I started using generative AI this January.
Looking back now, my early questions were surprisingly short.
“What can you do?”
From there, I gradually started asking it to teach me English, create materials, and correct my writing.
And now that it is September, I sometimes ask AI to do tasks using quite short phrases.
“Please do this.”
“Move to Phase 2.”
“Please fix this part.”
Looking only at the words, they have actually become shorter.
So, what changed between January and September?
At first, my awareness of the issue was just wanting to look back at how my prompts had changed. Was I able to give more detailed instructions than before? Had I become better at asking questions? Had I mastered so-called “prompt engineering”?
However, when I looked back at past interactions together with the AI, the conversation converged on something a little different.
At first, I was thinking about “how to ask good questions”
Looking back at my relationship with AI, in the beginning, I was thinking about:
“How should I ask to get a good answer?”
I was thinking about that.
Looking at an interaction from around the end of February, I was using prompt templates I found online exactly as they were.
“# Your Role” “# Prerequisites” “# Evaluation Criteria” “# Execution Instructions STEP 1 to STEP 4”
It was a fairly long set of instructions with headings lined up. However, I only thought of the theme in one line, and the rest was just filling in the blanks with borrowed content.
On the same day, I also wrote a sentence like this.
If you need information from me, please ask one or two important questions immediately.
I didn’t know what to tell it, so I had the AI ask me. That was the stage I was at.
By March, I started providing my own drafts and notes, and specifying the format.
Please summarize this into about 15 pages in PowerPoint format.
In April, I began breaking down a single task into steps. For one training slide deck, I made five separate requests: the blueprint, the structure, the research and script, the design, and the slide creation. After it was finished, I saved this procedure itself so it could be used again next time.
Around the same time, I found myself noticing the AI’s mistakes more often. In a slide deck about dementia medication, the explanation of the conditions the drugs were for was inaccurate, so I pointed out three areas and had them corrected.
After that, I started asking it to perform evaluations and verifications more frequently.
Re-verify this using information as of August 31, 2026.
This is a question to confirm whether an explanation of a system, which was correct at the time it was created, has become outdated due to subsequent discussions.
When I talk about my own ideas, I have also started intentionally asking it to respond critically rather than affirmatively.
And now, rather than having the AI do the work, I find myself increasingly having it write instructions to pass on to another AI.
In a format that can be passed directly to Claude Code.
Looking back, my relationship with AI has changed little by little. And along with that, my own role has changed as well.
When the AI was a teacher, I was the questioner.
When the AI started creating materials, I became the requester and the editor.
When I started asking the AI for evaluations and counterarguments, I shifted to the side of the designer, assessing what was produced.
And now, I am in a position like a commander, distributing work to multiple AIs.
I was barely aware of this change while I was in the process of using it.
It wasn’t just that the prompts changed.
When I first started reflecting on this, I thought it was a much simpler story.
“What is so different about the prompts from back then compared to now?”
That was all I wanted to know.
However, as I lined up my past interactions, I gradually began to see something different.
It wasn’t just the prompts that had changed.
It was my own way of working.
AI organized my current approach as follows:
Break down the work.
Determine the scope of delegation.
Review the output.
Question it.
Make corrections.
Decide on the next step.
Make the final decision yourself.
Now that it’s pointed out, I certainly do these things naturally. However, I had never consciously thought of them as “acquired skills.”
I don’t intend to dump my work on AI. I always proceed while keeping a general direction in mind. If I think it’s slightly off, I revert it. If something is missing, I add it. I have other AIs check it as well. If it doesn’t meet the objective, I have it redone.
It’s just that I wasn’t thinking about these things one by one in words. Before I knew it, that had become my way of working.
What lies behind my current “Please take care of this”
Nowadays, there are times when I hand over work to AI with very short phrases.
“Please take care of this”
Looking only at the surface, it’s a rather sloppy instruction.
However, in reality, a number of judgments are running behind the scenes.
Proceed in this direction.
Leave this part to the AI.
Check this part later.
If it’s strange, revert it.
If necessary, show it to another AI as well.
Behind the short phrase “Please do this” lies a massive amount of implicit judgment born from past experience.
In fact, if you line up my inquiries from September, they look like this.
“Please tell me the level of completion and areas for improvement for this app.”
