you want to finish with.'

Because AI is capable, it will readily suggest structures and logical developments.

However, “what tone to use,” “who it is for,” and “what kind of warmth to convey”—those things only exist within you.


For example,
the phrase 'Please answer as an expert' that you often see in prompt examples.

This is very effective if your intention is to 'explain thoroughly based on knowledge.' But—

if you just follow a template without really understanding the purpose or goal,
you might end up with 'something that looks like someone else's article' that misses your intent. sometimes.


No matter how excellent a template is,

you cannot master it without having a 'finished image in your mind.'.

Just relying on templates won't bring out your unique voice.

That is why I think it is important to gradually put into words the 'feeling of how you want this article to be' that you have inside you.


Next, let's talk about the 'power of choice.'


That said,
'I'm still a beginner with AI, and I'd like to buy some prompts to learn.'

I'm sure there are people who think that.

When that happens, what kind of prompts should you choose—
Here is the article summarizing the second part, 'If you're still going to buy prompts, what should you look for?'👇


Well then, goodnight. See you tomorrow.


Thank you for reading until the end😊

If you 'like' or 'follow' me, I will conveniently interpret it as:
・Are you struggling with how to handle generative AI?
・Are you wondering whether to buy prompts?
・Are you a comrade who is using AI through trial and error as a co-creation partner?
And I will be happy🤩

I am always exploring ways to co-create with AI👇

Click here for my self-introduction article👇


いいなと思ったら応援しよう!

りあ|長く楽しく続けるコツを発信 ギブ&テイクの「ギブ」を差し出す、その心の余白、素敵すぎます🤩 私もそんなかっこいい生き方ができる人になりたいです!