image-prompts

name	image-prompts
description	AI image prompt engineering for BFL Flux models (Flux 1.1 Pro via Replicate, Flux 2 via BFL API) and Ideogram v2. Covers prompt formula, model selection, hex color precision, no-negative-prompt patterns, JSON structured prompting, brand DNA integration. Adapted for Superdots brand.

Image Prompts — AI Image Generation Creative Director

Core Principle

Act as a Creative Director that constructs optimized prompts for image generation. Never pass raw text directly to the image generation script. Always interpret, enhance, and construct an optimized prompt using the 5-Component Formula from references/prompt-engineering.md.

Superdots Brand Constraints (MANDATORY)

These constraints override any domain mode defaults:

• Color palette: warm tones (reds, ambers, teals) on deep navy backgrounds
• Style: abstract geometric — NOT photorealistic, NOT stock photo
• Banned subjects: robots, AI clichés, generic tech imagery
• Allowed domain modes: Editorial, Abstract, UI/Web ONLY
• Image generation: use scripts/generate-ai-images.mjs --slug <slug> for hero images
• Social images: use scripts/tools/generate-social-image.mjs --platform [linkedin|facebook]
• Image differentiation: when generating multiple images in a batch, each MUST have visually distinct composition (vary perspective, color emphasis, subject)

The 5-Component Prompt Formula

Before constructing ANY prompt, read references/prompt-engineering.md for full details.

Quick Reference

1. SUBJECT — Who/what is the primary focus. Be specific: physical characteristics, material, texture. Never "a person" or "an abstract shape."
2. ACTION — What the subject is doing or its visual state. Use strong present-tense verbs.
3. LOCATION/CONTEXT — Where the scene takes place. Environmental details, atmosphere.
4. COMPOSITION — Camera perspective, framing, spatial relationships. "tight close-up", "bird's-eye view", "medium shot centered."
5. STYLE — Visual register, aesthetic, lighting. Reference real cameras, art movements, publications.

Critical Prompt Rules

• Name real cameras for photorealistic looks: "Sony A7R IV", "Canon EOS R5"
• Include micro-details: textures, reflections, material qualities
• Use prestigious context anchors for quality: "National Geographic cover", "Kinfolk editorial"
• NEVER use: "8K", "masterpiece", "ultra-realistic", "high resolution"
• NEVER describe the concept — describe the SCENE
• For critical constraints use ALL CAPS: "MUST contain exactly three elements"
• Write as natural narrative paragraphs, NOT comma-separated keyword lists

Domain Modes (Superdots-approved only)

Editorial Mode

• Publication refs: Kinfolk, Monocle, Wired, MIT Technology Review
• Styling: clean, minimal, warm tones, geometric elements
• Composition: intentional whitespace, rule of thirds, strong focal point

Abstract Mode

• Forms: geometric patterns, flowing gradients, data-inspired shapes
• Colors: Superdots palette (amber #F59E0B, teal #14B8A6, navy #0F172A, red #EF4444)
• Techniques: overlapping shapes, depth layers, motion blur, light trails

UI/Web Mode

• Styles: flat vector, isometric 3D, line art, glassmorphism
• Backgrounds: deep navy or transparent
• Sizing: design at 2x for retina

Aspect Ratios

Use Case	Ratio	Dimensions
Blog hero image	16:9	1920×1080
LinkedIn landscape	1.91:1	1200×630
Facebook square	1:1	1080×1080
Story/Reel	9:16	1080×1920
Instagram portrait	4:5	1080×1350
Pinterest pin	2:3	1000×1500

Prompt Templates

Abstract/Geometric (primary for Superdots)

A [geometric composition] of [abstract elements with specific shapes/forms],
featuring [color palette from Superdots brand] with [texture/material quality].
[Depth and layering description]. [Lighting: warm directional light creating
shadows and depth]. [Mood: sophisticated, modern, tech-forward without clichés].

Editorial/Conceptual

A [editorial-style composition] showing [conceptual subject related to article topic],
[arrangement and spatial relationship]. [Warm tones: amber highlights, teal accents
on deep navy]. [Lighting setup]. In the style of [publication reference].

Post-Processing

After generation, if needed:

# Crop to exact dimensions
magick input.png -resize 1200x630^ -gravity center -extent 1200x630 output.png

# Convert format
magick input.png output.webp

# Resize for specific platform
magick input.png -resize 1080x1080 instagram.png

Reference Documentation

Load on-demand — do NOT load all at startup:

• references/prompt-engineering.md — Domain mode details, modifier libraries, advanced techniques
• references/post-processing.md — ImageMagick pipeline recipes

Quality Checklist

Before submitting any generated image:

• No text artifacts or garbled characters
• Specific to the article topic (not generic)
• Matches Superdots brand (warm tones, abstract/geometric, no robots)
• Correct aspect ratio for intended platform
• Visually distinct from other images in the same batch

---

Flux Model Mastery (BFL)

Critical Flux Rules

1. NO negative prompts — Flux does not support them. Describe what you WANT, not what to avoid. See "Positive Alternatives" below.
2. Prose over keywords — Write descriptive narrative paragraphs. "woman, portrait, beautiful, 8k" produces worse results than a flowing description.
3. Lighting is king — Lighting has the single biggest impact on output quality. Always specify it.
4. Word order matters — Flux prioritizes elements that appear earlier in the prompt. Front-load the most important subject.
5. Hex colors work — Use #RRGGBB (color name) for precise brand color matching. Always pair hex with a human-readable name.
6. Optimal length — 30-80 words. Too short = generic. Too long = unfocused.

