An OG image generator
Every page needs a share picture, and one render from one model tells you nothing. So I asked my agent for a tool: a style, an idea, four models at once, my brand on every result. This video came out of it too.
- Where
- In the terminal, with Claude Code
- Size
- A longer one
- Short link
- piy.show/s106
Every page needs a picture
An OG image is the picture a link shows when somebody shares your page. Every episode, every guest, every article needs one, and I had none I liked. A picture from one model is a coin flip. I did not want to generate one picture. I wanted to compare sixteen.
So in August 2026 I asked my agent for a tool, in the terminal on my own machine. A style, an idea, several models at once, and my brand on every result in the same place. The first version ran the same day.
That is the whole method of this Short. Do not ask for the picture. Ask for the tool that makes picking easy, and then pick.
What is an OG image, and why does a page need one? Prompt it yourself
Build me a tool that takes a style and an idea, renders it through four image models at once, and puts my logo and my title on every result the same way.
Two axes: the look and the idea
A style is a look: the medium, the light, the palette, the camera, how a real face is treated. It says nothing about what happens. An idea is what happens. It says nothing about the look. The tool keeps them apart and joins them only at the moment it sends the prompt.
Why does that matter? Found a look you like? Keep it and feed it new ideas forever. Got an idea and no look? Ask the agent to invent five styles from it. Not sure? Run the idea through every style and look at them side by side.
Separate what changes from what stays. That rule works for pictures, and for most prompts.
How do I describe a style without the scene? Prompt it yourself
Invent five visual styles for this idea. A style describes only the look, the medium, the light, the palette and the camera. It never says what happens in the scene.
Sixteen pictures, one page, one decision
The grid is ideas by styles by models by tries. The first real run was four styles on four models: sixteen branded cards on one page, under two minutes, every one a success, ninety cents. One idea, sixteen answers, and the right direction was obvious in a second.
The four models are chosen because they fail differently. One holds a likeness, one follows instructions, one reads a scene's layout best, one comes from another aesthetic family. Four of the same kind would agree with each other and teach me nothing.
Before the run, a dry run prints the grid and the total price. The agent spends nothing until I have seen the number.
Why run four models instead of one? Prompt it yourself
Run this idea through all four styles on all four models. Put the sixteen results on one page with the cost of each, and show me the total price before you spend anything.
The logo is drawn by code, not by the model
The model makes the artwork. The brand band, my logo, the headline and the red rule, is drawn by code on top. A model cannot draw the same logo twice, and the text it renders is a lottery. Code draws it the same on every picture, in a blink, for nothing.
Two things follow. Every prompt ends with an order: render no text, and keep the bottom fifth of the frame calm, because that is where the band lands. And a card I already like can get a new title for free, with no model involved.
The lesson is the same one behind my film remix: let the model make the picture once, and let code do the part that must be exact.
Why not let the model draw the logo? Prompt it yourself
Never let the model draw my logo or my title. Render the art with no text and the bottom fifth kept calm, then put the brand band on by code, the same on every picture.
I can help you build yours
I build things like this with an AI agent, on my own computer. In a private session I help you do the same with your idea. Or bring it to the show, free.
Say who you are in the picture
I wanted my own face on the cards, so the tool takes reference photos. The first scene had two people in it, an interviewer and a candidate. One model cast me as the interviewer. Another put me on the monitor as the candidate, the exact opposite of the point. Nothing failed. The picture was wrong.
The fix is one sentence in the prompt: in this scene I am the interviewer, and everyone else is somebody else. A model given a reference and a crowd has to guess who the reference is. Tell it.
For your own pictures: name your role in every scene with more than one person.
How do I get my own face into the picture? Prompt it yourself
Here are three photos of me. In this scene I am the interviewer. Everyone else in the frame is somebody else.
Measure before you pay more
I asked two questions I thought I knew the answers to. Is it using old models? And would the dearer quality setting look better? The agent did not guess. It read the run records: every call had gone to the newest model, at the middle quality tier.
Then it ran a real test. The same four ideas at the high tier and at the middle one, side by side. The difference was marginal, at almost three times the price. Verdict: stay on the middle tier and spend the difference on more tries. It also sent the same prompt straight to the model's maker, with a real bill: the same picture, more money.
Ask your agent to measure instead of assume, and to show you the two pictures and the two prices. A measured no saves more than a guessed yes.
Does the dearer quality tier show? Prompt it yourself
Before you tell me a model is old or a setting is better, run the same prompt both ways. Show me the two pictures and the two prices.
This video came out of the same tool
The Short you watched is four pictures from the tool, one per era: 1985, 2003, 2021, 2026. Four seconds each, one line of text per picture, drawn by code in the brand's own type. The music came from a music model, after six options. The end card is the same band the cards wear.
Behind the scenes, there was more. After the terminal version, I had a web studio built: projects, contact sheets, picks, layouts, a login. It took two weeks. Two weeks later I deleted it, with everything behind it. The terminal was enough.
That is the honest part of this story. Build the small thing first, and let the habit decide whether the big thing is needed.
Why did I delete the web version? Prompt it yourself
Make a short story video from these four pictures. Four seconds each, one line of text per picture drawn by code, music under it, and my end card at the end.
Start with your next share picture
Take one page you want people to share. Write what happens in the picture, in one line. Ask for four looks and four pictures each, on one page, with the price first. Pick one. Put your own logo on by code.
What do I do next? Prompt it yourself
Here is my page title and what it is about. Give me four looks and four pictures each, on one page, with the cost before you start.
Do it in the terminal
I do this with Claude Code, the AI agent in the terminal. It works in a folder on your own computer, so it reads your files every time and can look things up for you.
A web chat works too, but there you paste everything in again in every new chat.
-
Get Claude Code
It comes with Claude Pro, US$20 a month, paid to Anthropic. Claude's plans
-
Make a folder for it, and open it
mkdir -p ~/projects/my-share-pictures && cd ~/projects/my-share-pictures
-
Start the agent
claude
-
Ask it
I need a share picture for every page on my site. Build me a tool that takes a style and an idea, renders it through several image models at once, and puts my logo and my title on every result the same way.
Never opened a terminal? In a private session I put Claude Code on your computer with you, and you start your first idea.
Prompty