Skip to content

Latest commit

 

History

History
670 lines (413 loc) · 44.7 KB

visual-advertising-engine.md

File metadata and controls

670 lines (413 loc) · 44.7 KB

🎬 Visual Advertising Engine

The canonical operating standard for every agent that generates advertising. Applies to: ad generation · product photography · lifestyle visuals · Meta Ads creatives · social media graphics · e-commerce visuals · image-generation prompts · image-editing prompts · ad series & campaigns.

This file is the single source of truth for the rules. Every other file in this repo either points here (SKILL.md, design-rules.md) or condenses it for pasting (core.md). If two files disagree, this one wins.

Rule IDs are stable. Cite them in QA verdicts and code reviews: FAIL: R03 (product changed), FAIL: R30-logo. Never renumber a rule — deprecate it and add a new ID.

Rules R01–R55 below
Layout / type / color numbers references/layout-system.md
Headline & copy generation references/headline-system.md
Creative variation matrix (test-ready sets) references/variation-matrix.md
Hook & headline engineering references/hook-engineering.md
Performance feedback loop references/creative-performance-loop.md
Platform compliance & multi-ratio references/platform-compliance.md
Video / UGC motion track references/video-ugc-track.md
Model routing & cost discipline references/model-routing.md
Competitor ad teardown references/competitor-ad-teardown.md
2026 visual languages (style atlas) references/style-atlas-2026.md
Static ad formats (persuasion skeletons) references/static-ad-formats.md
Creative diagnostics (results → next brief) references/creative-diagnostics.md
QA gate (scored, machine-checkable) references/qa-gate.md
Worked end-to-end examples examples/

Scope and precedence

R50 and R51 take precedence over everything below: final ads are generated by an AI image model after research and a short round of questions, never coded. This standard teaches model-independent advertising judgment and prompt writing. R45–R48 add the 2026 layer: one committed visual language, a format chosen by proof and funnel, concept diversity and the current placement system. R53–R55 add the late-2026 layer: check whether a trend cue has decayed, design for Meta's machine remix, and never let a generated image pass as a record. R42–R44 qualify older recipes: choose the medium for the message, and distinguish a style preference from a fidelity requirement. Photography-specific rules apply only to photography. Numerical layouts are starting points. User brief and supplied identity outrank category defaults. Tool setup is optional production guidance, never a prerequisite for writing a prompt.

R01 · MAIN GOAL

We do not generate "pretty AI pictures."

We generate images that could have been produced by:

  • a professional advertising photographer
  • an art director
  • a good graphic designer
  • a product stylist
  • a media buyer who understands scroll and advertising

Every visual must fulfil at least one goal:

  • stop attention
  • show the product
  • show the problem
  • show the use case
  • show the result
  • build product desire
  • increase brand credibility

If an image is only "pretty" but it's unclear what it sells — the creative is weak.


R02 · PRODUCT FIRST

The product is the main character.

The viewer must understand within about a second:

  • what they see
  • what is being advertised
  • why it should interest them

The product must be:

  • clearly visible
  • appropriately large
  • properly lit
  • sharper than its surroundings
  • shown at an attractive angle

Don't hide the product in a huge set. Don't build a beautiful interior where the product has to be searched for.


R03 · REFERENCE IMAGE = SOURCE OF TRUTH

If you receive a product photo, treat it like a technical document.

NEVER change on your own: shape · proportions · color · construction · material · logo · lettering · number of elements · mechanism · how the product works.

AI MAY change: environment · light · frame · perspective · set design · styling · mood.

But not the product itself.

If the product has an unusual mechanism, construction or way of working, the scene must respect the product's physics. Example: if a drop-down shelf slides vertically out from under a cabinet, there must be logical space it slides out of. Don't generate a construction that looks good but couldn't exist.

Always name each reference's role in the prompt: Image A = subject (preserve exactly), Image B = style/grade only. An unlabelled reference gets averaged into the output.


R04 · COMMERCIAL REALISM

The target style: professional commercial photography. Not: "obvious AI-generated advertisement".

Pay special attention to:

  • correct perspective
  • scale
  • gravity
  • contact of objects with surfaces
  • correct shadows
  • reflections
  • textures
  • materials
  • natural depth of field
  • correct human anatomy

Materials must look like materials. Metal like metal. Wood like wood. Glass like glass. Food should be moist, crispy, creamy or juicy when it actually is.


R05 · ANTI AI-SLOP RULE

AI tends to add things nobody needs. Automatically avoid: random neon · holograms · glowing lines · futuristic HUD · pseudo-interfaces · tiny icons · infographic bubbles · random gradients · arrows · fake logos · pseudo-lettering · excessive particles · random decorations · too-perfect interiors · excessive bokeh · plastic surfaces · aggressive HDR.

