When I felt AI's answers were shallow, it was my own articulation that was shallow
It wasn't the AI that was at fault.
For a while, I constantly felt that AI output was lacking. It would return something that sounded plausible. It was polished. But it never became something I could actually use. It would just return harmless generalities in neat bullet points.
So, I looked up how to write prompts. Give it a role, write constraints, make it think step-by-step. I tried them all. It got a little better, but not as much as I had hoped.
I found the cause in a completely different place.It wasn't that the AI's output was shallow, but that the questions I was asking were shallow.
AI is an amplifier, not a generator
What triggered a change in my way of thinking was viewing AI as an "amplifier."
An amplifier is a device that makes incoming signals larger. Conversely,if the input is close to zero, no matter how much you amplify it, the output will remain close to zero. It just results in more noise.
When you look at it this way, it becomes clear what prompt techniques are actually doing. They are about transmission efficiency. They are techniques for passing what you have on hand without loss, but they do not create the content itself.
And the problem I was having wasn't transmission, but that I had nothing to pass. When I asked it to "organize the points on this matter," I didn't even really understand what I was struggling with myself. If you don't understand it, you can't pass it on, no matter how much you refine your writing style.
The reason I was getting generalities back was because I was only providing generalities.
The day I asked for an organization of points and got generalities back
Let me talk about a specific example.
The other day, I needed to do some tedious organizing for my main job. It was the kind of thing involving how to take over facility management and how to handle contract transfers following an organizational merger. There were many stakeholders, and it was a mix of things that had been decided and things that hadn't.
At first, I asked it like this: "I want you to organize the points on this matter." As a premise, I added a few lines explaining the situation.
What came back was a textbook-like list of points. Nothing was missing. But every one of them ended with "Well, obviously." The parts I was really stuck on weren't included.
So I closed the AI for a moment and wrote three lines myself. What do I want to decide? What has already been decided? Where am I stuck?
As I tried to write the third line, my hand stopped.I hadn't been able to put into words what I was stuck on.After thinking for a while, it finally came to me. What I was stuck on wasn't the points themselves, but the fact that "it hadn't been decided who should make this judgment."
When I wrote that down and asked again, the response changed. This time, it was useful.
I didn't change a single character in how I wrote the prompt. The only thing that changed was how much I understood before I sent it.
AI becomes a mirror that reflects your own level of articulation
From this experience, one of my perspectives on how to use AI has changed.
The quality of the output directly reflects the quality of your own input. Therefore, AI can be used as a device to detect how much you truly understand something.
It is difficult to judge for yourself whether you only "think you understand." As long as things are in your head, most things feel like they are understood. However, when you throw them at an AI, this becomes blatantly obvious. If you throw a vague question, you get a vague answer. When the returned content is thin, it is usually a notification that your own understanding is thin.
Since I started thinking this way, I changed my procedure when I feel dissatisfied with the output. Before fixing the prompt, I test myself to see if I can explain the theme to someone in 3 minutes. If I cannot explain it, the cause is not on the output side, but on the input side. Tinkering with the prompt there just wastes time.
One more thing, I want to write this as a reminder to myself: The more accumulated knowledge a person has, the more they can notice contradictions and deceptions in AI output.In areas where you do not have your own criteria for judgment, you have no choice but to take what is returned at face value. If it sounds plausible, you let it pass. As AI accuracy improves, this gap will widen. The difference is not in the skill of using it, but in how deep the original articulation is.
What I do—3 lines before throwing, and 1 extra step after throwing
I currently have two main habits.
Write 3 lines before throwing.The 3 lines mentioned earlier are now my rules. 1. What I want to decide, 2. What has already been decided, 3. The points I am stuck on. It takes about 2 minutes, and if I cannot write the third line, I know at that point that the problem is on my side. Having these 3 lines clearly changes the quality of the output. It was more effective than memorizing 10 prompt templates.
After throwing, rewrite it into your own words.Do not just paste the organization returned by the AI into your documents or notes. Rewrite it by pulling it back into your own context once. If you skip this, only the deliverable remains in your hands, and nothing remains within you. When the same topic comes up next time, you will start from the same place again. Conversely, if you include this extra step, the resolution of the questions you throw next time will increase. The signal you put into the amplifier gradually becomes thicker.
Also, a small thing, but when I want an opinion, instead of saying "argue against this," I make it a point to ask, "Where is the weakness in my premise?". When you ask for an argument, the AI will line up general opposing opinions. When you ask for the weakness of a premise, it goes to read the premise you wrote, so the holes in your own thinking are returned.
To avoid misunderstanding—it is not that you should not use it until you have articulated it
Just to be clear, this is not about saying "use it only after you have organized your thoughts."
Rather, it is the opposite; I think using AI as a sounding board to advance articulation is the most compatible way to use it. You talk to it while in a fuzzy state, and while reacting to the returned words by saying "that's not it," you grasp your own contours. I do this daily as well.
The problem arises when you skip that process and adopt the words produced by the AI as your own conclusion as they are. Conclusions written in someone else's words do not generate the next question. Always place the task of rewriting into your own words at the exit of the sounding board process. If you keep only that, AI will not be an enemy of articulation, but the fastest accelerator.
If you want to improve the output, the thing to fix is not the input, but the original source
To summarize.
The amount you can draw from AI is determined by how well you have articulated the theme yourself. A prompt is a transmission technology, and it does not create what you are handing over. Therefore, when you feel the output is shallow, the thing you should work on is not the prompt or the choice of model, but your own understanding.
And this is probably the part where the difference will be most quietly made in the AI era. Since anyone can use the same model, the only place where a difference can emerge is on the input side. The more you have your own words, the more you can draw from the same tool.
Things to try this week
Please recall one question you recently asked an AI and thought, "this is kind of thin."
Before re-prompting, try writing just three lines. What do you want to decide? What has already been decided? Where are you getting stuck? If your hand stops on the third line, that is the answer. If you re-prompt after that and the output changes, it wasn't the AI that was shallow.
Follow and like to receive more.