Static Visual Design

Static Visual Design

Stay within the commercial-creative scope

Use this skill only when the requested deliverable communicates marketing, promotional, campaign, event, or brand-facing material. A request to explain a concept with an image, comparison graphic, educational infographic, technical diagram, chart, or data visualization is outside this skill even when the user asks for a polished PNG or JPEG. Use the general image-generation workflow or a more specific explanatory-visual skill instead.

Do not use this skill to invent a logo, create video, manage ad accounts, generate a standalone illustration or photo with no designed commercial layout, or answer copy-only requests.

Produce the finished visual with image generation and editing. Let the image model solve the composition, type treatment, atmosphere, and graphic language together instead of reducing the work to a generated background plus an HTML layout.

For advertising creatives, campaign posters, promotional social posts, announcement graphics, and other commercially designed static visuals, read references/static-ad-art-direction.md with read_skill_doc before defining directions or rendering. Use its communication structures, layout specification, one-canvas contract, and rejection rules. Do not treat the reference as a menu of decorative styles.

Establish the brief

Identify the minimum useful brief before generating:

• communication goal, audience, and intended response;
• primary message, supporting copy, CTA, and required disclaimer;
• output channel and target aspect ratio or dimensions;
• available logo, QR code, product, brand-guide, campaign, and reference assets.

When available, ground creative direction in real brand materials, product facts, customer language, prior winning creative, and performance evidence, each from a source this skill sanctions below. Treat those inputs as evidence, not as permission to invent adjacent claims.

Distinguish exact supplied copy from facts that still need copywriting. Preserve copy the user marks as exact. When the user supplies facts but not final wording, write concise, idiomatic campaign copy instead of mechanically restating the brief. Avoid defensive or generic claims such as "you cannot be wrong," vague superlatives, and urgency unsupported by the offer.

Do not create a separate brief document unless the user asks for one. Infer low-risk creative choices, but never invent prices, milestones, performance claims, offer mechanics, eligibility, dates, URLs, legal copy, or brand rules. Verify time-sensitive claims when necessary. If exact campaign terms are absent, keep the visual to the claims the user actually supplied.

Develop campaign directions before rendering

For an open-ended brand campaign, poster, promotion, or advertising request, turn the brief into two or three genuinely different communication angles before committing to a visual. Useful angle families include milestone pride or social proof, offer-led value, product benefit, emotional identity, and urgency, but choose only those supported by the brief.

Keep this initial direction-setting in one coherent planning pass. One creative lead should define and compare the full set of directions; do not delegate open-ended ideation to independent agents or parallel nodes. Independent ideation tends to converge on the same obvious brand cues while changing only surface decoration. Parallel execution is useful only after the briefs are locked: each executor receives a distinct proposition, structure, focal device, production finish, and explicit exclusions, and must not reinterpret the other directions.

Represent this order in the execution plan. Brand/reference acquisition is a shared prerequisite, creative direction and design specification depend on that grounding, and every render depends on the locked specification. Do not search for identity assets in parallel with artifact generation. Inspection depends on the render results. Independent renders may run in parallel only after all shared inputs are resolved.

Never bake asset-availability conclusions into planned work before the acquisition step has run. Do not write instructions such as "use the brand name in typography since no logo asset is available" into a step definition; whether a verified logo exists is an outcome of asset resolution, and how the brand is represented must stay open until that outcome is known.

Give every direction one single-minded proposition, one visual device, and one structural approach. The proposition explains why the audience should care; the visual device creates the memorable image; the structure controls how the message is scanned. Evaluate directions for brand fit, stopping power, glance-level clarity, offer comprehension, and factual safety. Do not let a creative become two disconnected ads stacked in one canvas.

Invent each direction's visual device deliberately instead of jumping from an angle name to a layout. Use generative moves — dramatize the verb in the proposition, give the proof number physical form in a real place, literalize the offer as an object or event, stage a recognizable cultural moment, shift scale, or cast a brand asset as the protagonist — and sketch more device candidates in writing than you have direction slots before keeping the strongest. Those sketches are sentences, not renders, so they cost no budget. An angle name like "milestone pride" is not a device; "the milestone staged as a rally the viewer is invited to join" is. The art-direction reference expands these moves.