Don't use an effect just because the generator can create it. Every element must have a function.

Full compendium: references/anti-slop-registry.md.


R06 · ONE CREATIVE = ONE IDEA

Each ad has one main message. Don't try to show at once: product + 7 benefits + promotion + reviews + instructions + parameters + company story + CTA.

Pick one reason for a person to stop the scroll.

One creative. One idea. One focal point.


R07 · VISUAL HIERARCHY

  • PRIMARY — what we see first. Most often: the product.
  • SECONDARY — an element that explains context: a hand using the product, a kitchen, a person, food, a car, a bathroom, a garden.
  • TERTIARY — subtle atmosphere elements. They must not compete for attention.

Two elements of equal weight = no hierarchy = FAIL. Type scale that enforces this: references/layout-system.md §2.


R08 · NEGATIVE SPACE

Don't fill the whole frame. Leave space. Negative space gives: premium feeling · room for a headline · easier ad reading · stronger hierarchy · calmer composition.

A good ad often has fewer elements than a weak one.

Minimum safe margin: 8% of the short edge on every side (86px on 1080px). Nothing crosses it.


R09 · LIGHTING IS PART OF THE PRODUCT

Don't write "professional lighting". Say exactly what the light does.

  • CLEAN COMMERCIAL — large soft key light, subtle fill, controlled shadows, clean highlights. For: e-commerce, electronics, cosmetics, home products.
  • PREMIUM DRAMATIC — directional key light, controlled rim light, darker environment, strong material definition. For: premium products, automotive, alcohol, design objects, food hero shots.
  • NATURAL LIFESTYLE — soft daylight from one side, realistic ambient bounce, natural shadows. For: home, furniture, kitchen, family products, wellness.
  • FOOD COMMERCIAL — strong directional light emphasizing texture, steam, gloss, sauce and crispy surfaces. Food must look so the viewer practically feels its texture.

Always name: the key light's size, direction and quality (e.g. "large soft window light from camera-left, 45°").


R10 · THINK LIKE A PHOTOGRAPHER

Every prompt makes a camera decision. Decide: camera position · angle · distance · lens character · depth of field · foreground · midground · background.

  • HERO — low or eye-level perspective. The product looks more important.
  • LIFESTYLE — 35–50mm natural perspective. The scene looks like real photography.
  • DETAIL — macro / close-up. Shows texture and quality.
  • ENVIRONMENTAL — wider, but the product stays the focal point.

R11 · BUILD DEPTH

Avoid flat images. Good commercial photography has foreground → subject → background. A softly blurred element in front → the main product sharp → subtle environment behind. Gives natural depth and a photographic character.


R12 · SHOW PRODUCT IN USE

A packshot alone isn't enough. Also generate images showing: who uses the product · where · when · what problem it solves · what result it gives.

The face doesn't always have to be visible. A hand, silhouette, gesture, POV or detail often gives enough context.


R13 · THREE MANDATORY MARKETING ANGLES

For every product prepare at least three directions:

  • PROBLEM — the frustration or situation before the product. Example: a hard-to-reach upper cabinet.
  • EFFECT — the result. Example: the shelf drops exactly to the user's height.
  • LIFESTYLE — how the product fits into a better life. Example: elegant kitchen, calm morning, user comfortably reaches for spices.

R14 · VISUAL CREATIVE LIBRARY

Generate many different directions, not the same packshot ten times. Basic set:

  1. HERO SHOT — product dominates the frame, lots of air, strong light, minimal scene.
  2. PREMIUM PACKSHOT — clean studio, controlled light, perfect materials.
  3. LIFESTYLE — product in the user's natural environment.
  4. PRODUCT IN USE — the product's most important function in one moment.
  5. MACRO DETAIL — detail of material, mechanism or texture.
  6. PROBLEM / SOLUTION — the problem visually contrasted with the solution.
  7. RESULT — the effect of using it, above all.
  8. UGC STYLE — more natural photography, less perfect, like a real customer recommendation.
  9. PREMIUM EDITORIAL — big-brand campaign style, more character, less classic e-commerce.
  10. SCROLL STOPPER — unusual perspective, action or moment, still a readable product.

R15 · FOOD REQUIRES DIFFERENT RULES

Food advertising must build appetite. Priorities: texture · steam · sauce · gloss · crunchiness · juiciness · freshness · layers · ingredients.

Good directions: cinematic hero food · dark premium food photography · ingredients suspended in motion · sauce splash · exploded burger/pizza layers · macro texture · steam and heat · chef finishing moment.

For dynamic scenes use Frozen-Time / Bullet-Time — ingredients, sauce, crumbs and drops frozen at the peak of motion. Food must still look physically credible.

