Artificial Lexicon
LLM guide

Using ChatGPT to write image prompts

A chat model knows the vocabulary of art. It read the same books the walls came from. This guide turns it into your prompt writer.

Free5 min read

Why this works

Midjourney and Stable Diffusion respond to keywords. ChatGPT, Claude, and Deepseek respond to requests. Put them together and you get a loop: you describe the picture in plain words, the chat model turns it into a keyword stack, the image model paints it.

ChatGPT and Gemini can also render the image themselves now; the chat image models guide covers that path. This guide is for when the keyword stack itself is the product. The chat model is not magic. It writes vague prompts if you make vague requests. The fix is the same one the walls teach: name the subject, the style, the light, and the frame. Give it slots to fill.

Give it the recipe, not the wish

"Write me a beautiful prompt about autumn" gets you mush. Hand it a template instead. The template below is the same anatomy the Midjourney page teaches: subject, style, mood, details, parameters.

recipe requestWrite 3 Midjourney prompts using this structure: [subject doing something], [one art style], [one lighting term], [2-3 concrete details] --ar 3:2 Subject: a lighthouse keeper on a stormy coast Keep each prompt under 40 words. No adjectives like "beautiful" or "stunning". Concrete nouns only.

The last two lines do the real work. Word caps stop rambling. Banning filler adjectives forces concrete choices.

The prompt builder runs this recipe on the site: one slot per wall, dice for the empty ones. Fill it by hand, then paste the result here for the model to polish.

Feed it wall words

The model picks better words when you give it a shortlist. Copy a category from a wall and paste it in. The lighting wall and the styles wall work best for this.

shortlist requestHere are lighting terms: rim lighting, Rembrandt lighting, god rays, dappled light, blue hour, sodium vapor lamps. Write one prompt for a night market scene using exactly two of these. Pick the two that fit a night market and tell me in one line why.

Asking why forces a choice instead of a shuffle. You also learn which words the model thinks belong together, which is half the craft.

Ask for variations, one axis at a time

The walls teach one change at a time. The chat model can run that experiment for you in a single message.

variation requestTake this prompt: "a tea house in the mountains, ukiyo-e, golden hour lighting, paper lanterns, mist --ar 3:2" Write 5 versions. Change ONLY the art style each time. Keep every other word identical. Use: watercolor, art deco, low poly, gouache, linocut print.

Generate all five and compare. You now know what each style word does to your subject, and it cost one message.

Make it critique your prompt

The model reads prompts better than it writes them. Paste your draft and ask what is vague.

critique requestHere is my image prompt: "a warrior in a forest, epic, cinematic, highly detailed, 8k" Point out every word that carries no visual information. Suggest one concrete replacement for each. Do not rewrite the whole prompt.

"Epic" and "highly detailed" will be the first against the wall. Good. Those words spend tokens and buy nothing.

Keep a prompt notebook

When a prompt works, save it. The chat model formats your keepers into a reusable library.

notebook requestFormat these 3 prompts as a table with columns: subject, style words, lighting words, parameters. Then list which words appear more than once. Those are my habits.

The repeat list is the honest part. It shows the rut you are in, and the walls are how you climb out of it.

💡 The loop in one line: recipe in, keywords out, image made, critique back, revise. Each pass through the loop is one lesson.

Published here first.

Keep going