Do not create directions by accumulating a universal blacklist of subjects or styles. No device is inherently right or wrong: a person, product, phone, mascot, gradient, confetti field, typographic composition, or empty color field can all work when it performs a clear role in the proposition. Exclude an element only when it conflicts with the chosen direction, brand evidence, or message—not because it appeared in a generic list.

When client taste is unknown, build a creative-risk ladder instead of betting everything on one aesthetic:

1. Brand-safe evolution — immediately recognizable, with a cleaner and more disciplined use of familiar brand cues.
2. Contemporary reinterpretation — preserves identity while changing the composition, image language, or type system meaningfully.
3. Bold exploration — uses a more ownable campaign metaphor or unexpected art direction while remaining factually and strategically on brief.

Do not make every option safe or every option experimental. The ladder gives a client a comfortable choice and reveals how much change they will accept.

Do not interpret the singular nouns "an ad," "a poster," or "a social post" as an instruction to explore only one direction. When the creative direction is open, render two or three candidates so the user can choose. When the user explicitly requests exactly one final asset, still compare possible directions and render the strongest one without exposing unnecessary internal deliberation.

Choose a designed structure

Do not default every request to a stock portrait or product mockup beside large headline text. Select a structure that serves the message, such as a dominant stat, a single hero scene, a product or object spotlight with callouts, a before/after contrast, a problem/solution transition, an editorial typographic composition, a proof or quote card, or an information-rich event poster. Treat these as a varied vocabulary, not fixed templates.

Make the headline and image divide the communication work. The image should add proof, tension, emotion, scale, contrast, or surprise instead of literally captioning the headline. Prefer one ownable campaign device over a collage of unrelated symbols. Match information density to the placement: feed ads and banners usually need a short hook and one support line; event and information posters may carry more detail if the hierarchy remains obvious.

Treat the visual device as a mechanism, not a medium label. Photography, illustration, objects, characters, expressive type, spatial relationships, and material treatments are all valid, but the direction must state what the device makes the audience notice or understand. “Bold typography,” “brand gradient,” “large number,” and “premium styling” are treatments, not complete ideas on their own.

Before rendering, turn each direction into a compact design specification: canvas and placement, communication structure, focal subject, visual-weight distribution, three-level information hierarchy, scan path, the depth stack (environment, atmosphere, focal subject, supporting graphics, typography, brand — each layer decided, not defaulted to empty), typography character and line count, color roles, production finish, reference responsibilities, and explicit exclusions. If the scan order is not clear in the specification, the direction is not ready to generate. Keep the specification compact enough to survive intact into the render prompt. Do not create a separate directions or strategy artifact unless the user asks for it, and do not make reading a required reference document its own deliverable step.

Use brand and reference assets intentionally

On vision-capable calls, inspect standard raster reference images already attached to the conversation directly from visual context. Use understand_media when the active model cannot receive native visual context, a reference is only available by file_id, was recovered from the workspace, or requires video or SVG source inspection. Pass useful product, campaign, style, or layout references to image generation or editing so the result belongs to the intended visual world.

For work naming a real brand, resolve the brand identity before rendering final candidates. Three sources are legitimate: an asset the user gave you in this task — including one attached in an earlier turn, whose file_id is no longer in context and which you recover by finding it with list_all_user_files — the current task workspace, and the brand's own official web presence, retrieved with the web tools available to you. If none of them yields a verified asset, ask the user for it; a visually plausible search result is not proof that a logo is authentic.

The user's other tasks and earlier outputs are not a source. Do not list them looking for brand material, and do not reverse-engineer a brand's identity by reading images found that way — a file that turned up in a listing proves nothing about whose brand it shows.

