The first image I made for this site was the AI image you have seen everywhere. The blue brain. The glowing orb. The abstract human figure holding light in her hands. I deleted it. I have been starting over, with increasing precision, for a year.
Every image on this site is one I prompted. I do not buy stock. I do not hire a designer. The pictures are mine, the way a written sentence is mine.
I thought I was learning to make images. I learned something different. I learned how to say what I mean before anyone else sees it. The image arrives as evidence. The prompt is the work.
This essay is what a year of prompts taught me about writing the sentences that go above them. It is not a tutorial. It is the story of a year I spent making pictures with a machine and discovering, halfway through, that I was learning to write. I will tell you what I made. I will tell you what I deleted. I will tell you what I noticed about my own writing that I had been missing for thirty years.
The first failure was not the model, the mirror#
The model is doing what you said. The model is good at doing what you said. The problem is what you said.
The blue brain is what you get when you ask the model for “an image about artificial intelligence.” It is also what you get when you ask for “an image about my essay on artificial intelligence.” The model learned from a thousand essays whose authors did not know what they meant. It has learned the picture that matches the vagueness. The picture is the blue brain. The blue brain has a cousin. The cousin is the human silhouette against a gradient sky. The cousin’s cousin is the businessman in a suit shaking hands with a robot. These are the three faces of every AI image that has ever been posted. They are also what the model wants to make when you have not given it anything more particular than “AI.”
I did not know this for the first six months. I thought the model was generic. The model was holding up a mirror.
The first thing I had to learn: the failure was not the picture. The failure was the prompt. The failure was the sentence that came before the prompt. The failure was the argument that came before the sentence. The blue brain is the average output for an average input. The model was telling me my input was average. I had thought it was not.
I had written essays I believed were sharp. I asked the model to make images for them. The model returned blue brains. The blue brains were correct. They were the most likely output for the prompts I had given. The prompts I had given matched the arguments I had made. The arguments were less sharp than I had thought.
That was the year. Not learning to prompt. Learning what I had been writing without seeing it. The image is a diagnostic for the sentence.
The prompt is a statement, not a request, speaking before the listener exists#
A prompt is not a request. It is a statement.
When you write a prompt, you are not asking the model for anything. You are telling the model what you mean. The model does its best to draw what you said. If what you said was clear, the image is clear. If what you said was a thousand things at once, the image is the average of the thousand things. The image is honest. The prompt was not.
A prompt can be a hundred words. A prompt can be three thousand words. There is no right length. There is only the moment when you have said what you meant, and the model has heard you. Some days that moment arrives in a sentence. Some days it takes me a paragraph. Some days I write three thousand words to a model and the model returns an image that surprises me, because three thousand words was what it took to find the one thing I was trying to say.
This is the discovery the year gave me. Prompting is not a craft of asking. It is a craft of stating. You are speaking before the listener exists. You are saying what you mean to nobody, and then a picture arrives.
This is also where the engineer and the artist in me met for the first time. The engineer specifies. The artist chooses what to specify. The prompt is the only place I have found that uses both at the same time. Most of my career has used one or the other.
Once I noticed it, I could not unsee it. Prompting was the form I had been waiting for without knowing I was waiting.
I have written prompts in voice memos while driving. I have written prompts on receipts. I have written prompts at the kitchen counter at six in the morning when nobody else is awake and a sentence is forming. The form is portable because it is thinking out loud, in writing, with a listener who is not yet there.
A year of prompts, the vocabulary I built#
The vocabulary I built over the year was not technical. It was personal.
The phrases I keep using now are not industry phrases. They are the ones that turned out to mean something to me. “A kitchen at six in the morning.” “Light from a single source.” “The hand of a person who has done the work for a long time.” “Not the conclusion. The moment before.” None of these are art-direction terms. All of them are what I would have written in an essay if I had been writing the essay with care.
The phrases that stopped working were the ones I had thought would work. “AI orchestration.” “Enterprise software.” “A diverse team in a modern office.” Each is a request to the average. Each returns the average. The average is the blue brain, in a different costume.
What I started doing instead: writing the prompt the way I write a sentence. With a specific subject. Particular detail. The witnessed thing, not the abstract one. Then the model gives me an image that earns its place above the words.
The images I have made along the way. The chip essay: a single hand pausing above a keyboard, the pause longer than the keystrokes. The mother essay: a kitchen at the end of the day, an empty bowl, the light from outside. The Roman legion essay: an empty road and a single figure who is not running anymore.
Each took dozens of prompts. Each of those prompts taught me what I had not yet said in the writing.
I keep a file of prompts. Not the failed ones. The ones that worked. The file is short. There are fewer than twenty entries after a year of work. Each entry is the prompt that made an image I kept. I open the file when I am stuck on a sentence. The vocabulary is in the file. The vocabulary is the sentence I have not yet written.
What making images taught me about writing, the model is the first reader#
Three months into prompting, my writing changed.
The change was structural. I had to know what I meant with a precision I had never needed before. The model is more honest than an editor. The editor will smooth over the fuzzy place. The model will draw the fuzzy place. The fuzziness becomes an image you can see. You cannot un-see it.
I started writing each essay with a question I had not used before. Could I prompt the image of this sentence? If the answer was no, the sentence was not yet specific enough. Back to the desk. Find the version that has an image inside it.
This is what the decision before the sentence feels like when the model is your second reader. The decision is no longer abstract. The decision has a consequence you can see in a few seconds. It is the most honest writing partner I have ever worked with, because it cannot pretend to understand you.
Every essay on this site now passes through two readers. The first is the model. The second is the person who arrives at the page. The first reader is rougher and more accurate. The first reader does not have manners.
There is a revision the model demands that I never asked of myself before. It does not let you leave a sentence vague. The image will show you the vagueness. You will fix the sentence. The next image will be closer.
The image is not the point. The image is the consequence of the sentence. The sentence is the work. The work happens before the image and is the reason the image works when it works. I had not been treating my own sentences this way for thirty years. The model treated them that way from the first prompt. I followed it.
Making images did not turn me into a designer. Making images turned me into a writer who knows what she means before she writes.
What the model still cannot do, the face and the room#
The model can do many things. There are images it cannot make.
It cannot make the face of a person I love. The face I want is not in the training set. The face I want is a face that exists once. The model gives me the average face that resembles her. I delete it.
It cannot make the specific room I sit in to write. The lamp in the corner. The light at four in the afternoon. The notebook open to the page where I crossed out three sentences before I kept one. The model gives me a room. It is not my room. I cannot get it to be.
It cannot make the thing I want the reader to feel before I tell her what to feel. The model is a tool of pictures. The image arrives. The reader sees the image. The reader does not feel what I felt at the desk while I was writing the prompt. That feeling does not transfer. It is also the thing the writing is for.
This is the part of the work where the model is not the part that decides whether the work gets through. The model is the part that arrives early. The part that lands is still the sentence.
Tomorrow I will write a new prompt. It will be too long or too short. The image will arrive. I will look at it. I will rewrite the sentence the prompt was for. I will write a new prompt. The image will be closer. Or it will not.
Either way I will be back at the desk. I will have learned something I could not have learned by reading.
If you want to learn this for your own work, reach out. The vocabulary I can teach you. The year is the part you have to spend.
The light is on in the room. The keyboard is warm. The next prompt is the next sentence. The next sentence is the next prompt.
Again.
