How to automate image prompt writing using AI and five simple inputs

Executive overview

Writing effective prompts for AI image models like Imagen 3 requires rich context, not keyword stuffing. Most people waste time tweaking prompts manually when AI can do it better.

Give an AI five structured inputs and it generates three ready-to-paste prompts in JSON format — optimised for Imagen 3. Save this as a reusable project in ChatGPT and your whole team gets consistent, high-quality image outputs without prompt engineering.

You don't need to know how to talk to image models — outsource that job to a text AI.

The five inputs

1. Purpose — why the image is being created (Instagram ad, YouTube thumbnail, hero section, pitch deck). Determines aspect ratio and visual conventions.
2. Audience — who the image is for (luxury buyers, tech executives, Gen Z consumers). Shapes tone and style.
3. Subject — what's in the image (product, person, character). The primary focus.
4. Brand rules — colours, fonts, style guidelines. Optional attachment if a brand guide exists.
5. Reference image — an existing photo, product shot, sketch, or competitor ad to anchor the output. Optional but commonly used.

How the system prompt was built

• Feed a Google blog post on Imagen 3 prompting tips to Claude and ask for the 80/20 distilled into five inputs.
• Ask Claude to synthesise that blog post into a full system prompt that accepts those five inputs and outputs three JSON-formatted prompts.
• The system prompt includes a guard: if any of the five inputs are missing, the AI asks for them before doing anything.
• Output structure: three variants — literal, creative, and premium — each as a JSON block ready to copy-paste into Imagen 3.
• JSON format produces better image outputs than plain text prompts in Imagen 3.

Using the system on a recurring basis

• Save the system prompt into a ChatGPT project so it's always active — no repasting required.
• Add business context and default inputs at the top of the project so the AI already knows your product, audience, and brand.
• Use the highest-reasoning model available (extended thinking on) for the best prompt quality.
• Paste the copied JSON prompt plus a reference image directly into Imagen 3.

Editing images without restarting

• Imagen 3 handles minor in-context edits better than other models — no need to rewrite the full prompt.
• Describe edits in plain language: "make the background warmer", "zoom out and place in a coffee shop", "move text to the left".
• Only reroll the full prompt if the output is fundamentally wrong.
• Remove watermarks in Canva using the Magic Eraser tool.