Treat identity-critical assets differently:

• Use an official supplied logo as the source of truth. Include it as a generation reference whenever the image tool supports references so palette, brand language, proportions, and reserved placement influence the whole design. Pass the actual path, URL, or file_id through the image tool's images argument (or use edit_image directly); naming the asset only in the prompt does not attach it. Still do not trust a generated or edited recreation as the final logo.
• Treat QR codes, certification marks, sponsor marks, UI screenshots, and other exact assets as non-generative inputs: generative rendering cannot preserve them pixel-for-pixel. No image tool composites deterministically. If the final placement requires exact reproduction, say so plainly and ask the user to arrange deterministic post-processing; never claim a generated candidate is one.
• Unless the user explicitly requests an unbranded or logo-free concept, a brand-specific final requires a verified logo. If none is available, ask for it and keep any interim output clearly labeled as a concept draft. Do not mark the requested branded asset complete, and never typeset or invent a substitute logo.

Separate stable identity cues from temporary campaign styling. Stable cues may include the official logo, recurring color relationships, typography character, product imagery, graphic proportions, and tone of voice. Temporary styling may include a particular gradient, metallic 3D type, bevel, glow, ribbon, confetti, swoosh, or seasonal campaign motif. Use several recent official references when available to infer what recurs; do not copy every effect from one old banner.

When official references look dated, familiarity still matters, but datedness is not a brand requirement. Preserve recognition through the stable cues and modernize hierarchy, whitespace, type discipline, image quality, depth, and the number of competing effects. A brand-safe direction should feel like the next campaign from the same brand, not a replica of its oldest promotion.

Generate the complete creative

Use generate_image to create the full designed asset, including the intended composition, typography, hierarchy, graphic elements, and user-supplied copy. Prompt with:

• the organizing visual idea and emotional tone;
• the chosen structure, focal subject, and specific visual device;
• exact text to render, quoted clearly;
• hierarchy and approximate placement, without over-constraining every pixel;
• target aspect ratio and viewing context;
• relevant brand, product, campaign, and style references, with a clear statement of what to borrow from each;
• the intended production finish, such as documentary photography, tactile collage, screen print, editorial type, clean product render, or bold flat art;
• any quiet zones required around identity-critical asset placement;
• exclusions such as fake logos, fake QR codes, watermarks, and unrelated text.

Each call must request one finished placement on one continuous canvas. Render each direction separately. Explicitly exclude contact sheets, moodboards, option grids, multiple versions, presentation mockups, repeated layouts, split-frame compositions, and duplicated headlines. Do not put words such as "variations," "option A/B," or "layout exploration" in a render prompt.

Keep copy load proportional to the format. A typical feed ad should have one exact headline of no more than two short lines, at most one short support line, one concise CTA, and only required fine print. The offer is expressed through the headline or the CTA, never as an additional standalone "offer highlight" element; if a brief or step description enumerates headline, support, offer, and CTA as four separate text blocks, merge them back into this budget before rendering. Never send alternative copy, strategy labels, markdown, rationale, or production notes for rendering.

Write a concept-specific prompt for every direction. Do not reuse a generic "modern professional ad" prompt with only the colors and headline changed.

Before generation, compare the locked briefs as a set. Each direction must differ on at least three of these axes: focal subject, structural approach, visual metaphor, production finish, palette balance, and type strategy. Do not reuse the same gradient, display treatment, and decorative motif across the set. Compare the actual render prompts, not merely their direction names or strategy labels. If the prompts describe the same focal mechanism and layout, revise them before rendering. Distinct propositions do not make distinct directions when every focal subject is the same device: a set in which each direction's hero is a large typographic number or headline is one direction wearing three copy variations. For an open brief, at most one direction in a set may use pure typography as its focal subject; the others must be carried by a scene, person, place, object, or material device. The same limit applies to production depth: a deliberately flat, minimal treatment with an empty environment and atmosphere is one direction at most, never the unexamined default for the whole set. When the user explicitly requests a typographic, minimalist, or otherwise constrained system, follow the brief and create the variation inside that system — through composition, scale, material, and color — instead of forcing scenes into it.

For campaign directions developed above and for plural creative requests, make the concepts materially different before making size variants. Vary the idea, composition, subject, image treatment, and hierarchy—not merely crop or accent color. Cosmetic resizes are not distinct concepts.

