DALL-E Prompting

DALL-E Prompting

When to Use

Use this skill when:

• User explicitly asks to generate, create, or design an image using DALL-E 3, GPT-image-1, ChatGPT image generation, or any OpenAI image generation model
• User wants help writing, improving, or debugging a DALL-E prompt that is not producing the desired output
• User needs to iterate on an existing DALL-E result through revision dialogue -- they have a first image and want to refine it
• User asks for a prompt that will work specifically with OpenAI's ecosystem (e.g., "make a prompt for the ChatGPT image button" or "write this for the Images API")
• User wants to generate text reliably inside an image and is using OpenAI models
• User needs a sequence of related images with consistent visual elements using DALL-E
• User is building an application that calls the OpenAI Images API and needs prompt engineering guidance for programmatic generation

Do NOT use when:

• User wants Midjourney prompts with --ar, --v, --cref, --sref, or --style parameter syntax (use midjourney-prompting)
• User wants Stable Diffusion prompts with CFG scale, sampler selection, negative prompt weight syntax, or LoRA triggers (use stable-diffusion-prompting)
• User wants to convert or adapt an existing prompt from Midjourney or Stable Diffusion syntax to DALL-E natural language (use prompt-translation)
• User wants to maintain a character or face across a Midjourney series using --cref (use midjourney-consistency)
• User is asking about image editing, inpainting, or outpainting workflows within DALL-E specifically for the edit endpoint (use dalle-editing if available)
• User wants to generate video, audio, or animated content -- OpenAI image models produce static images only

Process

Step 1: Identify the Model and Understand Its Behavior

Before writing a single word of prompt, determine which model is in use or which to recommend. The behavioral differences are fundamental and change how the entire prompt is written.

• DALL-E 3 is accessed via ChatGPT's image button, the Images API with model: "dall-e-3", and third-party integrations. It internally rewrites ("revises") your submitted prompt before generation. This means subtle, figurative, or poetic language sometimes works well -- the model interprets intent. However, literal instruction-following is unreliable. DALL-E 3 tends to add compositional elements, background details, and artistic embellishments that were not requested.
• GPT-image-1 is the newer model accessed via the Images API with model: "gpt-image-1" or the ChatGPT canvas/image generation UI in newer deployments. It follows prompts more literally, handles multi-element compositions better, renders text inside images with significantly higher reliability, and responds well to layout and spatial instructions. It does not perform internal prompt rewriting to the same degree.
• To choose between them: if text-in-image is critical, if precise layout matters, or if the user needs literal interpretation -- recommend GPT-image-1. If the user wants stylized artistic output and is comfortable with interpretive variation -- DALL-E 3 is acceptable.
• Note the API size parameters that each model supports. DALL-E 3 supports: 1024x1024, 1792x1024, 1024x1792. GPT-image-1 supports: 1024x1024, 1536x1024, 1024x1536, and auto. The landscape and portrait proportions differ slightly between models.
• The quality parameter matters: DALL-E 3 accepts standard or hd. GPT-image-1 uses low, medium, high. For a polished output, always recommend hd (DALL-E 3) or high (GPT-image-1).

Step 2: Gather Requirements with Targeted Questions

Do not write a prompt without understanding these six dimensions. If the user's request is thin (e.g., "make an image of a mountain"), ask the targeted questions below rather than guessing. If the request is detailed, extract these from what the user has provided.

• Subject: What is the primary visual element? Include relevant attributes -- size, color, age range, material, state of action. A "dog" is underspecified; "a golden retriever puppy mid-leap, ears flying, wet fur" is not.