Core logic: upload photo → extract four representative colors and match them reasonably → cut out the real person subject → freely compose each time according to the theme, title, and aspect ratio → process fonts, materials, and decorations in the selected style. The four colors come from the current photo; there is no fixed template for composition; the person retains photographic texture; actions and expressions can be modified.
The task workflow handles input, four-color palette, real-person cutout, free layout, and delivery; the style library provides glyphs, text effects, materials, title carriers, and decorative language. Fixed coordinates, reference palettes, natural backgrounds, or no-cutout suggestions in old style files are examples from extraction rather than execution requirements, and these core rules prevail. When the user explicitly requests the original scene, a fixed composition, or a specific palette for that instance, follow the request for that instance.
Styles are saved in their own styles/ directory: retain the user-provided Y2K, clothing color-sampled real-person torn-paper collage, and 23 design directions organized with user authorization. Do not automatically introduce other Punk skills, and do not restore previously rejected old experimental templates. File integration does not mean the generation effect has been verified.
Research is distilled only into text design rules. The skill package does not include collected reference covers; during generation, do not read the research gallery, screenshots, or blogger photos, and do not pass them to image tools. The output is only the current cover, with no reference images, research annotations, case numbers, schematic frames, or mobile platform interfaces.
Resources
- To choose a style, recommend, or list existing styles: read references/style-catalog.md. This directory is an index; actual styles are subject to
styles/*/STYLE.md. - To write cover prompts: read references/cover-prompt-blueprint.md, then read the selected style's
STYLE.mdandPROMPT.md。 - To extract the four colors for this issue and pair them: read references/photo-palette.md, record the source and use of C1–C4; all styles share this process.
- To arrange space for the person and text, make multiple ratios, or layout variants in the same style: read references/layouts.md。
- To determine character height, glyphs, strokes, backing, curves, and decorations: read references/design-system.md。
- To choose or redesign the person's actions, expressions, and handheld objects: read references/person-remix.md。
- To adapt the same design to other ratios: read references/ratios.md。
- For users adding new style documents, prompts, or references: read references/add-style.md。
SOURCE.mdis raw material, read only when verifying the original text or resolving ambiguity; it is not the complete prompt for every generation.
Input
- Person photo: used to determine the person for this issue. If no photo is provided in a new round, an existing photo explicitly specified by the user may be reused; do not swap in a different person from style references. If there is no photo, you may first choose a style or write a plan; an identity photo is required before actually generating a person cover.
- The user has allowed modification of the uploaded person's actions and expressions, applicable to all styles. Poses, gestures, orientation, gaze, and emotions may be adjusted according to the theme; preserving identity does not mean locking the original posture or original smile. Only when the user explicitly requests in a given task that the original action/expression remain unchanged should they be partially locked.
- Theme or video title: if the user has a title, preserve it verbatim; if there is only a theme, the cover title may be distilled, without fabricating data, results, identity, or quotations. If there is only a photo and no content, ask for the theme, and do not arbitrarily invent a slogan for this photo.
- Aspect ratio: use the user-specified value. When unspecified, use 3:4 and state so; supports 3:4, 16:9, 4:3, 9:16. When the user requests only one aspect ratio, do not automatically output four aspect ratios.
- Optional: style, number of versions, large background text, character name, supporting short phrase, additional requirements. Large background text is only distilled or carried over when the style uses that element; character names and supporting short phrases are omitted when absent, and names are not invented nor copy filled in.
Choose style and multiple versions
Based on the theme's expressive intent, audience, tone, photo shot scale and available space, as well as each style's best_for、avoid_for、style_anchors、supported_ratios selection. The index has candidates routed by theme, and the actual rules of a style are subject to the file.
| User intent | Behavior |
|---|---|
| Specify a style | Use the specified style and do not make the user choose again |
| Give a theme and photo, let you choose; or directly request production without specifying a style | Automatically choose a suitable style, briefly explain the choice and continue, without forcing a wait for confirmation |
| Recommend first, let me choose first | Show at most three existing differentiated candidates, each with one reason fitting the theme; wait for the choice before generating |
| Make a few versions for me to choose from | When no number is given, default to three; prefer different suitable styles and generate them separately |
| Specify the same style, multiple layouts | Use one style and change the layout, line breaks, character proportion, or image-text hierarchy |
| Only modify the skill, integrate a style, or only want a plan or prompt | Complete the corresponding work and do not additionally generate a cover |
Multiple versions reuse the same character, theme, title and aspect ratio; only change the copy when the user requests it. Different schemes should preferably be differentiated in glyphs, text paths, and the relationship between the character and text, not merely by changing the background color. Each version uses one style, generated and saved independently, without compositing a comparison nine-grid.
