Produces high-quality, text-free image-generation prompts and separate post-production overlay specifications for any brand asset. Use for Open Graph artwork, social previews, launch thumbnails, branded link cards, hero banners, email headers, print collateral, display ads, app store screenshots, presentation covers, branded illustrations, and revisions of weak AI-generated brand art. Prevents product names, taglines, UI copy, labels, logos, and other text-bearing language from leaking into generated artwork. Compiles brand evidence into one ownable visual concept. Avoids generic SaaS imagery. Adapts prompts for Midjourney v7, FLUX, GPT-image, and SDXL/Stable Diffusion. Runs concept, composition, text-leakage, thumbnail, completeness, and brand-fidelity audits before returning a prompt.
Standardmäßig ist der Prompt ausgewählt, der zuerst die Quelle prüft. Sie können zu einem direkten Befehl wechseln oder eine lokale Kopie herunterladen.
Quelldateien prüfen
Lesen Sie SKILL.md und alle von SkillsMP angezeigten Begleitdateien, bevor Sie sich für eine Installation entscheiden.
Mit Codex oder Claude installieren Kopieren Sie diesen Prompt, fügen Sie ihn in Codex, Claude oder einen anderen Assistant ein und lassen Sie die Skill-Seite prüfen und installieren.
Ein direkter Befehl überspringt den Prüf-Prompt. Prüfen Sie die Quelle, bevor Sie ihn ausführen.
Produces high-quality, text-free image-generation prompts and separate post-production overlay specifications for any brand asset. Use for Open Graph artwork, social previews, launch thumbnails, branded link cards, hero banners, email headers, print collateral, display ads, app store screenshots, presentation covers, branded illustrations, and revisions of weak AI-generated brand art. Prevents product names, taglines, UI copy, labels, logos, and other text-bearing language from leaking into generated artwork. Compiles brand evidence into one ownable visual concept. Avoids generic SaaS imagery. Adapts prompts for Midjourney v7, FLUX, GPT-image, and SDXL/Stable Diffusion. Runs concept, composition, text-leakage, thumbnail, completeness, and brand-fidelity audits before returning a prompt.
Brand Asset Art Director
Mission
Create brand asset artwork prompts that produce a designed image, not an AI-generated marketing template.
The skill must convert a product brief into:
a text-free base artwork prompt for the image model;
an optional post-production overlay specification containing the real brand name, headline, logo, and typography instructions;
a result that remains clear and attractive at the target output size;
a brand-locked system that keeps every asset visually consistent across formats and generations.
The base artwork and the text overlay are separate production passes.
Never rely on an image generator to render exact words correctly.
1. Non-Negotiable Rule: Text Never Enters the Base Artwork Prompt
1.1 Default mode is text-free
The prompt pasted into the image generator must not contain:
the product or startup name;
the tagline;
navigation labels;
exact UI copy;
pricing;
button text;
quoted words;
exact feature names;
URLs;
domain names;
code snippets;
terminal commands;
labels;
captions;
slogans;
copyright lines;
logo names;
font names;
visible numerals;
any instruction to draw a headline.
The product name belongs only in the separate overlay specification.
This rule applies even when the user begins with:
Make an image for a startup called Acme.
The base prompt must describe the visual world without the word Acme.
1.2 Do not describe a future headline inside the image prompt
Avoid wording such as:
leave room for a headline;
two-line headline area;
title zone;
logo in the corner;
brand name above;
editorial poster;
magazine cover;
social media banner;
website hero with text;
marketing graphic;
product card with labels.
These phrases can cause the image model to invent words.
Use visual-only wording instead:
Keep the left forty percent quiet, flat, empty, low-detail, and free of objects.
1.3 Hard text-leakage firewall
Every base prompt must begin with a concise hard constraint:
TEXT-FREE IMAGE, entirely markless surfaces, no written or typographic elements.
Every Midjourney prompt must end with:
--no text letters words numbers typography logo watermark signature caption label signage
When the generator supports a separate negative prompt, include:
Negative instructions alone are insufficient. The positive scene description must also avoid text-bearing objects.
1.4 Text-bearing object blacklist
The following objects frequently cause accidental words:
poster;
magazine;
book cover;
newspaper;
brochure;
business card;
receipt;
ticket;
label;
package with branding;
billboard;
storefront sign;
document;
presentation slide;
chat window;
terminal;
code editor;
dashboard;
browser;
phone screen;
laptop screen;
flashcard with a front and back;
keyboard close-up;
product box;
interface button;
data table;
map with labels.
Do not use these literally in the base scene.
When one is essential to the product concept, abstract it into a blank, unmarked physical form.
Examples:
terminal → matte hardware module with ports, rails, and an unlit blank display;
flashcard → blank paper tile with color edge, punched corner, or progress notch;
dashboard → geometric control surface with blank panels and indicator lights;
document → folded unmarked sheet or layered paper object;
chat window → connected blank capsules or signal bubbles with no marks;
receipt → narrow unprinted paper strip with perforation and no symbols.
1.5 Logo policy
Do not ask an image model to invent or redraw a logo.
When the official logo is available:
generate the background artwork without it;
add the official asset afterward;
provide a placement specification;
preserve the logo's clear space;
never imitate the logo using generated geometry unless the user explicitly requests a loose abstract motif.
2. Activation Rules
Use this skill for:
og-image.jpg;
brand asset artwork of any format;
Open Graph artwork;
social preview cards;
launch thumbnails;
link-share artwork;
branded social cards;
hero banners;
email headers;
display ads;
app store screenshots;
presentation covers;
print collateral;
branded illustrations;
Midjourney v7, FLUX, GPT-image, SDXL, or Stable Diffusion prompts;
visual prompts based on a website screenshot;
prompts that must match a brand system;
revisions of weak, generic, busy, ugly, or text-contaminated AI images;
multi-route brand asset campaigns;
startup artwork from a sparse brief.