Authenticity beats stylization. If the client has real dish photos, they are the hero (R03 applies to food). AI-invented dishes the venue doesn't serve = instant reject (R30).

Layouts, panel heights, gradient values: references/layout-system.md §3. Niche depth: references/hospitality-food-services-playbook.md.


R16 · PEOPLE SHOULD SUPPORT THE PRODUCT

Don't add a person just to have one. A person should: use the product · react to it · show scale · create context · represent the customer. The product or the effect of using it stays most important.

Watch: hands · fingers · gaze · hand-product contact · anatomy · natural gestures.


R17 · TYPOGRAPHY COMES AFTER THE IMAGE

First a strong photograph, then the ad. Hierarchy:

  • HEADLINE — short, bold, readable.
  • SUPPORT — optionally one short sentence.
  • CTA — only if actually needed.

Avoid: many fonts · big text blocks · badges everywhere · small descriptions · random buttons · 6 different text sizes. An ad should not look like a dashboard.

Type scale, pairings, tracking: references/layout-system.md §2. What the words say: references/headline-system.md.


R18 · DON'T LET THE MODEL INVENT TEXT

Two production modes — decide before generating (see references/layout-system.md §5):

  • Mode A · Native in-render text — the copy is baked into the AI render. Only when the model spells reliably. Quote every rendered word, keep strings short, append the spelling directive, vision-QA every variant.
  • Mode B · Deterministic composition (only on explicit user request, R50) — generate a clean background (no text, no logos, no signage), then compose typography and the official logo file with code/design tool.

Either way: never let the model invent logos · prices · product names · slogans · labels · contact details. A "plausible" logo is a FAIL (R30).


R19 · MOBILE-FIRST COMPOSITION

Design for the phone first.

  • META FEED — 4:5 (1080×1350). The default for static ads. Use 4:5 unless the user explicitly asks otherwise. Design on 1080×1350; export 1440×1800 where the placement recommends it (R48).
  • REELS / STORIES — 9:16 (1080×1920). Keep key elements out of the top 14% and bottom 35% UI zones and 6% from the sides (Meta's unified 2026 safe zone, R48).
  • MARKETPLACE / E-COMMERCE — 1:1 (1080×1080), or whatever the marketplace requires.

Don't generate one image and hope random cropping solves every placement. Compose for the specific format. When text or a logo sits near an edge, resize with scale+pad, never a hard crop.


R20 · SERIES MUST HAVE CONSISTENCY

If generating 5–10 images of the same product: the product must be identical. Change: context · frame · mood · light · situation · perspective. Don't change the product. A series should look like one campaign shot during one shoot.


R21 · CREATE VARIATION, NOT RANDOMNESS

Don't create 10 × almost-the-same photo. Create: hero · lifestyle · feature · close-up · problem · result · premium · UGC · unexpected angle · emotional use case. That is real creative diversity.

A colour swap is not a variant. Two creatives are different when they test different promises, not different palettes.


R22 · BEFORE / AFTER

Use Before/After only when it instantly shows the product's value: chaos → order · dark → light · dirty → clean · inconvenient access → product at hand height. Don't do before/after just because it's a popular format.


R23 · GOOD SCROLL STOPPERS

Scroll-stopping doesn't mean "add more effects". Stop attention through: an unusual perspective · extreme close-up · a moment of action · frozen motion · interesting foreground · strong scale contrast · unexpected context · a very simple composition · a face or gesture · showing the problem.


R24 · BUILD THE SCENE AROUND THE BENEFIT

Don't start with "make a beautiful kitchen". Start with "what should this ad say?". Convenience → build a scene of convenience. Premium → a scene of quality. Speed → the time-saving moment. Taste → the food's texture. Message first. Scene second.


R25 · PROMPT ARCHITECTURE

Every final prompt has this structure:

  1. OBJECTIVE — what should the image communicate?
  2. SUBJECT — who is the main character?
  3. ACTION / CONTEXT — what is happening?
  4. ENVIRONMENT — where are we?
  5. COMPOSITION — how are the elements arranged?
  6. CAMERA — perspective and lens character.
  7. LIGHTING — exact light character.
  8. MATERIALS / TEXTURES — what must look especially good?
  9. BRAND MOOD — premium, playful, minimal, natural, etc.
  10. OUTPUT — format and use case.
  11. CONSTRAINTS — what absolutely must not change.

A finished prompt has no placeholders. If any bracket [like this] survives into the final prompt, it isn't finished. Fully worked examples: examples/.


R26 · EXAMPLE OF A WEAK PROMPT

"Create a beautiful premium ad for this product in a modern luxury environment with cinematic lighting."