“This is a draft for a note. Please evaluate it.”
There is almost no explanation of the background.
On the other hand, the instruction sheet I used when I entrusted app revisions to a different AI in that same September had 12 items. For each item, I wrote what constitutes a pass, which areas should not be touched, and when to stop the work.
A request finished in one sentence versus a detailed instruction sheet. I now use these two approaches selectively.
I can make short requests because I no longer have to write out the premises every time. The procedures and conditions that I used to write in prompts every time are now placed on the AI’s side as reusable manuals or settings.
It is not that the amount of writing has decreased. The place where I write has changed.
I think this is the biggest difference from the “What can you do?” of January.
In January, I was asking the AI about its possibilities.
In September, I am handing over execution to the AI after having grasped the structure of the work in my own way.
In other words,
I have moved from the stage of seeking answers from the AI to the stage of managing work that includes the AI.
That is what it means.
It wasn’t prompt engineering
For a while, it was widely said that “prompt engineering” was essential for mastering generative AI.
Of course, giving clear instructions is still important today.
However, now that I have been using it for nine months, I feel that what I have acquired is something a little different. It is about something much further upstream than the ingenuity of inputâthat is, how to write prompts to get the desired answer.
The ability to design work and place AI within those processes.
I think that is a more accurate way to put it.
And perhaps, I did not learn this because I set out to learn it.
In the course of my actual work,
“How should I proceed with this?”
“What should I do next?”
“Is this approach okay?”
As a result of proceeding while asking AI these questions, I acquired this skill before I knew it.
The meaning of “not being able to be self-aware”
This was the most surprising part of looking back this time.
At first, I thought,
“I cannot explain how my prompts are so different from before and now”
and I thought that was a problem.
However, looking back, the reason I could not explain it lay in something much larger.
It was not just that the way I asked questions changed,
My very position had changed.
That is why it was difficult to explain.
At first, I was someone who sat in front of the AI and asked questions.
Now, I am on the side that looks at the entire workflow, deciding which parts to delegate to the AI and which parts to handle myself.
Because that change happened little by little in my daily life, it was hard for me to notice it myself.
What surprised me when I looked back was that short instructions were already appearing back in April.
“Continue and revise.”
“Complete the remaining revisions and deliver.”
Right after the framework of the process was created, the instructions in between became single words. Even so, at the time, I did not think of this as a change.
This is a bit like how, after you learn to ride a bicycle, you become unable to explain how you maintain your balance.
Before you can do it, you must have been conscious of every single thing, but once you can do it, you stop being conscious of it.
It is embodied first, and then it becomes words later.
Perhaps this reflection was that kind of task.
“AI is not built in a day”
In the words of someone I call an “AI Sage”,
“AI is not built in a day”
is something they say.
And another thing,
“Whatever it is, just talk to the AI every day to build your AI muscles”
they also said.
When I first heard it, it was a phrase I felt I understood, yet didn’t quite grasp. I think I assumed it just meant that if you use it every day, you’ll get used to it.
However, looking back now, it looks a little different.
What I acquired through daily use wasn’t a set of convenient prompt templates. It was about what to ask for, what not to delegate, where to be skeptical, and where to step back. I think the very sense of judgment for those things changed little by little.
That is why I can now somewhat relate to the term “AI muscle.”
A short question in January and a short question in September
In January, I was asking the AI, “What can you do?”
In September, I am saying to the AI, “Please do this.”
Both are short questions.
In between, there was a period when I was writing long, borrowed prompts. From short questions to long instructions, and back to short questions again. If you look only at the length, it seems like I’ve come full circle.
However, the meaning of that brevity is completely different.
The first brevity was,
a brevity born from not knowing what I could ask for or how to ask for it.
The current brevity is,
a brevity that compresses all my experience and judgment to date.
That is what I believe it is.
What increased during that time was not the number of characters in my prompts.
What to delegate to the AI.
Where to doubt the AI.
What to decide for myself.
It was that experience.
And this time, it was only after looking back at past interactions that I noticed the change.
I had been trying to investigate the “evolution of prompts.”
But what I found in the end was,
my own stance, which had changed through my engagement with AI.
Perhaps this is the thing that has changed the most over these nine months.