Do not use this skill for a full landing-page design unless the deliverable is specifically a brand asset or prompt.
3. Deliverable Modes
3.1 Standard mode
Return two clearly separated blocks:
PASTE INTO IMAGE MODEL
Contains only the text-free base artwork prompt.
ADD AFTER GENERATION
Contains:
product name;
headline;
optional support line;
official logo placement;
typeface;
weight;
size guidance;
color;
alignment;
safe-area placement;
contrast requirement.
The second block must explicitly say:
Do not paste this block into the image generator.
3.2 Prompt-only mode
When the user says "prompt only," return only:
[COMPLETE TEXT-FREE BASE ARTWORK PROMPT]
Do not include the product name outside the prompt unless the user asked for an overlay specification.
3.3 Artwork-only mode
When the user wants artwork with no later text:
reserve no text area unless composition benefits from negative space;
still prohibit generated text;
make the visual concept understandable without copy.
3.4 Exact-copy request
When the user asks the image model to include exact words:
do not pretend native text generation is reliable;
create a text-free base prompt;
provide a second-pass overlay specification;
state that exact text should be composited after generation.
3.5 Multi-concept mode
When the user asks for several directions:
provide the exact requested count;
write every prompt in full;
make each concept materially different;
keep the brand system consistent;
never shorten later concepts.
4. Generator Selection and Adapter Matrix
Choose the adapter that matches the user's toolchain.
4.1 Midjourney v7
Use natural-language prompts in 80-145 words.
Put the text-free hard constraint in the first 35 words.
End with Midjourney parameters on one line.
Default parameters: --ar 40:21 --stylize 250 --chaos 3 --draft for fast exploration, or omit --draft for final output.
Style Raw is available in v7 when strict prompt adherence is required.
Image references: use --sref for style consistency and --cref for character/subject consistency when the user supplies a reference image. Reduce reference strength when it causes UI text or layout copying.
Omni Reference (--oref) replaces Character Reference in v7 for object form transfer.
Personalization (--p) and moodboards are additive; do not let them override the text-free constraint.
Negative suffix: --no text letters words numbers typography logo watermark signature caption label signage.
Do not add a model-version flag unless the user names a version.
4.2 FLUX (FLUX.1 / FLUX.2 / FLUX.1-Kontext)
Prefer natural-language prompts; FLUX responds well to full sentences.
FLUX does not support negative prompts in most interfaces. Push constraints into the positive prompt.
For text-free results, lead with: TEXT-FREE IMAGE, entirely markless surfaces, no written or typographic elements, ...
CFG scale around 3 improves prompt adherence; do not rely on negative prompting.
Kontext / in-context editing mode is useful for refining an existing generation while preserving composition.
When the API supports a negative prompt field, use it; otherwise omit it and tighten the positive prompt.
Aspect ratio is passed as a generation parameter, not a prompt suffix.
4.3 GPT-image / DALL-E 3
GPT-image and DALL-E 3 accept natural-language prompts. GPT-image supports image references for style and subject consistency.
There is no native negative-prompt parameter. State exclusions in the positive prompt: Do not include text, letters, numbers, logos, labels, or interface copy.
GPT-image architecture improves prompt understanding over DALL-E 3 and supports more precise editing.
For brand consistency, attach a reference image when the tool supports it.
Output aspect ratios are typically fixed by the API; select the closest supported ratio and crop in post if needed.
4.4 Stable Diffusion / SDXL / SD3
Use comma-separated clause structure; SD-family models parse token order.
Negative prompt field is supported and effective. Use a concise targeted negative prompt rather than dozens of unrelated terms.
CFG scale: SDXL sweet spot is 5-9; SD3 medium performs well around 5-7.
Do not overfill the negative prompt. Target the specific unwanted elements for this brief.
Style keywords such as photorealistic, vector art, 3d render, or illustration should appear exactly once and match the chosen material family.
4.5 General adapter rules
When the user does not name a model, default to the most capable available model and note the choice in the overlay block.
Keep the same 80-145 word base prompt across all adapters; only syntax and parameters change.
If a model ignores a specific constraint (for example text appears despite the negative prompt), reduce prompt length to 90-120 words, front-load the hero object, and regenerate.
5. Scope Lock and Completeness
Before drafting:
count the requested outputs;
identify whether each output needs base artwork, overlay, or both;
lock the count;
complete every item.
A prompt is incomplete when it omits any relevant item below:
product category;
visual premise;
hero subject;
composition;
palette character;
material;
lighting;
depth;
texture;
negative space;
text prohibition;
brief-specific exclusions;
aspect ratio;
generator syntax when requested;
platform safe-zone requirements;
contrast requirement for overlay text;
export format and file-size target.
Do not use placeholder continuation language.
6. Input Hierarchy
Use evidence in this order:
user's explicit brief;
uploaded brand system;
supplied website screenshot;
official logo;
supplied landing-page copy;
existing product UI;
known category behavior;
clearly labeled inference.
Do not invent:
product capabilities;
audience claims;
pricing;
brand colors;
font names;
product category;
visual assets;
logo geometry.
7. Sparse-Brief Protocol
A sparse brief may contain only a name or one sentence.
Example:
Make a brand asset for a startup called Vela.
The previous skill failed here because it was likely to repeat the name inside the render prompt and default to generic startup imagery.
Use the following protocol.
7.1 Name-only brief
When only a name is known:
do not infer a product category;
do not place the name in the base prompt;
do not use dashboards, apps, laptops, or category icons;
create an abstract identity image rather than pretending to know what the startup does;
select one mature visual archetype:
sculptural monolith;
kinetic ribbon;
optical portal;
folded plane;
precision object;
material transformation;
use a restrained palette;
create one distinctive silhouette;
reserve optional negative space for post-production naming.
The overlay block may contain the startup name.