Too little. The generator has to invent almost everything — and its defaults are the slop in R05.


R27 · EXAMPLE OF A STRONG PROMPT

"Create a premium commercial product photograph using the supplied product reference as the exact source of truth. Place the unchanged product prominently in the foreground of a minimal contemporary kitchen during natural morning light. Show one realistic hand interacting with the product to immediately communicate its core function. Use an eye-level 50mm commercial photography perspective, strong subject separation, subtle foreground depth and a softly defocused environment. Large soft directional window light from camera-left should define the product materials and produce physically realistic shadows. Keep the background simple, neutral and premium with generous negative space above the product for advertising copy. The product must remain identical to the reference in shape, proportions, colors, materials, logo and mechanical details. No additional product features, no fake text, no decorative UI, no neon effects, no excessive props, no visual clutter. 4:5 vertical Meta Ads composition."


R28 · CREATIVE GENERATION WORKFLOW

Do not write the prompt first. In order:

  1. Identify the product.
  2. Identify the most important benefit.
  3. Define the target.
  4. Choose the marketing angle — Problem / Effect / Lifestyle (R13).
  5. Invent a simple visual metaphor or situation.
  6. Choose the creative type (R14).
  7. Write the headline (references/headline-system.md) — the words decide the composition, not the reverse.
  8. Design the composition (references/layout-system.md).
  9. Define the light and camera (R09, R10).
  10. Add constraints (R03, R05).
  11. Only now write the final prompt (R25).

R29 · SELF-CRITIQUE

After generating a concept, evaluate it:

  • Does it look like a real ad campaign?
  • Does it look like typical AI generation?
  • Is the product correct?
  • Is the benefit clear?
  • Is the scene physically logical?
  • Can anything be removed?
  • Would the photograph still work without text?

If it looks like typical AI graphics — redesign. Don't try to save it by adding more effects.


R30 · HARD FAIL CONDITIONS

Automatically reject a visual if:

ID Condition
R30-product the product changed construction
R30-logo the logo is wrong, redrawn or invented
R30-text lettering is fake, misspelled or gibberish
R30-anatomy hands or anatomy are deformed
R30-physics the product's physics is wrong
R30-scale the product is too small
R30-chaos the image is chaotic
R30-ui there's too much UI or icons
R30-background the background draws more attention than the product
R30-stock it looks like stock AI
R30-function elements have no logical function
R30-overload the ad tries to convey too many things

A hard fail is not fixable by retouching. Regenerate.


R31 · THE REMOVAL TEST

"Can I remove something without worsening the message?" If yes — remove it. A simpler creative with a strong idea beats a complicated one without a clear purpose.


R32 · DESIGNER TEST

Imagine the visual goes to: Apple · Nike · IKEA · a premium restaurant · a high-end architecture brand · a professional ad agency. Don't copy those brands. Just ask: "Would an art director accept the photography quality, hierarchy and composition?" If not — improve it.


R33 · META ADS TEST

Shrink the image to phone-screen size (or 150px wide). Check: do I still understand the ad? · is the product visible? · does the focal point still work? · can the headline be read? · does the scene stop attention? If the ad only works on a big monitor, it isn't designed for social media.


R34 · FINAL QUALITY CHECK

Every visual passes the scored gate in references/qa-gate.md before delivery — 10 criteria × 0/1/2, threshold ≥16/20 with zero hard fails. The checklist below is the human-readable version of that gate:

  • One clear goal (R01)
  • One main focal point (R06, R07)
  • Product immediately visible (R02)
  • Product 1:1 with the reference (R03)
  • Correct perspective, materials, light, shadows, physics (R04)
  • Clean background, no AI slop (R05)
  • No fake text, no invented logo (R18, R30)
  • Appropriate negative space and margins (R08)
  • Mobile-first readability at thumbnail size (R19, R33)
  • Clear marketing angle; headline is specific to this business (R24, headline test)
  • Visual works even without copy (R29)
  • Format matches the placement (R19)
  • The whole looks like a professional ad (R32)

R35 · CREATIVE VARIATION MATRIX

A campaign is a set built for testing, not a single hero. Before generating, define the variation matrix: what stays locked (brand-lock) and which test axes rotate between variants.

Stays locked (brand-lock): product (R03) · logo · brand palette · core promise · headline intent. Rotates (one axis per variant): hook wording · visual metaphor · format · lifestyle scene · creator/face · emotion · ratio/placement.

  • Lock the brand, rotate the message. A variant differs by the promise it tests, not by a colour swap (R21).
  • Never rotate two axes in one variant — you won't know what won. One variable per cell, everything else identical.
  • Series consistency (R20) applies to tests too — all variants read as the same campaign, not ten random ideas.

