How do you write prompts for DALL·E 3 inside ChatGPT?

Since October 19, ChatGPT Plus and Enterprise users can make images inside a normal chat. You describe what you want, ChatGPT passes a detailed version of your request to OpenAI's DALL·E 3 image model, and a selection of images comes back. You can then ask for changes in the same conversation. That changes how you write prompts. Instead of packing keywords into one line, you hold a conversation with something closer to a designer who takes notes.

A prompt here is simply your request for an image. With older image tools, people learned a kind of shorthand: strings of style words, camera terms and quality tags. DALL·E 3 is built to follow long, detailed descriptions closely, so ordinary sentences now work better than shorthand.

What a good first message contains

Think of it as a design brief. A useful first message answers six questions:

1. Purpose. Where will the image be used? A poster, a slide, a website banner?
2. Subject. What exactly is in it, and what is it doing?
3. Setting. Where is it, what time of day, what season?
4. Style. Photo, watercolour, flat illustration, pencil sketch? Describe qualities such as colours and line, not famous names: OpenAI built DALL·E 3 to turn down requests that copy the style of a living artist.
5. Shape. Wide (landscape) or tall (portrait)? DALL·E 3 supports both.
6. Text. Any words that must appear, in quotes. DALL·E 3 renders text much better than earlier image models, but always check the spelling.

A worked example: a community poster

Here is a first message for a real-world job:

I need a poster for our community garden's open house on Saturday morning.
Show a small vegetable garden in late summer with raised wooden beds, tomatoes,
sunflowers and a few neighbours of different ages chatting with coffee cups.
Style: friendly flat illustration, warm greens and yellows, simple shapes, no shading.
Format: tall, portrait orientation.
Leave clear space at the top for the words "Open House" in large letters.

ChatGPT returns a few versions. Now refine one thing per message:

I like the second one. Make the sky a soft morning blue and add a hand-painted
wooden sign by the gate that says "Jardin des Voisins".

Good. Keep everything the same, but make the people a little larger so faces
are visible from across a room.

Small, specific requests like these tend to keep what already works. If a result drifts, say what to keep: "Keep the layout and colours exactly as they are, only change the sign."

One more useful habit: ask ChatGPT to show you the exact prompt it sent for the version you like. Save that text. It is a detailed description you can reuse next year, or adapt for a flyer in the same style.

Limits worth knowing

DALL·E 3 inside ChatGPT runs safety checks on both your request and the images. OpenAI says it limits violent, adult and hateful content, reduces images of public figures, and declines living artists' styles. Text inside images is improved, not perfect, so read every letter before printing. For anything that will be shown publicly, also look closely at hands, faces and small details, which OpenAI says are better than before but are still where errors hide.

How is it different from prompting Midjourney?

Midjourney, the other popular image tool, works through the chat app Discord. You type the command /imagine followed by your prompt, and the image arrives in the channel. Version 5, introduced in March as an alpha test for paying subscribers, is known for very realistic, detailed images, and Midjourney promised it would respond more closely to the wording of prompts than earlier versions.

The prompting style is quite different:

In practice, Midjourney rewards people who learn its vocabulary and parameters, and many artists like the control that gives. DALL·E 3 in ChatGPT rewards people who can describe what they want clearly, which suits beginners and anyone working on a practical job like a poster. A good brief helps with both: the six questions above work as a checklist for a Midjourney prompt too, you just compress the answers into keywords and parameters.