by Serguey Shinder
For a while I had a habit of blaming the tool. I would ask a model to do something, get back a mediocre result, and mutter that it just did not understand what I wanted. I rewrote prompts, added instructions, layered on constraints, all in the spirit of forcing a stubborn machine to finally get it. The results improved a little and my frustration stayed exactly the same.
The uncomfortable turn came when I read one of my own prompts aloud, the way I would read it to a new colleague. It was a mess. It asked for three things that quietly contradicted each other. It assumed context I had never stated. It used a word two different ways in the same sentence. The model had not misunderstood me. It had understood me precisely, and what I had actually said was muddled.
That was a humbling mirror. I had been treating my own thinking as clear and the tool as the source of confusion, when the truth ran the other way. The vagueness I kept blaming the model for was mine. I simply had not been forced to confront it before, because people are generous. A human colleague would have quietly guessed what I meant and covered for my sloppiness. The model just did exactly what I said.
I have come to see these tools as unusually honest mirrors of the clarity, or lack of it, that I bring to them. A confused request produces a confused answer, and the confusion is legible in a way it never was when it lived only in my head. Writing a good prompt turns out to be indistinguishable from thinking a clear thought, and I had been avoiding the second by fiddling with the first.
Now, when a result disappoints me, my first move is not to add more instructions. It is to ask whether I actually know what I want, precisely, in words, before I demand that a machine produce it. Half the time the answer is no, and the act of getting to yes solves the problem before the model is even involved.
This has changed how I work far beyond any single tool. The discipline of saying exactly what I mean, of noticing my own contradictions before I ship them to something that will take them literally, has made me a clearer thinker in general. The machine did not teach me anything I could not have known. It just refused, politely and relentlessly, to cover for the parts of my thinking I had never bothered to finish.
– Serguey Asael Shinder
Leave a Reply