Prompt Structure Formula

[Subject] + [Action/Pose] + [Style/Medium] + [Context/Setting] + [Lighting] + [Camera/Technical]

Example for Superdots:

Bold geometric composition of overlapping angular planes in #E8363B (brand red)
and #F59E0B (warm amber), arranged with confident layered depth on deep navy
background #0B1222. Subtle dot-grid texture visible in darker areas. Warm
radial glow from bottom-left edge. Flat design with generous negative space.
Centered focal point filling 60% of frame. No text, no letters.

Model Selection

We use Flux 1.1 Pro via Replicate (default). For future BFL API direct access:

Model	Best for	Speed	Cost
Flux 1.1 Pro	Production social images (our default)	Medium	$0.04/img
FLUX.2 [klein] 4B	Fast drafts, previews, iteration	Fastest	$0.014/img
FLUX.2 [pro]	Production balanced (upgrade path)	Medium	$0.03/MP
FLUX.2 [max]	Maximum quality, complex compositions	Slowest	$0.07/MP
FLUX.2 [flex]	Typography rendering (text-on-image)	Medium	$0.05/MP

Current setup: REPLICATE_API_TOKEN → Flux 1.1 Pro. If BFL_API_KEY is set, consider using FLUX.2 [pro] or [max] for better results.

Flux Parameters (via Replicate)

Parameter	Default	Range	Effect
guidance_scale	3.5	1.5-5.0	Prompt adherence. 2.0-3.0: more creative/atmospheric. 3.5-4.0: balanced (our default). 4.5-5.0: strict literal interpretation.
num_inference_steps	28	20-50	Quality vs speed. 20-25: fast drafts. 28: production default. 35-50: maximum detail (diminishing returns past 35).
output_format	webp	webp/png/jpg	Always webp for social (best quality/size). Use png only for transparency needs.
output_quality	85	1-100	WebP quality. 85: our standard (good balance). 95: when fine detail matters.
seed	random	0-2^32	Reproducibility. Set a seed to regenerate the same image. Useful for A/B variants — keep seed, change one prompt element.

Hex Color Precision

Flux accurately renders hex colors. Always pair with a descriptive name:

Good: dominant accent in #E8363B (brand red), background #0B1222 (deep navy)
Bad:  #E8363B background #0B1222

For Superdots brand palette:

#0B1222 (deep navy) — background, always
#E8363B (brand red) — primary accent for marketing, legal, design, strategy
#14B8A6 (teal) — primary accent for engineering, operations, hr, technology; also used for secondary UI elements (badges, links) across all departments
#F59E0B (warm amber) — primary accent for sales, finance, customer-support

Limit to 3-5 colors per prompt. Associate each color with a specific object or area. The department accent color from image-style-config.json is the authoritative source — always check it.

Positive Alternatives (No-Negative-Prompt Patterns)

Instead of	Use
"no text, no letters"	"clean unmarked surfaces, text-free composition"
"no people"	"empty, solitary, deserted"
"not photorealistic"	"abstract geometric illustration, flat design"
"no blue"	"warm tones: red, amber on deep navy"
"not cluttered"	"minimal, generous negative space, clean composition"
"no robots"	"abstract shapes, human-centered geometric forms"

JSON Structured Prompting (Advanced)

For complex compositions, use JSON to plan then flatten to prose:

{
  "scene": {
    "setting": "abstract brand composition",
    "mood": "confident, bold, modern"
  },
  "subjects": [{
    "type": "geometric",
    "description": "overlapping angular planes with layered depth",
    "position": "center, filling 60% of frame"
  }],
  "colors": {
    "primary": "#E8363B (brand red)",
    "background": "#0B1222 (deep navy)",
    "accent": "#F59E0B (warm amber)"
  },
  "technical": {
    "texture": "subtle dot-grid pattern in background",
    "lighting": "warm radial glow from bottom-left",
    "composition": "centered, generous negative space"
  }
}

Flatten to natural language before sending to the model.

Superdots Brand DNA in Every Prompt

Every generated image MUST include these brand signatures in the prompt:

1. Deep navy background #0B1222 — never pure black, never gray, never blue
2. Dot-grid texture — subtle radial dot pattern (1.5px dots, 18px spacing) visible in background. Use: "subtle dot-grid texture overlay"
3. Warm red glow — soft radial gradient from bottom-left, brand red at 12% opacity. Use: "warm radial glow from bottom-left edge"
4. Bold geometric shapes — confident, layered, with clear depth hierarchy
5. Generous negative space — compositions breathe, not cluttered. Use: "minimal, generous negative space"
6. Department accent as PRIMARY color — not a subtle hint, the dominant visual element
7. No text, letters, logos — ever. If text artifacts appear, regenerate immediately. Use: "clean unmarked surfaces, text-free composition"