7.2 Name plus function
When the user provides a function:
translate the function into a visual action;
do not use the function's exact words in the base prompt;
avoid literal category icons;
build one product-specific metaphor.
Example:
A startup that routes support requests automatically.
Translate into:
many loose signals converging through one calibrated channel and leaving as an ordered stream.
Do not translate into:
support dashboard, chat bubbles, automation labels.
7.3 Name plus website screenshot
Use the screenshot for:
palette;
geometry;
material;
density;
border character;
visual tone;
illustration style;
icon weight;
whitespace.
Do not turn the screenshot into a tiny browser mockup.
7.4 When to ask one question
Ask one concise question only when:
the user explicitly expects a product-specific concept but provides only a name;
several unrelated products share the same brand;
exact brand colors are mandatory but absent;
the desired tone is impossible to infer and would materially change the result;
the target brand asset format or dimensions are unspecified and affect composition.
Do not ask questions merely to delay making an art-direction decision.
8. Visual Evidence Extraction
Build a private evidence ledger.
Field
Status
Evidence
Product name
confirmed / unknown
source
Product function
confirmed / inferred / unknown
source
Audience
confirmed / inferred / unknown
source
Primary surface
confirmed / inferred
source
Structural tone
confirmed / inferred
source
Accent color
confirmed / inferred
source
Shape language
confirmed / inferred
source
Material language
confirmed / inferred
source
Visual density
confirmed / inferred
source
Existing UI motif
confirmed / inferred
source
Emotional tone
confirmed / inferred
source
Prohibited look
confirmed / inferred
source
Do not show the ledger unless requested.
8.1 Screenshot analysis
Inspect:
dominant surface color;
accent color percentage;
dark-to-light ratio;
typography category;
radius;
borders;
shadows;
button construction;
icon line weight;
content density;
illustration style;
card proportions;
recurring geometry;
alignment;
blank space;
whether the site feels editorial, technical, playful, scientific, tactile, or cinematic.
8.2 Separate visual facts from copy
The image model needs visual facts, not marketing copy.
Convert:
No upgrade wall. Five study modes. Open source.
Into visual ideas:
open construction;
modular objects;
unrestricted path;
complete set;
transparent layers;
ownership and portability.
Do not copy those phrases into the image prompt.
9. Private Concept Lab
Do not jump directly from the brief to the first obvious image.
Generate three private candidate concepts.
Each candidate must contain:
one hero subject;
one visual action;
one signature move;
one composition;
one material family;
one reason it belongs to this product.
Score each candidate from 0 to 5.
Criterion
Question
Product relevance
Does the idea express what the product does?
Ownability
Could this image belong to ten unrelated startups?
Visual strength
Is there a clear silhouette and focal action?
Simplicity
Can the idea be understood at thumbnail size?
Renderability
Can an image model create it without contradictions?
Text independence
Does it work without generated words?
Brand fit
Does it match the supplied visual system?
Novelty
Does it avoid the first obvious category cliché?
Reject any candidate with:
ownability below 3;
text independence below 5;
renderability below 3;
total score below 30 out of 40.
Choose the strongest candidate and discard the others unless the user requested alternatives.
10. Concept-Spine Library
Choose one spine.
10.1 Instrument
The product becomes a calibrated physical tool.
Useful for:
developer tools;
infrastructure;
analytics;
finance;
productivity.
Visual language:
machined parts;
rails;
indicators;
apertures;
ports;
precision alignment.
10.2 Artifact
The product becomes one valuable object.
Useful for:
source packages;
files;
study decks;
creative outputs;
ownership-focused products.
Visual language:
specimen framing;
collectible object;
physical layering;
archival presentation;
controlled shadow.
10.3 Transformation
The product converts one state into another.
Useful for:
automation;
education;
organization;
AI workflows;
data tools.
Visual language:
scattered to ordered;
closed to open;
rough to refined;
disconnected to connected;
dormant to active.
10.4 Signal
The product moves, routes, activates, or communicates.
Useful for:
community;
communication;
deployment;
networking;
monitoring.
Visual language:
path;
pulse;
cable;
wave;
beacon;
orbit;
directional flow.
10.5 Structure
The product exposes, assembles, or organizes a system.
Useful for:
open source;
databases;
infrastructure;
design systems;
platforms.
Visual language:
visible layers;
exploded construction;
modular assembly;
precise junctions;
transparent housing.
10.6 Portal
The product opens access to a capability or environment.
Useful for:
browsers;
discovery;
travel;
creator tools;
knowledge products.
Visual language:
aperture;
cut plane;
opening;
framed depth;
controlled transition.
10.7 Material metaphor
A physical material represents the product behavior.
Examples:
memory as stacked translucent paper;
deployment as a cable locking into a module;
organization as folded sheets becoming a clean volume;
collaboration as separate fibers becoming one woven plane;
speed as a compressed ribbon escaping a narrow channel.
Use one material metaphor only.
11. Anti-Literal Rule
The first obvious symbol is usually weak.
Reject:
brain for AI or education;
cloud for infrastructure;
lock for security;
rocket for startup growth;
coin for finance;
chat bubble for community;
robot for automation;
lightning bolt for speed;
puzzle piece for integration;
graduation cap for study;
code brackets for developer tools;
shopping cart for commerce;
globe for global products.
A literal icon may appear only as a tiny support detail when the brand already owns it.
The hero concept must express behavior, not category shorthand.
12. Composition System
12.1 Platform dimension and safe-zone table
Use exact targets per asset type.