Choose each aspect ratio from the actual placement, never as a habit, and let the medium set the vocabulary: social campaigns default to 4:5 feed and 9:16 story (1:1 only when a channel specifically requires square); print, poster, out-of-home, and display work follows its placement's actual dimensions, such as standard banner and billboard sizes. When the user has not named channels, either ask or cover the likely placements deliberately and say which asset serves which placement. Do not silently default every render to one square canvas. Compose natively for the chosen ratio: a square canvas wants a centered or radial composition built around its focal device, not a vertical poster stack with dead side margins.

Generate each materially different aspect ratio for that format. Do not force a landscape master into square, portrait, or story placements when a fresh composition would be stronger.

Do not forbid familiar devices merely because they are common in generated ads. A gradient, portrait, phone, giant headline, glow, wave, confetti field, or sparkle can be the right choice. Reject it only when it is an unexamined default, does no communication work, or could move unchanged to a competitor. Favor one clear focal point, a recognizable silhouette at thumbnail size, and a visual idea specific enough that another brand could not use it unchanged.

Treat stacked display effects as a warning sign: metallic 3D lettering plus bevels, glow, drop shadows, ribbons, swooshes, confetti, and multiple headline blocks rarely become stronger by accumulation. Keep only the effects that serve the chosen visual device and remove the rest.

Inspect and iterate with image tools

Inspect every candidate with understand_media, checking:

• exact spelling, numbers, dates, CTA, offer, and disclaimer;
• whether the copy reads naturally and expresses the intended campaign angle;
• hierarchy, contrast, and thumbnail-size legibility;
• whether the headline and image complement rather than repeat each other;
• whether the visual device feels ownable instead of stock or interchangeable;
• crop, balance, edge clearance, and platform safe zones;
• consistency with the supplied references and recognizable brand language;
• accidental pseudo-logos, fake QR codes, watermarks, malformed objects, or unrelated lettering.

Only one failure is a coverage failure: a required asset with no candidate at all. That one is not budgeted — render it.

Everything else is a quality failure, including every structural one: a contact sheet or several ads in one image, duplicated layouts or headlines, fake or duplicated logos, a missing required brand asset, a specification layer the render dropped, wrong or invented copy, unverified claims, unclear hierarchy, clipped or overlapping essentials. Reject them while the repair budget lasts, and never rationalize them as stylistic choices. When the budget runs out, name the defect in your answer rather than accepting it silently — and do not reclassify a defect as missing coverage to buy another render.

Use edit_image only to refine a strong, single-canvas candidate with a localized defect. While the budget lasts, regenerate from the locked design specification when the organizing idea, focal subject, hierarchy, copy load, or canvas structure is wrong, or when several text errors appear. Correct permitted copy errors and inspect again. A successful generation call alone is not proof that the asset is finished. Compare the candidates side by side and discard the safest generic option even when it is technically clean.

Required assets, coverage, and repairs

Three definitions govern every rule in this skill, so that no rule has to restate them.

A required asset is one deliverable this work must produce: one design direction at one placement, either named by the brief or required by this skill — the two or three directions an open brief calls for, at the placements this skill lists for the channel. Anything past that count is an optional asset, whether you frame it as another direction, another variant, or another crop.

Coverage is producing the first candidate for a required asset that has none. Coverage is never budgeted; it is what the deliverable is. Coverage means an asset is absent — not that an existing candidate is wrong, however badly. A render that merged several required assets into one image leaves the others absent, so rendering those is coverage; fixing the merged image itself is not.

A repair is any generate_image or edit_image call on an asset that already has a candidate. Repairs are budgeted: at most two on the same asset, and four across the run. Three things also cost a repair, so that no relabeling buys a free render: any call on an optional asset, including its first; re-rendering a direction you already delivered in order to get a better version of it, even at another required placement; and any call you would otherwise describe as a variant, a fresh angle, or a retry.

Where the two definitions both fit one call, coverage wins: a required placement with no candidate is covered free even when the direction it reuses was already delivered elsewhere. Charge it as a repair only once that placement has its own candidate and you are rendering it again.

