Free Guide

My AI Image Prompts Actually Work Now

(Steal This)

I've been generating a lot of images lately. Hero graphics for articles, LinkedIn scroll-stoppers, slide deck visuals - the works. After months of experimentation, I landed on a system that consistently delivers. Not "pretty good" results. Predictable, brand-aligned, on-message results.

The Problem With Most AI Image Prompts

Most people approach AI image generation like a conversation:

"Make me an image of a leader looking confident but approachable, maybe in an office setting, professional but not boring..."

This doesn't work. Rambling prose triggers inconsistent output. Models like Gemini do far better with structured data than meandering descriptions.

The secret? JSON prompts.

The 6-Key Structure

Every image I generate uses this exact JSON structure. Six keys that trigger the model's best internal reasoning:

{
  "purpose": "What is this image for?",
  "aspect_ratio": "Platform-specific dimensions",
  "audience": "Who is viewing this?",
  "subject": "The core visual subject - be CONCRETE",
  "brand_rules": "Your non-negotiable style guide",
  "reference_image": "Optional visual reference, or null"
}

Why Each Key Matters

Purpose tells the model what job the image is doing. "Substack Header" generates differently than "LinkedIn Scroll-Stopper." Obvious once you see it.

Aspect Ratio is crucial for platform fit:

14:10 for Substack hero images
4:3 for LinkedIn posts
16:9 for presentation slides

Audience changes everything. "Skeptical executives" yields different imagery than "junior developers." Think about who you're reaching before you describe what you want.

Subject - be painfully concrete. Not "leadership" but "A single figure standing confidently at a podium, spotlight illuminating their face." The model can't read your mind. Give it specifics.

Brand Rules are your visual guardrails. Anti-patterns too - what you never want to see. Document both.

Reference Image is optional, but useful when you have a specific composition in mind.

A Real Example

Here's an actual prompt I used for an article about "Zombie Scrum":

{
  "purpose": "Article Hero Image",
  "aspect_ratio": "14:10",
  "audience": "Frustrated Scrum teams and change agents",
  "subject": "A group of 'Zombie' LEGO minifigures (pale skin, vacant expressions, tattered business suits) shuffling continuously around a kanban board that has no exit column. One 'alive' minifigure in bright colors looks at the camera with concern.",
  "brand_rules": "Cinematic lighting, spooky but professional atmosphere, high detail, photorealistic texture.",
  "reference_image": null
}

Notice the specificity. Not just "zombies at a meeting" - pale skin, vacant expressions, tattered suits, and a kanban board with no exit column. That last detail tells the whole story.

Make It Your Own: 5 Customization Tips

Define YOUR Brand Rules

My brand rules emphasize photorealistic LEGO minifigures with cinematic lighting. Yours might be:

Minimalist line art with bold accent colors
Watercolor aesthetic with warm, organic tones
Retro-futuristic with neon gradients, geometric shapes
Flat vector illustrations with playful proportions

Whatever you choose, document your anti-patterns too. I explicitly ban "generic stock photo styles" and "uncritiqued teamwork hands."

Build a Mood Vocabulary

Create a set of go-to mood descriptors:

Professional, confident, forward-looking
Warm and inviting, approachable
High-contrast, bold, attention-grabbing
Contemplative, introspective, thoughtful

Having these pre-defined prevents the "uh... make it look good?" trap. You've been there.

Create Platform Presets

Don't think about dimensions every time. Build templates:

For Substack:

aspect_ratio: "14:10"

For LinkedIn:

aspect_ratio: "4:3"

Use Accent Colors With Meaning

I assign colors to concepts:

Warm gold = success

Cool blue = data

Soft red = tension

Deep green = growth

Visual consistency across your body of work. People start recognizing your content before they read your name.

Save Your Best Prompts

When something nails it, save it. I keep a reference file organized by theme:

Leadership/credibility images
Data/metrics visualizations
Collaboration/team dynamics
Conflict/tension moments

These become starting points. Don't reinvent the wheel every time.

The Full Template

Here's the complete system prompt I use. Steal it, fork it, make it weird:

image-prompt-generator.md

---
name: image-prompt-generator
description: Generate structured JSON prompts for AI image generation
---

## Brand Visual Guidelines

- **Style:** [Your signature aesthetic]
- **Mood:** [Your emotional tone]
- **Lighting:** [Your preferred lighting style]
- **Palette:** [Your color system with meanings]
- **Anti-Patterns:** [What you NEVER want to see]

## Output Structure

For each image, generate:

1. **Substack Hero** (14:10) - Main visual anchor
2. **LinkedIn Teaser** (4:3) - High contrast, attention-grabbing
3. **Supporting Images** (Optional) - Data/concept visualizations

## JSON Template

```json
{
  "purpose": "",
  "aspect_ratio": "",
  "audience": "",
  "subject": "",
  "brand_rules": "",
  "reference_image": null
}
```

Try It This Week

Pick an article you're working on. Before you start hunting for stock photos or sketching ideas, write a JSON prompt using this structure. The best prompts aren't the most creative. They're the most specific.

P.S. - If you're using a different image model (DALL-E, Midjourney, Stable Diffusion), this structure still works. The key is giving the model structured constraints rather than rambling instructions.