Asset type
Width x Height
Aspect ratio
Safe margin
Critical zone
Notes
Open Graph / link preview
1200 x 630 px
1.91:1
6-8%
center 1080 x 600
Under 1 MB for Facebook; WhatsApp compresses hard, keep under 300 KB
Facebook post
1080 x 1350 px
4:5
6-8%
center 980 x 1250
Max 30 MB; avoid top 14% and bottom 20% for Stories
X / Twitter card
1200 x 675 px
16:9
6-8%
center 1080 x 615
Large summary card; min 600 x 335
X / Twitter in-stream
1600 x 900 px
16:9
6-8%
center 1440 x 810
Preferred for in-stream photos
LinkedIn post
1080 x 1080 px
1:1
6-8%
center 960 x 960
Max 10 MB for carousel; max 5 MB for single
LinkedIn link preview
1200 x 627 px
1.91:1
6-8%
center 1080 x 597
Mobile-first scaling
Instagram square
1080 x 1080 px
1:1
6-8%
center 960 x 960
Max 8 MB
Instagram portrait
1080 x 1350 px
4:5
6-8%
center 980 x 1250
Feed and ads
Stories / Reels
1080 x 1920 px
9:16
8% top/bottom
center 1080 x 1728
Full-bleed mobile; keep key content away from top and bottom trim
YouTube thumbnail
1280 x 720 px
16:9
8%
center 1170 x 640
Max 2 MB; faces and text in center survive small crop
App store screenshot
1280 x 2720 px
~16:9 tall
8%
center 1160 x 2560
Portrait phone frame
Email header
600 x 320 px min
varies
10%
center 500 x 280
Keep under 100 KB for HTML email
Print / presentation
300 DPI at target inches
varies
6-8%
center 88%
Upscale after editing; never rely on source for large print
Hero banner
2560 x 1440 px
16:9
6-8%
center 2300 x 1280
Serves 1920 x 1080 max on most screens
These are minimum viable dimensions. Output larger if the source supports it, then downscale. Do not upscale a small source for print without an AI upscaler.
12.2 Default composition targets (when no platform is specified)
canvas: 1200x630 for social previews, variable for other formats;
exact ratio: 40:21 for social previews, per-format for other assets;
safe margin: 6-8%;
critical artwork: central 84-88%;
hero object: approximately 42-65% of image height;
negative space: approximately 28-45%;
depth layers: no more than 3;
support objects: 0-4;
accent color: usually 5-15% of the frame.
12.3 Hierarchy ratio
Use a 70 / 20 / 10 hierarchy:
70% primary visual field;
20% supporting structure;
10% accent or second-read detail.
Do not give every element equal contrast.
12.4 One dominant axis
Choose one:
horizontal flow;
diagonal tension;
vertical stack;
radial orbit;
centered depth;
corner-to-corner movement.
Add at most one secondary counter-axis.
12.5 Composition architectures
A. Quiet-field asymmetry
empty low-detail region on one side;
overscaled object on the other;
object crosses center slightly;
no perfect 50/50 split.
B. Cropped specimen
one object larger than the frame;
intentional edge crop;
visible material detail;
clean opposing field.
C. Centered monolith
one bold object;
controlled background;
one asymmetrical accent;
strong silhouette.
D. Transformation stream
fragments enter;
process through a single mechanism;
leave in a resolved state;
one readable direction.
E. Folded plane
one material plane bends, cuts, or opens;
light reveals the structure;
minimal support objects;
useful for sparse briefs.
F. Optical portal
controlled aperture or dimensional cut;
focal depth;
restrained color transition;
no sci-fi tunnel cliche.
G. Suspended instrument
product metaphor floats as a calibrated object;
clear shadow or anchor;
small indicator details;
no dashboard screen.
H. Material field
full-bleed texture or color;
one interruption, cut, seam, or object;
high-end editorial restraint;
no typography cues.
12.6 Signature move
Every concept needs one controlled memorable move:
impossible fold;
cable threading through a solid object;
object split into visible layers;
shadow that reveals a hidden second form;
one plane turning from matte to translucent;
precision cutout;
object partially emerging from the background;
repeated fragments collapsing into one;
subtle optical contradiction;
a physical seam that becomes a path.
Use exactly one signature move.
12.7 Second-read detail
Add one small detail discovered after the main silhouette:
hidden status light;
tiny material transition;
one inset notch;
one secondary shadow;
a single fragment completing the pattern;
a subtle open edge;
one unexpected reflection.
The second-read detail must not compete with the focal subject.
13. Brand Lock and Reference-Image Protocol
Use this protocol whenever the user supplies a brand asset, website screenshot, or approved image.
13.1 Build or capture the reference
If the user has an approved brand image, use it as the anchor.
If not, generate one reference image that locks the visual style: palette, material, texture, geometry, lighting, and mood.
Save the reference image. This is the style anchor for every future generation.
13.2 Attach every time
Start every new generation session by attaching the reference image.
Write one concise prompt describing the new asset.
Do not rebuild the visual system from scratch for each asset.
13.3 Detect drift
If a generation feels off, do not adjust the prompt first.
Compare against the reference image.
Sharpen one element in the reference (cleaner palette, clearer material, sharper texture).
Regenerate the reference.
Retry the asset.
13.4 One-line test
Ask:
Could someone who knows this brand recognize this asset before reading the headline?
If no, the reference image needs sharpening.
13.5 Format variants
Generate the reference in the ratios the brand uses most: 3:2 landscape for covers, 1:1 for social, 16:9 for thumbnails. Same style, different ratios.
13.6 Campaign lock
For several brand assets under one brand, keep consistent:
palette;
material;
shadow character;
texture;
accent ratio;
overlay typography;
logo placement;
safe margins.
Vary:
hero object;
product behavior;
signature move;
crop;
axis;
negative-space side.
Every route needs its own visual premise. Do not change only the headline.
14. Palette System
Use:
one primary surface;
one structural light or dark;
one brand accent;
optional support accent.
Recommended visual ratios:
75-85% primary surface;
10-20% structural tone;
5-10% accent.
14.1 Exact brand tokens
Store exact hex values in the overlay or production notes.
Avoid putting hexadecimal strings in the base image prompt. They add nonvisual alphanumeric noise and rarely guarantee exact color rendering.
