| name | nano-banana-skill |
| description | Nano Banana image generation for marketing assets. Artist-level visual direction
for brand graphics, thumbnails, Substack covers, X post backgrounds, and LinkedIn
headers. Covers prompt architecture, lighting direction, composition rules,
style consistency, and what to never generate. Stops AI from producing generic
Midjourney-looking stock imagery.
|
Nano Banana Skill
The Artist Mindset
Nano Banana is not a prompt box. It is a collaborator. The quality of output is entirely determined by the quality of direction. Vague prompts produce vague images. Specific direction produces specific results.
Before generating any image, answer these four questions:
- What is the exact subject? (not "a developer" — "a terminal window on a dark glass desk, partially reflected, 3/4 angle from the left")
- What is the lighting source? (not "good lighting" — "single cool blue rim light from the top-left, ambient fill from below in warm amber, no direct frontal fill")
- What is the camera relationship to the subject? (close-up, wide establishing, isometric, top-down, eye level, slightly below)
- What mood or decade does this reference? (Apple keynote 2020, early Stripe.com, Vercel 2023, editorial NYT tech section)
If you cannot answer all four, you are not ready to generate.
Prompt Architecture
Structure every prompt in this order:
[Subject description] + [spatial positioning] + [lighting] + [surface/material] +
[background treatment] + [color palette] + [mood/atmosphere] + [technical specs] +
[negative prompt]
The negative prompt is as important as the positive
Always include a negative prompt. The default negative prompt for all brand assets:
--no stock photography, watermark, lens flare, excessive bokeh,
saturated gradients, cartoon, illustration style, human faces,
text overlaid, banner ads, corporate clipart, plugin icons,
emoji-style graphics, purple-to-blue AI gradient
Asset Types and Prompts
Substack Header Background
Goal: abstract, editorial, dark. The text sits on top — the background should not compete.
Prompt template:
Abstract architectural geometry, [specific material: brushed titanium / oxidized steel / dark concrete /
smoked glass] surface, extreme close-up macro photography, shallow focus, cool ambient light from
the left at 15 degrees, rich dark shadows, [accent color]-tinted atmospheric haze in the midground,
no horizon line, no recognisable objects, mood: late-night studio, editorial, considered
--ar 16:9 --no text, faces, plants, logos, purple gradients, stock look
Examples of specific subjects that work:
- Macro of circuit board traces, deep green with amber highlights
- Oxidized metal surface with fine horizontal lines, lit from left
- Dark concrete with a single crack, top-down, blue ambient fill
- Glass prism face, black background, narrow teal beam refraction
- Brushed aluminum sheet, specular highlight along the top edge only
X Post Background / Frame
Goal: visual tension without distraction. The background should make the text pop, not compete with it.
Prompt template:
[Subject] centered in negative space, isolated on near-black background (#080810),
single [color]-toned rim light from the [direction], no secondary light sources,
photorealistic material rendering, [specific material quality], extreme high resolution,
no motion blur, sharp focus on subject, atmosphere: confident, technical, quiet power
--ar 1:1 --no gradient backgrounds, text, logos, watermarks, soft vignette
What works:
- A single object with strong material quality (glass, metal, carbon fiber) in negative space
- Abstract particle fields at very low opacity (grain, not sparkle)
- Minimal geometric forms: a single line, a single plane, a single point of light
- Terminal or interface fragments, partially visible, blurred to abstraction
What never works:
- Busy scenes with multiple subjects
- Gradients as the hero element (they look dated in 12 months)
- Typography in the generated image (it will be wrong)
- Anything that looks like an ad
Thumbnail / Feature Image
Goal: strong, recognisable, shareable. Must read at 240×135px.
Prompt template:
[Clear subject with high contrast] on [simple/solid background], dramatic directional lighting,
subject fills [60-80]% of frame, strong [light/dark] value contrast, single dominant color,
graphic composition, editorial photography style, not a render, photorealistic,
mood: [technical authority / quiet confidence / maker energy]
--ar [16:9 or 1:1] --no busy backgrounds, multiple subjects, text, watermark
Thumbnail rules:
- High contrast between subject and background. If they're the same value, it won't read small.
