Punk Cover

Punk Cover

Core Rule

punk-cover is a cover-generation workflow, not a style owner. Select exactly one reusable style from the repository-level styles/ library, then compile that style atom into the cover shape.

The final image prompt is not a loose concatenation of "cover instructions + style instructions". It must be one integrated, cover-specific prompt that applies the selected style to the cover format.

Use three inputs:

1. punk-cover task fields: platform, aspect ratio, optional output dimensions and output mode, title clarity, article summarization, cover communication goals, and universal output constraints.
2. The selected style's META.md metadata and STYLE.md reusable visual style atom.
3. references/cover-prompt-blueprint.md, which defines the full cover prompt shape.

The style file defines the reusable visual language. The cover blueprint defines the output form. The final prompt must fuse them into one complete cover-generation prompt.

Resources

• Read references/style-catalog.md to list or choose cover styles.
• Read references/cover-prompt-blueprint.md before composing the final prompt.
• For the selected style, read both:
  • ../../styles/{style-id}/META.md
  • ../../styles/{style-id}/STYLE.md
• Only expose styles whose META.md metadata has outputs containing cover or poster.
• Do not expose photo, avatar, portrait, pet-only, polaroid-only, or image-remix styles in the punk-cover menu unless their metadata explicitly includes cover or poster.
• Do not use legacy cover templates. Use only the selected style's META.md, STYLE.md, and references/cover-prompt-blueprint.md.

Workflow

1. Analyze the source material:

   • Extract the core visual subject.
   • Separate three kinds of content before drafting any visible copy:
     • Source facts: facts, names, claims, and context found in the supplied material. These inform the visual concept but are not automatically approved as text to print on the cover.
     • User-authored copy: title, subtitle, label, or other visible wording the user explicitly supplied for the cover. Preserve every character, punctuation mark, letter case, and wording exactly unless the user asks for editing.
     • Suggested copy: every model-derived short title, subtitle, translation, slogan, label, or rewritten title. Mark each item as a suggestion and obtain explicit user approval before it can become visible copy.
   • Preserve a title supplied explicitly as cover copy. Treat a topic or a title detected only inside source material as a source fact until the user confirms it as visible copy. Do not silently replace a long approved title with a shorter one; either lay out the exact title or propose a shorter alternative and wait for approval.
   • Draft an optional subtitle only as suggested copy when the user did not provide one. A style atom's optional or automatically derived text field never grants permission to print that text.
   • Summarize context, audience, mood, metaphor, and banned elements.
   • Convert long source material into concise derived fields: title/topic, 1-3 sentence context summary, visual subject, audience, mood, metaphor, and banned elements.
   • Create an approved-visible-copy manifest containing the exact main visual title, a complete title only when it will be displayed separately, the subtitle, and any other text that may appear. Use none for absent fields and do not duplicate one title into multiple visible layers by default. Do not put source facts or suggestions in this manifest until the user approves them.

2. Confirm platform and aspect ratio before generating any prompt:

   • Xiaohongshu: 3:4
   • WeChat public account: 2.35:1
   • X: 5:2
   • Custom: keep the user's ratio exactly.
   • If the user only provides source content, ask which platform they want to publish to. Do not generate the prompt yet.
   • If the user provides a custom ratio, use it and do not ask for platform unless the platform matters for wording.
   • If the user provides exact width × height without a ratio, derive and preserve that ratio. If explicit dimensions and an explicit ratio conflict, ask which one should control before generating.
   • Default to a single image. If the user explicitly requests a multi-size suite, preserve every requested ratio or output dimension and require at least two targets before generation.
   • Never satisfy a multi-size request by cropping, stretching, padding, or placing multiple ratios in one grid; compile one independently composed prompt per target ratio.
   • Only skip this question when the user has already provided a platform, a ratio, or explicitly says to decide everything automatically.

3. Confirm style before generating any prompt:

   • If the user specifies one catalog style, use it.
   • If the user supplies a complete visual direction that matches a catalog style, including the 复古时代错位编辑封面 brief, a “超大标题 × 中央视觉主体 × 图文穿插” brief, a “真实纸雕层叠 × 极简留白 × 柔和光影 × 准确隐喻” brief, a “真实纸张质感 × 击凸压凹 × 文字构图 × 克制视觉隐喻” brief, an “抽象概念 → 具体场景 → 游戏机制 → 像素视觉隐喻” brief, or an “OSB 木板 × 工业蓝标识字 × 极简线条隐喻” brief, treat that style as specified and use the matching META.md and STYLE.md without asking the user to repeat the style.
   • If no style is specified, recommend exactly three eligible catalog styles based on the content and give a one-sentence reason for each, then ask the user to choose one or provide a custom style direction.
   • Do not show all eligible styles by default unless the user asks for the full menu.
   • Auto-select exactly one eligible catalog style only when the user explicitly delegates style selection (for example, "you decide the style" or "auto for style") or says to decide everything automatically. Providing an article alone is not delegation.

4. Use this confirmation gate:

   • When platform/ratio, style, or approval for any proposed visible copy is missing, stop after asking the question. Do not fill a style file, save prompt files, or generate an image.
   • The first response for article-only input should contain the platform question, three recommended styles, and any proposed visible copy clearly labeled as suggestions.
   • Continue without asking only when platform/ratio, style, and all visible-copy approvals are known.
   • User-authored copy is already approved when the user explicitly supplies it as cover text and has not requested rewriting. Asking the skill to decide platform or style automatically does not approve model-derived visible copy.