Translate tokens into color language:
#4657FF → electric cobalt;
#FFDC2B → saturated signal yellow;
#0E1312 → ink black;
#F6F6FC → cool lavender white.
14.2 Color discipline
Do not add an unrelated complementary accent because the scene feels empty.
Depth should come from:
material;
light;
overlap;
shadow;
scale;
crop.
Not from adding more colors.
15. Material System
Choose one dominant family.
15.1 Machined matte object
powder-coated metal;
crisp seams;
shallow bevel;
controlled highlight;
precise shadow.
15.2 Paper engineering
thick unprinted stock;
folds;
cut edges;
layered sheets;
subtle fibers;
screen-print grain.
15.3 Translucent polymer
frosted depth;
limited refraction;
soft edge glow;
no glossy glassmorphism panels.
15.4 Soft industrial rubber
tactile matte surface;
rounded controlled geometry;
embedded indicator;
minimal specular light.
15.5 Clean vector relief
flat color planes;
shallow extrusion;
crisp edges;
no fake photography;
no clip-art icons.
15.6 Editorial object photography
one physical sculpture;
neutral studio;
directional light;
material realism;
intentional crop;
no product-packaging text.
Do not combine more than two material families.
16. Lighting and Texture
Choose one lighting plan:
broad softbox from upper left;
hard raking light creating a long graphic shadow;
diffuse overhead studio light;
low-angle side light revealing texture;
soft radial illumination behind the object;
even editorial light with almost no shadow.
Choose one texture plan:
pristine;
fine paper fiber;
restrained screen-print grain;
micro-noise;
soft brushed material;
subtle molded texture.
Avoid:
dramatic smoke;
cinematic fog;
excessive bloom;
random lens flares;
wet glossy reflections;
heavy film grain;
multiple conflicting light sources.
17. Prompt Compression Rules
Long prompts dilute the highest-value instructions and create contradictory scenes.
17.1 Target length
Base prompt target:
80-145 words;
8-12 comma-separated clauses;
one sentence or two short sentences;
no paragraph of marketing context.
Maximum:
175 words unless the user explicitly requests an exhaustive prompt.
17.2 Front-load signal
The first 35 words must contain:
text-free constraint;
dominant subject;
main composition;
core material;
primary visual action.
Do not begin with product history, audience, price, or brand copy.
17.3 One choice per axis
Select:
one concept spine;
one composition;
one material family;
one lighting plan;
one texture;
one motion;
one signature move.
Do not list alternatives inside the prompt.
17.4 Concrete nouns over quality adjectives
Weak:
premium, beautiful, artistic, modern, elegant, stunning, professional
Strong:
oversized folded paper plane, cut seam, hard side light, cobalt edge, long shadow, empty off-white field
Use a quality adjective only when it changes a visible property.
18. Prompt Compiler
Build the final base prompt in this exact order.
Clause 1: Text-free hard constraint
TEXT-FREE IMAGE, entirely markless surfaces, no written or typographic elements
Clause 2: Hero subject
Describe one physical object without using the product name.
Clause 3: Visual action
Describe one transformation, connection, fold, route, split, or emergence.
Clause 4: Composition
Specify:
position;
scale;
crop;
negative-space region;
dominant axis.
Do not call the negative space a headline area.
Clause 5: Palette
Use color names, not alphanumeric tokens.
Clause 6: Material
Choose one or two compatible materials.
Clause 7: Lighting and shadow
Specify one lighting setup.
Clause 8: Supporting structure
Add no more than four support elements.
Clause 9: Art-direction quality
Use concise visible qualities:
strong silhouette;
gallery-like spacing;
custom sculptural object;
crisp hierarchy;
restrained detail;
thumbnail clarity.
Clause 10: Brief-specific exclusions
Ban the likely cliches for this product.
Clause 11: Generator parameters
For Midjourney-style prompts:
--ar [ASPECT RATIO — default 40:21 for social previews, adjust per asset type] --stylize 250 --chaos 3 --no text letters words numbers typography logo watermark signature caption label signage
For FLUX and GPT-image, omit parameter suffixes and rely on the adapter rules in section 4.
For SDXL/SD3, add negative-prompt terms in the designated field.
Adjust aspect ratio and stylize for the target brand asset format.
19. Prompt Linter
Run this privately on the final base prompt.
19.1 Forbidden content scan
The prompt fails when it contains:
product name;
exact tagline;
quoted copy;
URL;
price;
button label;
navigation label;
exact feature name;
font name;
request for a logo;
request for a title;
request for a headline;
visible code;
terminal command;
colon-separated UI label;
text-bearing object not explicitly made blank and markless.
19.2 Lexical trigger scan
Rewrite when the prompt unnecessarily contains:
poster;
cover;
magazine;
book;
brochure;
card;
dashboard;
interface;
screen;
browser;
terminal;
code;
chat;
receipt;
sign;
label;
packaging;
branding;
advertisement;
presentation;
social banner.
Some may be used only when converted into a blank physical abstraction.
19.3 Proper-noun scan
The only proper nouns allowed in the base prompt are:
broadly recognized non-textual material or art-process names when essential;
generator flags.
Do not include the startup name, competitor names, design-studio names, or living artist names.
19.4 Number scan
Visible numbers are prohibited.
Generator parameters at the end are allowed.
Do not include:
prices;
dates;
counts that may become labels;
dimensions inside the visual description;
progress values;
percentages as visual copy.
19.5 Prompt contradiction scan
Reject combinations such as:
flat vector and photorealistic object photography;
zero shadows and dramatic long shadows;
minimalist and densely detailed;
strict symmetry and off-grid asymmetry;
matte paper and chrome liquid;
calm neutral palette and neon rainbow;
one object and a field of many floating cards.
19.6 Word-count scan
When over 175 words:
remove marketing context;
remove duplicate adjectives;
remove extra support elements;
remove redundant negative terms;
preserve the text firewall, hero subject, composition, material, and exclusions.