- One dominant color. Not two. Not three.
- Subject is large. Not a figure in a landscape. A figure that fills the frame.
- Avoid symmetry — off-center compositions have more tension.
LinkedIn Header
Goal: professional without being corporate. Technical without being cold.
Prompt template:
Wide panoramic [subject], [1200×628 landscape composition], low horizon line,
atmospheric depth with soft focus background, cool blue-grey ambient light,
[single warm accent light source], near-monochromatic palette with one [color] accent,
editorial technology photography, cinematic, 2:1 aspect ratio
--ar 2:1 --no text, logos, corporate photography, stock look, cheesy handshake
Color Direction per Project
ystack
- Palette: deep dark navy (
#080810), teal accent (#00d4aa), white text
- Lighting: cool rim light, teal atmospheric haze
- Materials: metal, code, glass, terminal
- Mood: technical authority, open source, late-night builder
Focused AI
- Palette: deep dark (
#080810), violet accent (#7c6bff), soft white
- Lighting: soft violet rim light, warm fill from below
- Materials: glass, fabric (headphones, desk surface), acoustic foam
- Mood: quiet focus, considered, minimal
Lumina
- Palette: dark (
#080810), amber/gold accent (#f59e0b), white
- Lighting: warm amber directional, cool fill
- Materials: brushed gold, glass, clean data surfaces
- Mood: financial confidence, Canadian, modern not flashy
Standard Bio.
- Palette: black (
#080810), teal dot accent, editorial
- Lighting: editorial photography, high contrast, newspaper energy
- Materials: paper texture, ink, signal imagery
- Mood: considered, signal vs noise, infrastructure
Composition Rules
Rule 1: One subject, one relationship
Every image has one subject and one spatial relationship to the viewer. Do not add secondary subjects to "fill space."
Rule 2: Light has a source
Light does not come from everywhere. Specify exactly one primary light source direction and temperature. Fill light (secondary) is optional.
Rule 3: Negative space is not empty
Background is not just a void. It has texture, depth, and atmosphere. A very subtle grain at 0.03 opacity is different from nothing.
Rule 4: Materials have a story
"Metal" is not specific. "Brushed titanium with a horizontal grain, slightly oxidized on the edges" is specific. Material quality signals craft.
Rule 5: The thumbnail test
Before approving any asset, squint at it. Can you still read the composition? If it becomes undecipherable when blurred, the composition is too complex.
What to Never Generate
- Faces. They distract, they get weird, and they belong in photography, not brand graphics.
- Motivational imagery — silhouette of a person climbing a mountain, person looking at horizon, fist pump.
- Corporate diversity photography — diverse group of people around a laptop, smiling at a boardroom.
- The AI gradient — purple-blue linear gradient as background. Banned in all contexts.
- Lens flare as a style choice — this was cinematic in 2012. It is now a stock photo filter.
- Text in generated images — Nano Banana will misspell it. Text goes in the composition layer, not the generated image.
- Logos of any kind — cannot be generated accurately. Create separately and composite.
- Neon cyberpunk — bright neons on dark backgrounds with rain effects. This aesthetic ages poorly.
- Generic "tech" imagery — glowing neural networks, floating UI elements, hands touching holograms.
Iteration Protocol
- Generate 3 variants of the same prompt before choosing. First generation is rarely the final.
- Adjust one variable at a time. If you change lighting AND subject AND background simultaneously, you cannot learn which change improved the image.
- Write the winning prompt down in
.agents/brain/image-library.md with the asset it was used for.
- Build a prompt library. Every successful prompt is an asset.
Asset Sizing Reference
| Asset | Dimensions | Format | Notes |
|---|
| Substack header | 1456×816 | PNG | Export at 2x |
| X post frame | 1080×1080 | PNG | Safe zone 80px |
| X post video bg | 1080×1080 | PNG | Used as bg layer in Remotion |
| LinkedIn header | 1200×628 | PNG | |
| Thumbnail (16:9) | 1280×720 | PNG | Must read at 240×135 |
| README hero | 1200×630 | PNG | GitHub OG card size |
| Favicon | 512×512 | PNG | Centered subject, high contrast |