A test set is 1 locked promise × 1 rotating axis × N variants, not N unrelated ads.

Depth: references/variation-matrix.md.


R36 · HOOK & HEADLINE FIRST

The hook decides the composition, not the reverse. Before writing a single prompt, resolve what makes someone stop the scroll — in words.

Run the 20→3 hook gate (see references/hook-engineering.md):

  1. Write 20 candidate hooks for the one promise (R06).
  2. Rank them by stop-power and relevance to the target.
  3. Pick 3 structurally different survivors — different opening mechanism, not three phrasings of the same line.

Only after the hook survives do you build the scene around it (R24) and set the type scale (R17). A headline is not a placeholder you paste onto a finished picture.

If the thumbnail or the first second doesn't carry the message, the ad lost before the visual mattered.


R37 · PERFORMANCE FEEDBACK LOOP

The creative system learns from live results, not from taste alone. After a campaign ships, the metrics come back and feed the next brief.

  • Close the loop: publish → measure (CTR, CPA/ROAS, hook-through, 3s hold for video) → label the winning angle/format → carry it forward as the new default in the next variation matrix (R35).
  • Kill silently, promote loudly. Cut what underperforms without ego; give the winner more budget and more variants.
  • Creative → landing continuity: the ad's promise must be the landing page's first message. An ad that promises one thing and lands on another burns budget (R06).
  • Guard against over-fitting — one small win on a tiny sample is noise, not a new religion.

Depth: references/creative-performance-loop.md.


R38 · PLATFORM COMPLIANCE & MULTI-RATIO

An ad that passes the canvas but dies under platform chrome is a failed deliverable. Compose for the placement, including its UI overlays and safe zones.

  • Per-platform safe zones: Meta feed 4:5 keeps the lower caption/CTA area clear; 9:16 Reels/Stories keep the top and bottom UI zones clear (R19); TikTok and YouTube Shorts have their own text-safe bands. See references/platform-compliance.md.
  • Every ratio via re-layout, never a dumb crop. Scaling/panning to change ratio is a recomposition (headline, CTA, product re-seated), not a crop — a crop loses hierarchy (R07).
  • Deliver native files per placement (4:5 feed, 9:16 short, 1:1 marketplace, 16:9 video), not one image the user has to hack.

If the thumbnail is unreadable (R33) or the CTA is under a platform button, the creative wasn't finished.


R39 · VIDEO & MOTION TRACK

The rules of static advertising survive into motion; motion must serve the message, not decorate it. When the brief calls for a reel, TikTok, story or animated ad, add a motion track:

  • Static-first: the hero frame of the video must pass the same gate as a static ad (R02–R34). Motion is a layer on top of a design that already works.
  • Formats that have a job: product-image-to-video · UGC/avatar talking-head · 3–12s motion from stills · animated loop. Pick the format for the promise (R35), not for novelty.
  • Motion supports the hook (R36): the first second must carry the message; captions and on-screen text are large, clear, purposeful — never a HUD (R05).
  • Route the model to the job: a precise-motion model for product demos, a cinematic model for story-driven ads, a lip-sync/avatar model for UGC — see references/model-routing.md.

Depth: references/video-ugc-track.md.


R40 · ARTIFACT CONTROL

A dirty render is not bad luck — it is one of four named failure modes, and all four are decided before you spend. Diagnose the mode; the fixes do not transfer between them.

  • A · Texture dissolution — fine repeating or organic detail at scale (fur, knit, foliage, crumb, seeds, stone, particles) collapses into Voronoi cells, webbing or noise clusters. Fix by bounding the detail: write the prompt as a layout specification of discrete parts, not a holistic atmosphere; give micro-texture one sharp zone and defocus the rest (R10, R11); name the material, never the density ("ultra-detailed", "intricate", "8k" are direct triggers); and state the negatives explicitly — STRICTLY NO cellular texture, NO webbing, NO repeating Voronoi patterns, NO noise clusters, NO tiled texture fills.
  • B · Style collision — two descriptors that cannot coexist ("Impressionist" + "ultra-detailed"). The model doesn't choose between them, it tries to satisfy both, and the conflict surfaces as artifacting. One artist, one medium is the test. Styles sharing a visual logic combine freely and often beat either alone. An obscure named style is the same failure — describe its mechanics instead (R09, R10).
  • C · Context bleed — the editing memory that makes in-chat refinement work also ghosts earlier images into later ones. The first image in a session is almost always clean; every one after it is dirtier, even on an unrelated subject. This is the largest single cause. So R34's one finished ad per generation becomes one finished ad per session: iterate by opening a clean session with a better prompt, never by nudging the same image again. At scale, prefer an API call — independent context — over a chat thread.
  • D · Quality tier — draft-tier settings are for validating the angle, never for the shipped asset. Draft cheap, render the winner high. If raising the tier doesn't clear it, you are in A, B or C.