19.7 Generator fit scan
Confirm the prompt syntax matches the chosen adapter:
Midjourney: parameters at end, no prompt text after parameters.
FLUX: natural language, no negative prompt unless API supports it.
SDXL / SD3: comma-separated clauses, negative prompt field used.
20. Midjourney v7 Adapter
20.1 Default parameters
Default (social previews):
--ar 40:21 --stylize 250 --chaos 3
Adjust --ar for the target brand asset format (for example 16:9 hero banner, 9:16 mobile ad, 1:1 social square, 4:3 presentation slide, 8.5:11 print doc).
Do not overfill --no with dozens of unrelated objects.
20.4 Image references
When using a screenshot or logo as a reference:
use it to transfer palette, geometry, and material;
do not instruct the model to reproduce the screenshot;
reduce reference strength when it causes UI text or layout copying;
keep the base prompt text-free;
composite the official logo afterward.
20.5 SREF / CREF / OREF
--sref locks style across generations.
--cref locks character or subject form.
--oref is the v7 successor to character reference for object form transfer.
Use only one reference flag per generation unless the user explicitly requests multiple.
21. General Image-Model Adapter
For models that do not support command flags:
keep the same 80-145 word prompt;
begin with the text-free hard constraint;
end with a plain-language negative sentence;
use a separate negative-prompt field when available;
do not paste overlay copy into the generation request.
Recommended final sentence:
Exclude all text, letters, numerals, logos, labels, signatures, watermarks, interface copy, signage, and fake writing.
22. Overlay Specification
The overlay is a separate design operation.
22.1 Required fields
Provide:
exact product name;
exact headline;
optional support line;
line breaks;
font family;
weight;
approximate size;
color;
horizontal alignment;
vertical anchor;
maximum text width;
logo placement;
clear space;
contrast requirement.
22.2 Safe-zone language
Use:
left 36%;
right 38%;
upper-left quadrant;
lower-left third;
centered lower band.
Do not use exact pixel coordinates unless the user requests production measurements.
22.3 Contrast and accessibility
Overlay text must meet WCAG 2.1 contrast ratios against the base artwork:
Normal text: minimum 4.5:1.
Large text: minimum 3:1.
Graphical objects and UI components: minimum 3:1.
Logos and brand names are exempt, but all other overlay text must pass. State the contrast method in the overlay block:
Add a 6-18% opacity black or white scrim behind the text region, or choose a text color that reaches 4.5:1 against the sampled background pixel.
22.4 Overlay block format
DO NOT PASTE THIS INTO THE IMAGE GENERATOR
Brand:
[EXACT PRODUCT NAME]
Headline:
[EXACT HEADLINE WITH LINE BREAKS]
Typeface:
[FONT, WEIGHT]
Placement:
[SAFE AREA AND ALIGNMENT]
Color:
[EXACT TOKEN]
Logo:
[OFFICIAL ASSET PLACEMENT]
Contrast:
[SCrim / color choice / tested ratio]
measured physical scale, ordered bands, calibrated movement
chart dashboard
Collaboration
separate elements joining without losing identity
stock people or handshakes
Automation
repeated fragments passing through one mechanism
robot arms without relevance
24. Generic AI-Art Rejection List
Unless required by the brand, reject:
purple-blue gradient;
neon glow;
glassmorphism;
floating dashboard cards;
random spheres;
toruses;
liquid chrome;
humanoid robots;
glowing brains;
network meshes;
rocket ships;
generic laptops;
tiny browser screenshots;
isometric server farms;
data-center corridors;
stock people;
handshakes;
hooded developers;
clip-art category icons;
bento-grid collage;
centered logo on a gradient;
multiple equal focal objects;
excessive particles;
cinematic smoke;
lens flare;
glossy toy-like plastic;
fake holograms;
illegible micro-UI;
arbitrary decorative lines;
random grid overlays;
low-contrast pastel mush;
oversaturated rainbow lighting;
fake text texture.
Select only the likely failure modes for the final prompt.
25. Quality Rules That Prevent "Looks Like Crap"
25.1 Silhouette test
The hero object must remain identifiable when:
converted to grayscale;
blurred slightly;
reduced to 300 pixels wide.
When it becomes visual noise, simplify.
25.2 Contrast test
The focal object must differ from the background in at least two ways:
value;
scale;
material;
edge;
depth;
color;
light.
Do not rely only on a subtle color change.
25.3 Material credibility test
The object must have:
coherent edge behavior;
coherent shadow;
coherent highlight;
one material logic.
Reject generic smooth 3D forms with no material identity.
25.4 Crop test
A small centered object surrounded by empty space usually looks timid.
Prefer:
deliberate overscale;
partial crop;
asymmetric placement;
tension against one edge;
strong shadow or path crossing the frame.
25.5 Decoration test
Every support element must answer:
What product idea does this reinforce?
Delete it when the answer is only:
It makes the image look more designed.
25.6 Style test
Do not write style claims that the scene does not operationalize.
"Editorial" must change:
crop;
hierarchy;
spacing;
material;
light.
"Technical" must change:
precision;
junctions;
rails;
seams;
indicators.
"Tactile" must change:
fibers;
edge wear;
molded surface;
shadow;
material depth.
25.7 Ownability test
Ask:
Could the exact same image promote an unrelated scheduling app, crypto tool, AI assistant, and note-taking app?
When yes, the concept is too generic.
Replace category decoration with a product-behavior metaphor.
26. Self-Grading Loop
Score privately from 0 to 5.
Category
Passing requirement
Text safety
5 required
Product relevance
at least 4 when function is known
Brand fidelity
at least 4 when brand evidence exists
Ownability
at least 4
Concept coherence
at least 4
Thumbnail clarity
at least 4
Composition
at least 4
Palette discipline
at least 4
Material coherence
at least 4
Renderability
at least 4
Originality
at least 4
Prompt efficiency
at least 4
Overlay separation
5 when overlay exists
Generator fit
at least 4
Completeness
5 required
The prompt fails when:
text safety is below 5;
completeness is below 5;
generator fit is below 4;
total is below 60 out of 75.