When there are not enough suitable styles, state so truthfully. If the user vaguely asks for "a few versions", layouts of the same style can be used to fill the number, but they must be labeled as layout variants; when multiple different styles are explicitly requested, first state the current number and ask whether layout variants are acceptable or whether to continue providing styles, without passing them off as other existing styles. The number does not automatically multiply based on the number of aspect ratios.
Compile cover
- View the uploaded photo, extract four representative colors, and record their respective sources; allocate them to background, text, emphasis, and auxiliary uses as needed for this version, as well as derived colors in the same color family, without copying the style's fixed color values.
- Determine the text, canvas, and character protection zone for this issue. Cut the character out of the original photo environment as an independent subject, preserve photographic texture and stable identity, and design actions and expressions according to the theme; do not copy the original background by default.
- Read the selected style's metadata and visual atoms, and design using its glyphs, materials, title carriers, and decorative language. Re-decide the character position, shot scale, title area, line breaks, and text paths each time; the basic composition library is optional inspiration, and layouts are not fixed templates or random rotations by style. Multiple versions must have recognizable compositional differences.
- Compile into concrete prompts according to the blueprint, implementing the four-color sources and roles, real-person cutout, title character-by-character whitelist, font size, text effects, foreground-background hierarchy, actions and expressions, and decorative uses. Title boxes or background text can migrate with this issue's layout, and fixed positions, grid counts, or fixed color palettes from material references must not be treated as constraints.
- Style-specific elements like bubble text, title plaques, paper edges, and doodles only apply to that style; extracting four colors, reasonable color pairing, real-person cutouts, and free composition are core rules shared by all styles. The user's current request takes priority; do not incorporate research analysis or reference images into the generation prompt.
- The model names in the original document are only source identifiers and do not equal models or tool parameters available in the current environment.
Generation and Saving
Before generation, save in the current working directory punk-assets/punk-ip-video-cover/{slug}/brief.md, recording the input, selected style, aspect ratio, and differences for each version; save the prompts separately as prompts/{version}-{style-id}-{ratio}.md。
For actual image production, use the built-in image tools available in the current environment. If local images have not been viewed, view them first. By default, only pass in the identity photo for this session; product images or existing works are passed in as content assets only when provided by the user and genuinely needed for this session, and their role must be made clear. Style is based on the text rules of this skill; do not use collected covers for image-to-image. Use accurate local image paths and avoid using recent research screenshots as inputs in bulk; when the identity image cannot be obtained separately, request that the person's photo be reattached.
Generate one finished image separately for each version and each requested aspect ratio; the quantity does not automatically multiply by the number of styles or the four aspect ratios. When generating actions and themed props, preserve identity according to the selected style; the Y2K rule for preserving clothing/handheld items still applies only to Y2K.
When the user requests a finished product and tools are available, actually generate it; when the user only wants prompts, wants to choose a plan first, or wants to add a style, execute within that scope. If tools are unavailable, keep the prompts and explain, and do not call it a generated image. Do not fabricate tool parameters, and do not switch APIs because a source file mentions a certain model.
Use the current output explicitly returned by the tool to save images to this task directory, naming them by version, style, and aspect ratio. Do not guess files from the global output directory, and do not fabricate files that were not returned.
Checking and Delivery
Check whether the four colors come from the uploaded photo, whether the color roles and derivation relationships are clear, whether the person is treated as an independent cutout subject, and whether the new layout is determined by the content of this session; do not restore fixed coordinates or an old palette just because the style changed.
Check identity, character-by-character accuracy of Chinese text, title hierarchy, text paths, occlusion of eyes, nose, and mouth, contact between hands and props, decorative use, differences between versions, and style anchors. Actions and expressions may be modified; check their thematic fit, naturalness, and whether the identity is still recognizable after modification, and do not judge a change in smile or pose itself as an identity failure; only when the facial features become another person, the limbs are deformed, or it clearly does not fit the theme is correction needed. Confirm that there are no reference images, other people's avatars, original case titles, collection watermarks, research frames, or multi-version collages. Check according to the thumbnail impression of a 3:4 cover at about 270×360, and check other aspect ratios at an equivalent small size.
Read the actual width and height to verify the aspect ratio; if the dimensions cannot be confirmed, explain. When the tool cannot output an exact canvas size, use the image tool to adjust the composition or truthfully report the actual dimensions, and do not stretch faces or text. Do not generate an entire library of extra images just because all style generation validations have not yet been performed.
Only make targeted corrections when there is a specific problem; if the same problem remains unresolved after two corrections, keep the result and explain. Deliver the images and prompts for each version, indicating the style, aspect ratio, differences, and validation status; do not name layout variants as new styles.