/image-gen — On-Brand Image Generation

Describe what you need. Get an image that looks like your brand made it.

This skill reads your visual brand identity from brand/creative-kit.md, crafts a narrative prompt that bakes in your style constraints, and generates the image via Gemini API. No brand style defined yet? It still works — just at a lower enhancement level. Run /visual-style first for the best results.

---

On Activation

1. Check GEMINI_API_KEY environment variable.

   • If missing: "Image generation requires a Gemini API key. Set GEMINI_API_KEY in your environment. Get one at ai.google.dev."
   • Do not proceed without it.

2. Read brand files in priority order:

   • brand/creative-kit.md — look for ## Visual Brand Style section
   • brand/voice-profile.md — personality informs image tone
   • brand/positioning.md — angles inform visual metaphors
   • brand/landscape.md — Claims Blacklist (don't generate imagery that visually implies blacklisted claims)

3. Determine enhancement level:

Level	Context available	Image quality
L0	No brand files	Good generic images — uses discovery questions + prompt craft
L1	voice-profile.md	Personality-aligned — playful brand gets warm/bright images
L2	+ creative-kit.md colors/typography	Color-constrained — palette woven into prompts
L3	+ Visual Brand Style section	Fully on-brand — style anchors, lighting, mood, composition all applied

---

Phase 1: Discovery

Use AskUserQuestion. Ask one at a time. Skip questions the user already answered in their request.

Question 1: Purpose "What's this image for?"

• Blog header / article illustration
• Social media post (which platform?)
• Product shot / marketing asset
• Hero image / landing page
• Presentation / slide deck
• Something else (describe it)

This determines aspect ratio:

Use case	Default ratio	Resolution
Blog header	16:9	2K
Social square	1:1	2K
Social story	9:16	2K
Hero / banner	21:9	2K
Product shot	4:3	4K
Thumbnail	16:9	1K

Question 2: Feeling "What should someone feel when they see this?" (Free text — this drives the prompt's emotional anchor)

Question 3: Style override (only if L3 brand style exists) "Use your on-brand style, or something different?"

• On-brand (default) — applies Visual Brand Style from creative-kit.md
• Different — describe the style you want instead

If user picks "different," their description overrides the brand style for this image only.

Skip discovery when: The user's request is specific enough. "Generate a 16:9 blog header showing a glowing terminal on a dark background, warm rim lighting" — don't ask what they want, they just told you.

---

Phase 2: Prompt Crafting (Nano Banana Prompt Engineering)

You are an expert Nano Banana prompt engineer. Your job is to turn the user's brief into a single, high-quality prompt for Nano Banana 2 (or Pro), a "thinking" image model used for professional asset production.

Core principle: brief a senior art director, don't list keywords. Write natural language in full sentences. Never use "tag soup" like "dog, park, 4k, realistic."

The 10 Rules

1. General style. Be specific and descriptive about subject, setting, composition, camera/viewpoint, lighting, mood, materials, and textures. Full sentences, not comma lists.

2. Context and purpose. Always encode the purpose and audience (YouTube thumbnail, app icon, hero banner, tweet graphic, 4K wallpaper). Let purpose guide style, polish level, and framing.

3. Text and infographics. If text must appear, put it clearly in quotes in the prompt. Ask for legible, clean typography and specify style (bold sans-serif, monospace, handwritten). For data, ask the model to compress into infographics, diagrams, or whiteboards.

4. Character and brand consistency. When reference images exist, explicitly refer to them: "Keep the person's facial features exactly the same as Image 1." Allow changes in pose, expression, angle while preserving identity.

5. Grounding and realism. For real data, locations, or products, tell the model to rely on up-to-date factual knowledge. Encourage coherent details consistent with physics.

6. Editing and restoration. For edits to existing images, give semantic instructions: "remove," "replace," "add," "restore," "change the season." Maintain original structure, only change what's intended.

7. Dimensional and structural control. For floor plans, schematics, wireframes, grids, tell the model to follow that layout closely. For 2D↔3D, describe how the new representation should look while preserving key relationships.

8. Resolution, detail, and format. Specify resolution ("high detail suitable for 4K wallpaper," "clean 16:9 thumbnail"). Call out micro details and textures when needed (brushed steel, cracked paint, mossy stone).

9. Narrative and sequences. For multiple images, describe the story arc, emotional beats, what stays consistent across images. Specify count, format, and identity/style consistency.

10. Output rules. Do not ask follow-up questions about the prompt. Resolve small ambiguities with sensible professional defaults. Output a single flowing narrative prompt.