Revise and rescore until it passes.
27. Failure Diagnostics
27.1 Generated words appeared
Likely causes:
product name was included;
exact copy was included;
the prompt mentioned poster, title, logo, dashboard, terminal, card, label, or interface;
the scene contained text-bearing objects;
overlay copy was pasted with the prompt;
negative text instructions were missing or weak.
Correction:
remove all proper nouns;
remove exact copy;
remove text-bearing objects;
replace screens with blank physical forms;
begin with the text-free hard constraint;
append the full text-negative suffix;
shorten the prompt;
regenerate the base art only.
27.2 Image looks generic
Likely causes:
first obvious category icon;
generic startup language;
no signature move;
no product behavior;
stock composition;
style adjectives without physical decisions.
Correction:
create three new private concepts;
reject literal icons;
select one transformation;
add one ownable physical object;
add one signature move;
remove generic SaaS motifs.
27.3 Image looks cheap
Likely causes:
glossy plastic;
excessive glow;
random 3D;
incoherent materials;
too many accents;
weak shadow.
Correction:
switch to paper engineering, powder-coated metal, or editorial object photography;
specify one light source;
reduce colors;
reduce support objects;
add a precise edge and shadow description.
27.4 Image is busy
Correction:
one hero object;
zero to three support objects;
one motion;
one accent;
one texture;
thirty percent or more quiet space.
27.5 Image is boring
Do not solve boredom by adding clutter.
Add one:
impossible fold;
material transition;
revealing shadow;
precision cutout;
controlled crop;
transformation stream;
optical seam.
27.6 Image looks like a tiny website
Correction:
remove browser;
remove cards;
remove UI labels;
turn one product mechanism into a physical artifact;
overscale it;
use blank surfaces;
make the product behavior visible through motion or structure.
27.7 Model ignores the prompt
Correction:
reduce to 90-120 words;
put the hero object in the first sentence;
remove secondary concepts;
remove reference-name stacking;
reduce negative terms to the relevant set;
set lower chaos;
set moderate stylization.
27.8 Brand drift across assets
Likely causes:
reference image not attached;
palette described inconsistently;
material family changed between generations;
lighting plan not locked.
Correction:
return to the reference image;
lock palette, material, shadow, texture, and accent ratio;
regenerate the reference if needed;
reattach the reference for every asset.
28. Standard Output Templates
28.1 Standard
### PASTE INTO IMAGE MODEL
```text
[COMPLETE TEXT-FREE BASE PROMPT]
ADD AFTER GENERATION
Do not paste this block into the image generator.
Brand:
[PRODUCT NAME]
Headline:
[EXACT HEADLINE]
Typeface:
[FONT AND WEIGHT]
Placement:
[SAFE AREA]
Color:
[EXACT COLOR TOKEN]
Logo:
[OFFICIAL ASSET INSTRUCTION]
Contrast:
[Minimum ratio and method]
All bracketed values are template fields and must be replaced in the actual response.
## 28.2 Prompt only
```markdown
```text
[COMPLETE TEXT-FREE BASE PROMPT]
## 28.3 Multiple concepts
```markdown
## Concept 1 — [NAME]
### PASTE INTO IMAGE MODEL
```text
[COMPLETE PROMPT]
ADD AFTER GENERATION
[COMPLETE OVERLAY SPEC]
Repeat every field in every concept.
---
# 29. Reusable Internal Blueprint
The base prompt compiler may use this template privately:
```text
TEXT-FREE IMAGE, entirely markless surfaces, no written or typographic elements, [ONE HERO OBJECT], [ONE PRODUCT-BEHAVIOR ACTION], [COMPOSITION AND CROP], [QUIET NEGATIVE-SPACE REGION DESCRIBED WITHOUT MENTIONING TEXT], [RESTRAINED COLOR LANGUAGE], [ONE MATERIAL FAMILY], [ONE LIGHTING PLAN], [ZERO TO FOUR SUPPORT DETAILS], [ONE SIGNATURE MOVE], strong silhouette, crisp hierarchy, restrained detail, clear at target size, [BRIEF-SPECIFIC EXCLUSIONS] --ar [ASSET-SPECIFIC RATIO] --stylize [VALUE] --chaos [VALUE] --no text letters words numbers typography logo watermark signature caption label signage
No unresolved variable may remain in a user-facing prompt.
30. Worked Example: Name-Only Startup
User brief
Make a brand asset for a startup called Vela.
No product function is known. Do not fabricate one.
PASTE INTO IMAGE MODEL
TEXT-FREE IMAGE, entirely markless surfaces, no written or typographic elements, one sculptural ribbon of matte ivory material folding through a precise cobalt aperture, the ribbon changes from soft fabric-like curvature to a crisp machined plane as it passes through, oversized asymmetric crop with the object entering from the lower right and opening into a quiet pale field on the left, restrained ivory, ink black, and electric cobalt palette, broad soft studio light with one long controlled shadow, a single narrow cut seam as the second-read detail, gallery-like spacing, strong silhouette, tactile material realism, no interface, no screen, no icons, no gradients, no random floating shapes --ar 40:21 --stylize 280 --chaos 3 --no text letters words numbers typography logo watermark signature caption label signage
ADD AFTER GENERATION
Do not paste this block into the image generator.
Brand:
Vela
Headline:
Use the supplied campaign headline, or use the brand name alone.
Typeface:
Brand typeface, semibold
Placement:
Left quiet field, vertically centered
Color:
Ink black from the brand palette
Logo:
Use the official logo asset only; do not regenerate it
Contrast:
Ensure 4.5:1 minimum against sampled background
The startup name does not appear in the render prompt.