Anchoring is the same discipline. A model with no explicit rule set for your brand will interpret one prompt many ways — drifting tone, shadow direction and background across a catalogue. Anchor with 2–3 references maximum, each with a labelled role (R03); more references collide exactly like clashing styles. Anchor specifics — colour values, a lighting reference, an approved packshot — not a vibe.

Repair matches the damage: regenerate for major breakage and for anything touching the product itself (R03 is absolute); mask and repaint only local blemishes; upscale only soft detail. Upscaling a webbed image returns a bigger webbed image.

Artifacting is the one failure a script cannot call — grain, weave and webbing are the same high-frequency energy — so it is inspected at 100%, by eye, in the vision pass, never at thumbnail size.

Depth and the pre-flight checklist: references/artifact-control.md.

A clean render is the precondition for every other rule in this document. An artifacted ad fails R05 before anyone reads the headline.


R41 · MINIMAL EFFECTIVE EDIT

Preserve what already works. Change only what was requested. Most commercial image work is editing, not generation, and the dominant failure is not a dirty texture — it is the model rebuilding what nobody asked it to touch. An edit instruction with no preservation contract reads as a licence to re-solve the whole frame.

Asked to replace the stove, you keep the room. Asked to change the clothing, you keep the face. Asked to improve the lighting, you keep the composition.

  • Every edit prompt carries two slots, never one: EDIT INSTRUCTIONS (the one thing that changes) and PRESERVE (everything that must survive). Silence is not protection — anything unnamed is fair game.
  • Default preserve-list, unless the brief says otherwise: room layout · architecture · proportions · subject identity · facial features · product shape · product branding · furniture placement · camera orientation · the visual logic of the frame.
  • One element per turn. Simultaneous changes are how the four edit artifacts arrive together: reference drift (the scene changed too), edge halo/bleed (the edit was too big to integrate), duplication and warped geometry (the model re-solved a protected region), and overload (several ideas, no hero left — R06 failing at the edit layer). The meta-remedy for all four is the same: make the ask smaller. Two clean edits beat one ambitious one.
  • Integration is part of the instruction. Match scale, perspective, contact shadows and reflections to the existing scene, and ask for a clean boundary with no halo, glow or texture bleed. An object with no contact shadow reads as a sticker (R04).
  • Edit priority order: preserve identity and composition → make the change → match lighting and perspective → repair edge integration → improve realism → remove artifacts → polish. Reference accuracy outranks creative improvement, always. An edit that made the picture nicer while losing the room is a failed edit, not a bonus.
  • Iteration has a shape: one concept = one image family = one session. A direct iteration of what is on screen (angle, light, framing, one object, text fix) stays in the session; a new concept, scene or campaign direction opens a fresh one (R40 mode C).

Depth, the copy-paste preservation blocks and the symptom→remedy table: references/artifact-control.md §7, §8, §11.

The brief asked for one change. Every other change you shipped is a defect, however good it looks.


R42 · TASTE IS A DECISION, NOT A STYLE BAN

Choose a visual grammar that serves audience, message and brand: photographic, documentary, editorial, typographic, illustrated or graphic. A product, service action, event name or verified offer may be the dominant element. Product-first means the advertised subject is immediately understood.

Separate hard failures (false facts, identity drift, broken text, unreadable essential content) from preferences (colour, medium, serif versus sans, symmetry, texture). Cream, purple and gradients can belong to a deliberate identity. Arbitrary choices and incoherent execution are the failures. This qualifies older style prohibitions throughout the repo.

Use photographic depth only for photographic work. Layout numbers guide composition; inspect actual fit. A subline or button-shaped CTA is optional when the message and next action are already clear. Apply the removal test to the prompt as well as the image.

Depth: art direction.

R43 · MARKETING TO VISUAL CONTRACT

Before prompting, resolve audience and moment → verified offer → one takeaway → visible mechanism → action. State what the viewer understands because of what they see. Never invent proof, prices, urgency, reviews or business details. Omit unknown facts or ask when essential.

Explore distinct concepts when useful, then select a coherent direction. For controlled experiments, change one variable and hold the rest fixed; concept exploration may vary multiple decisions without claiming causal attribution. A single requested prompt does not need twenty hooks or a mandatory campaign.

Specify medium, dominant element, reading path, copy field, quiet space, type roles, palette roles and reference fidelity. Explain the few decisions that distinguish this brand from a generic category template.

R44 · PROMPT QUALITY AND IMAGE QUALITY ARE DIFFERENT

R25 is a semantic checklist, not mandatory headings. Combine it into concise prose. Add exact copy, type roles and reading order for finished ads; replace camera/light fields with graphic form when appropriate. The five-slot skeleton is an alternative packaging of those decisions, not a second required prompt.