When the per-asset limit is reached, stop repairing that asset and deliver its best candidate. When the run budget is spent, stop repairing altogether and deliver the best candidates you have, whatever the per-asset counts say. Neither limit ever stops you from covering a required asset that has no candidate at all.

Count within the run you are in; no counter is shared between plan steps. That is a limitation, not an allowance. Do not spend a step's fresh budget on an asset an earlier step already tried to repair, and do not split one asset's repairs across steps to buy more attempts. An asset that failed inspection in an earlier step is delivered with its defects named, not retried.

Handle identity assets without blind post-processing

Use official logos and other brand assets as generation or editing references when the image model supports them. Inspect the result closely. Never add a second logo over a generated pseudo-logo, and never treat a generic typed brand name as proof of fidelity; while the repair budget lasts, remove the artifact or regenerate the creative. Once it is spent, hand the candidate back with the unfaithful mark named — and, because a brand-specific final still requires a verified logo, hand it back as a concept draft, not as the finished branded asset.

Do not add a deterministic compositing step for ordinary branded visuals. Apply the identity-asset rule above whenever exact reproduction is required; do not pretend a generative result is exact.

Do not use HTML/CSS plus browser screenshots for ordinary poster, ad, banner, or social-creative generation. Use HTML only when the user explicitly requests an editable HTML/template deliverable; it is not the default fallback for text layout.

Apply the completion gate

The gate has two halves and the budget touches only one of them.

Coverage is unconditional. Do not enter final_answer until every required asset exists as a successful tool result, including every direction and placement the brief or this skill asks for. A spent budget is never a reason an asset is absent, because coverage is not budgeted: if one has no candidate, render it. A brand-specific final also needs its verified logo — when none was ever obtained, ask the user rather than shipping a substitute, and keep the interim output labeled a concept draft rather than marking the branded asset complete.

Quality is what the budget releases. Every asset should pass understand_media inspection with no automatic rejection; a successful generate_image or edit_image call is never completion evidence by itself. When the budget is spent and an asset still fails, deliver its best candidate and name the defect concretely — which text is misspelled, which element is clipped, which logo is not authentic — and let the user decide whether to spend another round. Handing back a flawed asset with its flaws stated is a valid outcome; looping until the task runs out of iterations is not. While the budget lasts, keep rejecting: identity marks that are fake, duplicated, misspelled or visibly distorted; a render that omits a required brand asset; output that is merely polished but generic, that carries two disconnected visual ideas, or that weakens the supplied fact into awkward copy. When the user explicitly requires an exact logo, a brand name rendered as ordinary text does not satisfy the requirement.

When render work is a plan step, write both halves into the step's completion criteria and termination condition, so that the step completes either on a clean pass or on a spent budget with the defects named. A stop rule you were handed cannot be edited: when it admits only a clean pass and the budget is spent, finish the step and report the condition as unmet with the defects named, rather than repairing past the budget to satisfy it. When no plan is involved, apply the same rule to the direct decision to stop.

A completion check may ask whether work remains before you finish, and its own instructions treat any named defect as remaining work. Answer it for what the budget settled — repairs are done, coverage is complete — and put the defects in the answer text, where the user reads them, with the answer's outcome partial rather than completed, not in the check's missing_verification field, which reopens the loop the budget just closed. This is a workaround, not a supported path: nothing in the runtime yet distinguishes a budget-exhausted hand-back from an unfinished run, so expect the check to push back and hold the budget anyway.

That check is the one inside your own run. When your run is one step of a plan, a separate assessment decides whether the whole task is complete; it cannot see this skill, it can send work back for more rounds, and it gives up after a few. Name the defects in the step result anyway, but expect no better than that: a defect-named hand-back may be replanned and the task may still fail on the replan limit. Complete coverage is what gives it the best chance, not a guarantee.

Deliver

Return only final PNG or JPEG files that were actually created successfully. Lead with the files, then identify the communication angle and dimensions of each asset in one concise line so the user can compare candidates. Do not present a prompt, brief, HTML intermediate, or claimed file path instead of the requested image.