SEO Images — Premium AI Image Generation for Content

Generate publication-quality images for blog posts, newsletters, and social media using structured JSON prompts and Nano Banana 2 (Gemini 3.1 Flash Image).

No stock photo slop. No generic AI art. Images that match the quality of your writing.

Quick Start

# Generate a dark dashboard hero image
/seo-image "OpenClaw Meta Ads: $1,400/mo Saved" --style dark_data --ratio 16:9

# Generate a founder portrait
/seo-image "founder working late" --style founder_editorial --preset office_morning --ratio 4:5

# Generate a product shot
/seo-image "premium leather wallet on marble" --style product_lifestyle --ratio 1:1

# Generate a social share card
/seo-image "Ridge Wallet: $200M/Year" --style social_card --ratio 16:9

Setup

Option A: Replicate (recommended)

# Get your API token from https://replicate.com/account/api-tokens
export REPLICATE_API_TOKEN="r8_your_token_here"

Model:

google/nano-banana-2

Option B: Google AI Studio (free tier available)

# Get your API key from https://aistudio.google.com/apikey
export GOOGLE_AI_API_KEY="your_key_here"

Model:

gemini-3.1-flash-image-preview

https://generativelanguage.googleapis.com/v1beta/models/gemini-3.1-flash-image-preview:generateContent

Option C: fal.ai

export FAL_KEY="your_key_here"

Option D: Any provider with Nano Banana 2 access

The skill generates structured JSON prompts. The prompt is the valuable part — pipe it to whatever provider you prefer.

How It Works

1. You provide: Article context + style selection + optional customization
2. Skill builds: A cinema-grade structured JSON prompt with full camera/lighting/color specs
3. Skill flattens: JSON into an optimized text prompt for the model
4. Skill calls: Replicate API (or your preferred provider)
5. Skill outputs: 1-3 variations at your specified aspect ratio and resolution

The 6 Visual Styles

Read the full style definitions in

styles/

• Full JSON schema with every parameter documented
• Presets for common use cases
• Negative prompts to avoid common AI failures
• Vertical targeting (which industries/content types it works for)

1.

founder_editorial

— Cinematic People Photography

Hyper-realistic editorial portraits. Real people in real environments. Magazine-quality lighting and composition. Shot on named cameras with specific lenses.

Best for: About pages, founder stories, team pages, interview headers, newsletter author images

Presets:

rooftop_golden

office_morning

street_candid

2.

dark_data

— Premium Data Visualization Cards

Dark glass dashboard cards with bold stats, clean typography, and subtle depth. Rendered text that's actually legible. Think Bloomberg Terminal meets Stripe marketing.

Best for: Stats callouts, performance reports, case study headers, metric hero images, newsletter stats

Presets:

revenue_card

performance_card

3.

product_lifestyle

— Editorial Product Photography

Premium product shots in lifestyle contexts. The product is hero but the environment tells a story. Phase One quality.

Best for: Product launches, comparison posts, e-commerce features, DTC content

4.

social_card

— Scroll-Stopping Typography

High-impact text-forward cards for blog OG images, tweet cards, LinkedIn headers. Bold type, high contrast, designed to stop the scroll at thumbnail size.

Best for: Blog share images, tweet cards, LinkedIn post images, newsletter hero, chapter headers

5.

warm_lifestyle

— Bright Editorial Spaces

Warm, aspirational lifestyle photography. Natural light, earth tones, beautiful spaces. Film-stock quality.

Best for: Wellness content, real estate, creator economy, coaching, course promos

6.

cinematic_scene

— Movie-Still Narratives

Film-quality scenes with dramatic lighting and deliberate composition. Named director references for specific visual languages.

Best for: Long-form editorial, investigative pieces, dramatic feature stories, newsletter covers

Prompt Architecture

Every style uses this structured JSON format:

{
  "prompt": "the flattened text prompt sent to the model",
  "style_id": "which style template was used",
  "subject": { ... },
  "environment": { ... },
  "camera": {
    "camera_model": "specific camera body (e.g. Leica M11)",
    "lens_focal_length_mm": 50,
    "aperture": "f/1.4",
    "shot_type": "medium_close_up",
    "perspective": "eye_level",
    "depth_of_field": "shallow"
  },
  "lighting": {
    "key_light": "direction, quality, falloff description",
    "fill_light": "source and intensity",
    "rim_light": "position and effect",
    "color_temperature": "warm | neutral | cool | mixed"
  },
  "color_and_style": {
    "palette": ["#hex1", "#hex2", "#hex3"],
    "grade": "specific color grading description",
    "grain": "film grain type and intensity",
    "overall_style": "cultural reference (e.g. 'A24 meets Acne Studios')"
  },
  "output_settings": {
    "aspect_ratio": "16:9",
    "resolution": {"width": 1920, "height": 1080},
    "format": "png"
  },
  "negative_prompt": ["array of specific things to avoid"]
}

Why this matters: Generic prompts produce generic images. Specifying the exact camera model, aperture, film stock, lighting rig, and color grade gives Nano Banana 2 the detail it needs to produce images that look like they were shot by a real photographer with real equipment.