A prompt specifies one output, one direction, exact quoted copy, composition and relevant constraints. Omit model names, unsupported parameter syntax and irrelevant rules. Check factual support, spatial feasibility, copy length and contradictions before delivery.

Prompting cannot guarantee exact fonts, pixels, spelling or preservation. For prompt-only work, deliver the complete prompt without claiming a rendered result. For an actual image, inspect small-size hierarchy, text and source fidelity before scoring it.

Depth: prompt craft and worked directions.

R45 · COMMIT TO ONE VISUAL LANGUAGE

A style is a complete language: type class and weight, colour roles, image treatment, layout skeleton and one signature device. Choose one per creative, name it in the working note and execute all of its parts. Mixing two dialects in one frame, such as direct flash with riso grain and sticker badges, reads as noise, not richness.

Choose in this order: brand identity → message mechanism → available source assets → placement and funnel → trend. A trend can tip the choice between valid options; it never overrides identity or message. Imperfection, grain and hand marks are process decisions, never filters sprinkled over a clean render. A portfolio presentation board (several tilted ads on one background) is packaging for designers, not an ad.

Depth: style atlas 2026.

R46 · FORMAT FOLLOWS PROOF AND FUNNEL

Before styling a static ad, choose its persuasion skeleton: bold statement, stat drop, review card or stack, product + callouts, comparison, before/after, numbered benefits, founder quote, offer stack, native interface, problem → solution, checkerboard, advertorial, or in-use. The style is the skin; the format is the skeleton. State both.

A format is available only when its proof exists: a verified number for a stat drop, real customer words for a review card, a real person for a founder quote, fair and checkable differences for a comparison, real terms for an offer. Never manufacture proof to fit a template. Match the format to the audience's stage: positioning and native formats for cold audiences, proof-dense formats for warm ones, verified offers for retargeting.

Depth: static ad formats.

R47 · DIVERSITY IS DELIVERY

Meta's retrieval system (Andromeda) groups creatives it judges similar and lets them compete as one. Ten ads that differ only in headline colour, a word or a crop behave like one ad. A campaign set earns reach when its members differ in concept: persona, motivation, hook, format and visual language.

Plan exploration as distinct concepts; plan a controlled test as one rotated variable (R35). They answer different questions, so do not present a cosmetic variant as a new concept. When a concept wins, iterate on it for new personas or contexts rather than recolouring it.

Depth: variation matrix §6.

R48 · DESIGN FOR THE 2026 PLACEMENT SYSTEM

Meta unified the Stories and Reels safe zone in March 2026: keep text, logos and CTAs out of the top 14%, the bottom 35% and 6% at each side of a 9:16 canvas. Feed statics are 4:5; design on 1080×1350 and export 1440×1800 where the placement recommends it. Meta's text fields show roughly 125 characters of primary text, a 40-character headline (fewer on small phones) and a 30-character description before truncation; put the essential message in the image and the first line.

Advantage+ creative enhancements can expand images, add overlays, rewrite text and generate backgrounds, and Meta's in-house image model is rolling into those tools. Keep brand-critical content inside the central area, leave clean extendable edges, and tell the user which enhancements to switch off when identity or product fidelity matters. Ads created or substantially edited with generative AI may carry an "AI info" label; never hide it. Never draw fake functional UI such as play buttons, close icons or notification badges; one on a still is a hard fail, R48-fake-ui. Platform numbers change: verify them in Meta's Ads Guide before final delivery.

Depth: platform guidance.

R49 · DIAGNOSE BEFORE YOU REDESIGN

When results exist, read them before touching the creative. Find the first broken step of the funnel: stopping power (hook), holding power, click, conversion. Change the creative decision that step points to, and only that one. Fatigue asks for a new concept, a weak hook for a new first frame, a good CTR with a bad CPA for a continuity check on the landing page. Metrics diagnose; they never excuse invented proof, and a correlation is not a controlled test.

Depth: creative diagnostics.

R50 · GENERATED, NEVER CODED

Every final creative is generated from scratch by an AI image model (an API model, Codex or the host's image tool), with its imagery, headline, copy and layout in the render. Do not build, template or finish ads with HTML, CSS, code or programmatic text overlays unless the user explicitly asks. This rule outranks every older instruction to use "Mode B", a "deterministic composition" or a "separate typesetting" step. Those passages now describe a fallback that runs only on explicit request.

Make generated text reliable instead: short quoted copy with exact diacritics, "no other text", clear hierarchy and position; supplied logo and product images as references with named roles; after each render, transcribe and compare every word, then regenerate or run a targeted model edit on any error. The facts rules still hold: never let the model invent logos, prices, claims or signage (R18).