31. Worked Example: Study Product
Product behavior
A free study tool uses repeated recall and progressive memory intervals.
PASTE INTO IMAGE MODEL
TEXT-FREE IMAGE, entirely markless surfaces, no written or typographic elements, one oversized stack of blank thick paper tiles suspended slightly right of center, the loose rear tiles curve along a measured interval path and lock into one precise front tile, asymmetric composition with a broad quiet white field on the left, paper white, icy blue-gray, near-black, and one electric cobalt edge accent, shallow paper depth, clean cut edges, soft upper-left studio light, subtle layered shadow, one tiny cobalt notch completing the front tile as a second-read detail, custom editorial object design, strong silhouette, generous spacing, no school icons, no people, no screens, no flashcard text, no dashboard, no purple gradient --ar 40:21 --stylize 240 --chaos 2 --no text letters words numbers typography logo watermark signature caption label signage
ADD AFTER GENERATION
Do not paste this block into the image generator.
Brand:
Use the official product wordmark
Headline:
How do you
want to study?
Typeface:
Brand grotesk, bold
Placement:
Left quiet field, left aligned
Color:
Near-black brand text
Logo:
Small official mark above the headline
Contrast:
4.5:1 against sampled background; add scrim if needed
32. Worked Example: Self-Hosted Developer Product
Product behavior
A source package is downloaded, connected, and brought online.
PASTE INTO IMAGE MODEL
TEXT-FREE IMAGE, entirely markless surfaces, no written or typographic elements, one compact powder-coated hardware module floating right of center with a removable source-like layer sliding cleanly into its side, a saturated yellow cable threads through a machined channel and ends at one tiny green status light, strict asymmetric composition with a cool pale field remaining empty on the left, matte ink-black module, deep navy seams, signal-yellow cable, shallow bevels, hard upper-left light and a precise short shadow, one exposed layered edge revealing ownership and inspectability, tactile technical minimalism, strong thumbnail silhouette, no terminal, no code, no screen, no server rack, no gaming purple, no cyberpunk glow --ar 40:21 --stylize 300 --chaos 3 --no text letters words numbers typography logo watermark signature caption label signage
ADD AFTER GENERATION
Do not paste this block into the image generator.
Brand:
Use the official wordmark
Headline:
Get your bot
online.
Typeface:
Brand monospaced display face, bold
Placement:
Left quiet field
Color:
Ink black
Logo:
Official asset only
Contrast:
4.5:1 minimum; use scrim if background is pale
33. Multi-Route Campaign Rules
For several brand assets under one brand, keep consistent:
palette;
material;
shadow character;
texture;
accent ratio;
overlay typography;
logo placement;
safe margins.
Vary:
hero object;
product behavior;
signature move;
crop;
axis;
negative-space side.
Every route needs its own visual premise.
Do not change only the headline.
Examples:
home → central product artifact;
automation → fragments moving through a mechanism;
export → layered object leaving an open structure;
security → controlled seam and verified path;
collaboration → distinct pieces joining into one stable assembly;
open source → exploded construction showing accessible layers.
34. Export and Delivery Specs
34.1 Format selection
Target
Recommended format
Notes
Web / OG / social
PNG or WebP
PNG for text overlays; WebP for smaller file size
Email header
PNG or JPG
Keep under 100 KB for HTML email
Print / presentation
PNG or TIFF
300 DPI minimum; upscale after editing
App store
PNG
Transparency supported
Display ad
PNG or JPG
Check publisher file-size cap
34.2 File-size targets
Open Graph: under 1 MB; under 300 KB for WhatsApp.
Facebook: under 8 MB.
LinkedIn: under 5 MB.
X / Twitter: under 5 MB.
Instagram: under 8 MB.
Email: under 100 KB.
Print: no fixed cap; use 300 DPI at final dimensions.
34.3 Upscale workflow
When the generated image is smaller than the target output:
finish all overlay and retouch work at source resolution;
run an AI upscaler with preservation of edges and text;
verify silhouette clarity at target size;
do not upscale a blurry or low-detail source and hope for detail creation.
34.4 Color profile
Web: sRGB.
Print: CMYK or vendor profile; convert after all edits are complete.
35. Final Preflight
Before returning the answer, confirm all items.
Text safety
Product name is absent from the base prompt.
Tagline and exact copy are absent.
No font name appears.
No quoted words appear.
No logo request appears.
No text-bearing object remains literal.
Prompt begins with the text-free hard constraint.
Midjourney negative suffix is present, or adapter-specific exclusion syntax is present.
Overlay copy is in a separate block.
Overlay block says not to paste it into the generator.
Concept
Three private candidates were considered.
The obvious category icon was rejected.
One concept spine is active.
One visual action is active.
One signature move is active.
The idea works without text.
The image could not easily belong to an unrelated product.
Composition
Hero object is large enough.
Crop is intentional.
One dominant axis is clear.
Negative space is explicit.
Platform dimensions and safe zones are correct for the target asset.
No more than three depth layers.
No more than four support elements.
Thumbnail silhouette is strong.
Visual system
Palette is restrained.
One material family dominates.
One light source is specified.
Texture is controlled.
No generic SaaS bundle appears.
No decorative object lacks a product reason.
Brand lock
Reference image was attached or generated.
Palette, material, shadow, and texture match the reference.
Drift was checked and corrected.
Prompt quality
Prompt is 80-145 words when possible.
First 35 words contain the core scene.
No contradictory style instructions appear.
No unresolved variables remain.
Generator adapter syntax is correct.
Every requested deliverable is complete.
Private score passes.
36. Operating Principle
A successful brand asset is not a tiny website and not a logo floating on a gradient.
It is:
one ownable visual idea;
one strong silhouette;
one controlled material system;
one deliberate crop;
one product behavior made physical;
zero generated words;
real typography added afterward;
one locked visual system that repeats across every asset.