Generating Images

Via the generation script

# Set your API token
export REPLICATE_API_TOKEN="your_token"

# Generate from a style + context
bash scripts/generate.sh \
  --style dark_data \
  --context "Article about Ridge Wallet scaling to \$200M revenue" \
  --headline "Ridge Wallet: \$200M/Year" \
  --subtext "4-Voice Content System · 500+ AI Creatives/Day" \
  --ratio 16:9 \
  --resolution 2K \
  --output ./output/ridge-hero.png

# Generate with a preset
bash scripts/generate.sh \
  --style founder_editorial \
  --preset office_morning \
  --context "Newsletter author headshot" \
  --ratio 4:5 \
  --output ./output/author.png

# Generate a social card
bash scripts/generate.sh \
  --style social_card \
  --headline "The Agent That Never Sleeps" \
  --subtext "How AI Replaced 90 Minutes of Ads Manager" \
  --ratio 16:9 \
  --output ./output/og-image.png

Via the agent (OpenClaw / Claude Code / Codex)

The agent reads the article context, selects the appropriate style, builds the full JSON prompt, and calls the API. Example:

/seo-image for my article "OpenClaw Meta Ads Complete Guide" — I need a hero image and a social share card

The agent will:

1. Read the article to understand the content
2. Select

   dark_data

   for the hero (tech/data content)
3. Select

   social_card

   for the share image
4. Build full structured prompts for each
5. Generate via Replicate
6. Save both images to the specified output directory

Reference Images

For style consistency across articles, maintain reference images:

references/
  founder_editorial/   → 3-5 gold-standard portrait references
  dark_data/           → 3-5 reference dashboard cards
  product_lifestyle/   → 3-5 reference product shots
  social_card/         → 3-5 reference typography cards
  warm_lifestyle/      → 3-5 reference lifestyle shots
  cinematic_scene/     → 3-5 reference film stills

Pass reference images to the API via

image_input

curl -s -X POST "https://api.replicate.com/v1/models/google/nano-banana-2/predictions" \
  -H "Authorization: Bearer $REPLICATE_API_TOKEN" \
  -H "Content-Type: application/json" \
  -d '{
    "input": {
      "prompt": "your flattened prompt here",
      "image_input": ["https://url-to-reference-image.png"],
      "aspect_ratio": "16:9",
      "resolution": "2K",
      "output_format": "png"
    }
  }'

API Reference

Replicate —

google/nano-banana-2

# Create prediction
curl -s -X POST "https://api.replicate.com/v1/models/google/nano-banana-2/predictions" \
  -H "Authorization: Bearer $REPLICATE_API_TOKEN" \
  -H "Content-Type: application/json" \
  -d '{
    "input": {
      "prompt": "string — the full text prompt",
      "image_input": ["array of image URLs — up to 14 reference images"],
      "aspect_ratio": "16:9",
      "resolution": "1K | 2K | 4K",
      "output_format": "jpg | png",
      "google_search": false,
      "image_search": false
    }
  }'

# Poll for result
curl -s -H "Authorization: Bearer $REPLICATE_API_TOKEN" \
  "https://api.replicate.com/v1/predictions/{prediction_id}"

Parameters:

prompt

image_input

aspect_ratio

resolution

output_format

google_search

image_search

Google AI Studio —

gemini-3.1-flash-image-preview

curl -s -X POST \
  "https://generativelanguage.googleapis.com/v1beta/models/gemini-3.1-flash-image-preview:generateContent?key=$GOOGLE_AI_API_KEY" \
  -H "Content-Type: application/json" \
  -d '{
    "contents": [{"parts": [{"text": "your prompt here"}]}],
    "generationConfig": {
      "responseModalities": ["TEXT", "IMAGE"],
      "imageSizeOptions": {"aspectRatio": "16:9"}
    }
  }'

Prompt Flattening

The structured JSON is for YOUR reference and consistency. The model receives a flattened text version. The

scripts/generate.sh

JSON structured prompt → Flatten to narrative text → Send to model

Example:
{
  "subject": {"age": 32, "pose": "leaning on railing"},
  "camera": {"camera_model": "Leica M11", "aperture": "f/1.4"},
  "lighting": {"key_light": "warm sunset from camera-left"}
}

→ "hyper-realistic portrait of a 32-year-old man leaning casually on
   a railing, warm sunset light from camera-left with soft falloff,
   shot on Leica M11 at f/1.4, shallow depth of field..."

The structured format ensures consistency and makes it easy to tweak individual parameters without rewriting the entire prompt.

Tips

1. Camera specificity matters. "Shot on Leica M11 at f/1.4" produces dramatically different results than "professional photography."
2. Name the color grade. "Cinematic teal and orange" or "desaturated Portra 400 emulation" gives the model a clear reference point.
3. Negative prompts prevent common failures. Always include "no warped hands" for people shots and "no illegible text" for typography.
4. Reference images are your secret weapon. Feed 1-3 reference images via

   image_input

   to maintain style consistency across a content series.
5. Text rendering works now. Use quotes around text you want rendered:

   "GROWTH PLAYS"

   . Specify the font:

   "bold white Impact font"

   . NB2 handles this well.
6. Don't over-prompt. If the prompt exceeds ~500 words, the model starts losing focus. Keep the flattened version tight. The JSON is for your records.