R51 · RESEARCH AND ASK BEFORE GENERATING

Before the first generation, research the brand, its Meta Ad Library presence, 2–3 competitors, reviews and the seasonal context. Then ask the user 3–6 questions in a single message, each with a recommended default, about what research could not settle: offer and proof, audience and moment, main objection, style direction, references, placements, language and model. Compress everything into a decision note (brief, insight, concepts, references, exact copy, output) and write prompts from it. After every generation, analyse the result (text transcription, fidelity, thumbnail test, idea, defects) and change one decision per regeneration. A user who asks to skip questions gets stated assumptions instead.

Depth: discovery and research.

R52 · THE WORDS MUST SELL

Every string on the image has a job. The headline sells the product: a concrete benefit or result the buyer gets, in the buyer's words, specific enough that no competitor could paste it onto their ad. The support line gives the proof or the reason to act now. The CTA names the action. A pretty, poetic or clever line that does not say what is sold and why it matters fails, however good the image is.

Run the sell test on every headline before generating:

  1. What is it? Can a stranger tell the product from the image plus headline in one second?
  2. Why me? Does the headline state a benefit, a result or a pain removed, not an adjective or a mood?
  3. Why believe it? Is there a concrete detail such as a number, time, place, material or real proof?
  4. Why now? Is there a verified reason to act (offer, season, date), or at least a clear action?
  5. Could a competitor use it unchanged? If yes, rewrite it.

Copy slop is a hard fail: mood lines with no product ("Poranek ma warstwy."), stacked adjectives, "najlepszy/premium/wyjątkowy" without proof, generic questions, invented urgency, and the banned AI vocabulary. Facts come only from the brief (R43).

Depth: headline system §0.

R53 · CHECK THE SIGNAL BEFORE YOU BORROW IT

A trend cue only works while it still signals something. When image generation costs nothing, cues get copied until they mean "template". Before you use a trend cue, look for it in the Meta Ad Library among the brand's category and 2–3 competitors (R51). If three of them already use it, treat it as decayed: choose a different device, or keep the cue only when it is part of the brand's own identity.

As of October 2026, treat these cues as decayed by default: serif headlines on warm paper-coloured pages (the AI-company look), dithering, ASCII art, pixel grids, scanlines, plain grain overlays and halftone backgrounds. The same test applies to motion: if Meta's static-to-video gives every advertiser the same pans and pacing, a distinctive first frame matters more, not less.

Depth: style atlas 2026 §5.

R54 · DESIGN FOR MACHINE REMIX

Meta's own systems now rework creatives after upload. Advantage+ can expand the frame, add text, generate backgrounds and variants, and turn statics into short videos. Design every static so it survives that:

  • Layers. A clear subject, a clean background plane and the copy area should not overlap in ways that break when the image is animated, extended or re-cropped. Never run a headline across the product or a face.
  • Centre. Keep the headline, logo, offer and the product's identifying details inside the central area. Leave edges that can be extended believably.
  • Control. Name which enhancements to review or switch off when product fidelity, food, real estate, regulated claims or exact copy matter. Use the Restricted Words setting for words generated text must never use. Preview every placement and every generated variant before launch.
  • Distinctive inputs. Generated variants inherit the input. Real demonstrations, real customer language and a committed visual language (R45) are what keep remixed output different from everyone else's.

Depth: platform guidance §6.

R55 · NEVER LET A GENERATED IMAGE PASS AS A RECORD

A generated ad may show a designed world. It may not stand in for evidence. A generated image must never be presented as a real customer, a real employee, a real place, a real event, a real before/after or the real making of the product. "Proof of process" and "documentary" styles are the most exposed: build them from the business's own photos passed as references, or make the staging obvious.

When a photorealistic image shows people, places or events that a viewer could take as authentic, disclose that it is AI-generated or AI-modified in the ad itself, in addition to whatever label the platform applies. In the EU, the AI Act's transparency duties for deepfake content apply from 2 August 2026. Keep the generation log (model, prompt, seed and request id) with every delivered ad, never strip provenance metadata, and tell the user when a brief needs a legal check. This is a production rule, not legal advice.

Depth: platform guidance §6.

🏁 FINAL PRINCIPLE — DON'T DECORATE. DIRECT.

Don't treat the image generator as a tool for adding more and more effects. Treat it like a production crew. First decide: what we show · why we show it · where the viewer looks · what emotions we want · what benefit must be understood. Only later choose: light · lens · set design · styling · color · effects.

The best AI ad shouldn't look like an AI ad. It should look like a very well-planned photo campaign.

One product. One idea. One strong visual.