Social Thumbnail Workflow

Social Thumbnail Workflow

Generate cover art for short-video platforms (TikTok / Reels / Shorts) or vertical social feeds. Optimized for 9:16, high contrast, big-headline space, and 1-second feed legibility.

Optional assistant scope

Use this planning workflow only when the caller asks for creative development. A supplied storyboard, prompt set or approved count/budget can call tools directly. Preserve explicit tool/model/provider, prompt, ratio, references and quality choices. Do not impose a new concept-selection or approval step on resolved inputs. Discovery is available to any workflow; delegation is optional. Return handles/results to the caller, which owns previews, downloads and presentation. Visual descriptions require actual inspection.

Choosing a dedicated Skill

For an unresolved marketing, event or campaign poster with designed text, call generate_marketing_poster directly. Use brand for the subject, content for supplied copy (autoCopy: false when exact wording matters), and extraNotes for verified details and layout constraints. Put style/design directions in extraNotes or customStyle; they are not words to print verbatim in the poster. A poster needs only a subject; images are optional. Do not force a generic image-generation agent or prompt-enhancement step. Preserve an upstream workflow's explicit tool/model/provider selection; a resolved generic illustration or image-only cover can still use generate_image.

Use list_skills for live styles, output specifications and purchased-credit prices. Choose styleId by its ID, not the displayed label; omit styleId and customStyle for Auto. Nonempty customStyle overrides the preset. styleImage supplies the primary visual style, with written style only as a compatible supplement; never copy its products, claims or wording. logo is an exact identity reference and productImages identify the subjects. Use actual accessible files or image URLs only. These dedicated tools require MeiGen account authorization and purchased credits; OpenAI-compatible providers do not run them.

Resolve only missing required information or scope. A supplied count, copy, quality and approved budget do not need another confirmation. Save a UUID requestId and exact inputs for each logical poster, then use check_skill with skill: "brand-poster" and the original ID after interruption. Follow nextAction; do not create paid replacements automatically.

When to trigger

• User says "video thumbnail", "封面", "短视频封面", "Reels cover", "TikTok cover"
• User asks for creative help planning a vertical thumbnail or poster
• User describes content meant for mobile feeds and asks for visual

Design constraints (apply to every prompt)

1. Default ratio — use aspectRatio: "9:16" with generate_image, or ratio: "9:16" with generate_marketing_poster, for an unspecified vertical cover; preserve an explicit caller ratio. Do NOT write --ar 9:16 in the prompt.
2. High contrast — bold foreground subject against a strong color-block or gradient background. Tiny details get lost when scaled to a phone-feed thumbnail.
3. Headline placement — for an image-only cover or a caller-requested later text overlay, reserve a clear zone for that overlay. For a dedicated marketing poster, supply the visible headline in content and any placement direction in extraNotes; let the poster workflow lay out the text instead of forcing a blank title area.
4. One focal point — feed thumbnails work when the eye lands in <1 second. Don't pack multiple subjects.
5. Brand-safe color — when the user gives a brand color, pin it; otherwise default to high-saturation primary palettes (red / cobalt / orange / electric green) over muted tones.

Workflow

1. Clarify intent

Ask only when essential headline or theme information is missing:

• "What's the headline text the cover needs to support?" (lets you reserve the right zone)
• OR if user already gave a clear theme: skip ask, plan directly.

2. Reference (optional but recommended)

Call search_gallery with the requested poster style or topic for live references. Each page returns at most three entries; use its nextOffset and snapshotId for another batch when requested.