Prompt Structure

Build the narrative in this order, woven into flowing prose:

1. Purpose and format — what this is for, aspect ratio, resolution
2. Scene — what's happening, where, environment
3. Subject — detailed description with textures, materials, poses
4. Composition — framing, focal point, depth of field, negative space
5. Lighting — source, quality, color temperature, interaction with materials
6. Mood — emotional tone, atmosphere
7. Text — any text in quotes with typography specification
8. Technical — camera/lens for photorealistic, style reference for illustrated

Brand Constraints (L3)

When Visual Brand Style exists, weave constraints INTO the narrative — don't add as a separate block:

• Primary Aesthetic → sets overall style direction
• Lighting → overrides generic lighting with brand-specific lighting
• Backgrounds → constrains background treatment
• Composition → constrains layout/framing
• Mood → anchors emotional tone
• Avoid → explicit exclusions baked into prompt
• Reference Prompts → use as structural templates, adapting subject matter

See references/prompt-patterns.md for proven patterns. See references/visual-metaphors.md for concept-to-metaphor mapping.

---

Phase 3: Generate

Gemini API Call

import os
from google import genai
from google.genai import types

client = genai.Client(api_key=os.environ["GEMINI_API_KEY"])

response = client.models.generate_content(
    model="gemini-3.1-flash-image-preview",
    contents=["<narrative prompt>"],
    config=types.GenerateContentConfig(
        response_modalities=['TEXT', 'IMAGE'],
        image_config=types.ImageConfig(
            aspect_ratio="<ratio>",
            image_size="<resolution>",
        ),
    ),
)

for part in response.parts:
    if part.text:
        print(part.text)
    elif part.inline_data:
        image = part.as_image()
        image.save("output.png")

Default to gemini-3.1-flash-image-preview (Nano Banana 2). Launched Feb 26, 2026. Pro quality at Flash speed/pricing. 4K support, up to 14 reference images for style consistency. Use gemini-3-pro-image-preview (Nano Banana Pro) only when text-heavy infographics or premium quality justify the higher cost.

Save and Log

1. Save image to project directory (e.g., images/, assets/, or wherever the project keeps visuals)
2. Append to brand/assets.md:

   | <date> | image | <file-path> | image-gen | <1-line description of what was generated> |
   

---

Phase 4: Iterate

After generating, show the image and offer:

"Image generated. Want me to:"

1. Adjust the lighting or mood
2. Change the composition or framing
3. Try a completely different approach
4. Generate more variations
5. Ship it

For adjustments, use Gemini's multi-turn chat API for iterative refinement — pass the previous image + adjustment prompt.

chat = client.chats.create(model="gemini-3-pro-image-preview")
response = chat.send_message("Make the lighting warmer and add more negative space on the left for text")

---

Image Editing

When the user provides an existing image to modify:

from PIL import Image

img = Image.open("input.png")
response = client.models.generate_content(
    model="gemini-3-pro-image-preview",
    contents=["<edit instruction>", img],
    config=types.GenerateContentConfig(
        response_modalities=['TEXT', 'IMAGE'],
    ),
)

Supports: style transfer, background replacement, object removal, color grading, text overlay.

---

Claims Blacklist Awareness

If brand/landscape.md exists, check the Claims Blacklist before generating:

• Don't generate imagery that visually implies "the only X" if that claim is blacklisted
• Don't generate competitor comparisons that misrepresent the competitive landscape
• When in doubt, describe the concept to the user and ask if it implies a blacklisted claim

---

Anti-Patterns

Anti-pattern	Why it fails	Instead
Listing comma-separated keywords	"professional, modern, tech, blue, minimal" produces generic AI art	Write a narrative: "A single glowing terminal on a dark slate desk, soft amber rim lighting catching the screen edges"
Ignoring aspect ratio	Generating 1:1 for a blog header means the user has to crop, losing composition	Always ask where the image goes FIRST — ratio determines everything
Overriding brand style silently	User defined a visual brand identity for consistency — ignoring it defeats the purpose	Default to brand style. Only override when user explicitly asks for something different
Generating without purpose	"A nice image" produces nothing distinctive	Every image needs a purpose (blog header, social, hero) and a feeling (trust, excitement, calm)
Saving as JPEG when PNG needed	Gemini returns JPEG by default — transparency and quality loss	Always convert to PNG with image.save("output.png", format="PNG")
Skipping the assets log	Future agents won't know what images exist, leading to duplicate generation	Always append to brand/assets.md after generating

---

Related Skills

• /visual-style — builds the Visual Brand Style section this skill reads
• /creative — produces multi-mode creative briefs (not pixels). Route here for ad concepts and video scripts
• /paper-marketing — designs in Paper MCP. Route here for carousels, slideshows, and social graphics designed as HTML
• /app-store-screenshots — specialized screenshot generator for App Store listings