xorbitsai/xagent/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.

1. Contemporary reinterpretation — preserves identity while changing the

composition, image language, or type system meaningfully.

1. 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.