5. Compile the final image prompt:

   • Use references/cover-prompt-blueprint.md as the required structure for prompts/cover.md.
   • Read exactly one selected style visual atom from ../../styles/{style-id}/STYLE.md.
   • Read the selected style metadata from ../../styles/{style-id}/META.md, including style_anchors, cover_shape_adaptation, must_preserve, and avoid_when_applying_to_cover when present.
   • Extract the selected style's non-negotiable visual anchors: materials, spatial logic, title treatment, typography behavior, texture, color system, and style-specific negative constraints.
   • Fuse the cover task fields and style anchors into one full cover prompt. The final prompt should read like a complete cover-generation brief, similar to the legacy complete prompts in exports/, not like two unrelated prompt fragments pasted together.
   • Include cover-shape sections for role/task, input fields, content understanding, title hierarchy, cover composition, style application, image-text relationship, typography, color/material/texture, negative constraints, and final standard.
   • Include the approved-visible-copy manifest and the selected rendering path. Copy approved strings into the prompt exactly; never normalize, translate, shorten, expand, or silently substitute them.
   • Every cover decision must be expressed through the selected style. For example, if the style is torn collage, the title, subject, support text, background, and metaphor must be implemented as torn paper, old newspaper, tape, grain, halftone, and paper layers; not as a generic cover with collage words appended.
   • Do not append the raw STYLE.md content as a standalone second section. Rewrite and adapt its style atoms into the cover blueprint.
   • Do not combine multiple styles or add a second custom style section.
   • Replace or resolve all explicit placeholders such as {{主题词}}, {{副标题，可留空}}, {{画幅比例...}}, {{语言...}}, {{用途...}}, {{补充背景，可留空}}, {{情绪倾向...}}, {{不想出现的元素，可留空}}, and any other {{...}} fields.
   • If a needed detail has no matching placeholder, merge it into the nearest cover section such as 补充语境, 风格落地方式, or 禁用元素.
   • For long articles, keep source summaries concise. A derived title or title layer may enter the prompt only after it has been shown as suggested copy and approved by the user.
   • Put only summarized context into the prompt; do not paste the original article body into the final prompt.
   • Leave optional fields blank only when the prompt says they can be blank.
   • Do not output analysis inside the final prompt.
   • For a multi-size suite, compile the blueprint separately for each target ratio or dimension. Keep the content, metaphor, style identity, material, and palette consistent, but rewrite composition, scale, typography, whitespace, reading direction, and spatial behavior for each target.

6. Choose the text-rendering path before image generation:

   • When exact visible wording is a hard requirement, default to deterministic_text_overlay:
     1. Generate a no-text base artwork with reserved safe areas. Instruct the image model to produce no letters, glyphs, pseudo-text, logos, labels, or watermarks.
     2. Use an available deterministic typography compositor, such as SVG/HTML/CSS rendering, Canvas, Pillow, or ImageMagick, to place only the approved strings. The typography layer must implement the selected style's title treatment rather than look like an unrelated caption.
     3. Inspect the final composited pixels, not just the text-layer source. If the environment has no suitable deterministic compositor, disclose that limitation before generating and ask whether the user wants prompt-only output or accepts the one-stage risk.
   • For text-as-object styles, preserve the style through deterministic geometry, masks, transforms, clipping, or layout when the available compositor supports them. If exact wording and the style's required deformation cannot both be achieved with available tools, present that tradeoff instead of silently weakening either requirement.
   • Deterministic overlay requires retrievable current-run base-image pixels. If the image tool exposes only an inline preview and no bytes, local path, or downloadable artifact, do not pretend the preview can be composited; disclose the limitation and ask the user to choose prompt-only delivery or the acknowledged one-stage risk.
   • Use one_stage_generated_text only when exact text is not a hard requirement or the user knowingly chooses it after being warned that an image model cannot guarantee spelling, completeness, or safe cropping.
   • Use prompt_only when requested or when image generation is unavailable. Prompt correctness does not test generated text.

7. Save files before image generation:

   • For a single image, create punk-assets/punk-cover/{slug}/prompts/cover.md with the complete filled prompt.
   • For an explicitly requested multi-size suite, instead create one file per target as punk-assets/punk-cover/{slug}/prompts/cover-{ratio-or-size}.md, using filesystem-safe ratio or size labels.
   • For a single image, if image generation returns an explicit local file path, downloadable URL, or image bytes for the current run, save that artifact as punk-assets/punk-cover/{slug}/cover.png.
   • Do not infer the correct artifact by scanning broad generated-image directories, because those directories may contain unrelated images from other runs.
   • If the image-generation tool only returns an inline preview with no explicit retrievable artifact for the current run, do not create a fake cover.png.
   • For a multi-size suite, save each explicitly retrievable artifact as punk-assets/punk-cover/{slug}/cover-{ratio-or-size}.png; never combine the suite into a contact sheet unless the user separately requests one.
   • Do not persist intermediate backgrounds, typography files, OCR output, or alternate covers outside these established paths without user authorization. Temporary tool-managed data is not a deliverable.

8. Generate, inspect, and report:

   • Generate an image by default after saving the prompt when a usable image-generation tool is available, such as image_gen. Generate one independent image per target when the user explicitly requests a multi-size suite; otherwise generate one image. An inline preview proves only that an image was generated; it does not prove that visible text is correct.
   • Inspect the actual current-run image or inline preview at sufficient resolution. Compare every visible title, subtitle, and other legible string character-by-character against the approved-visible-copy manifest, and check that each is complete, readable, unobscured, and inside the crop-safe area. Also check that no unapproved or pseudo-text is visible.
   • Never use the prompt's correct wording, a compositor input string, or an intended layout as evidence that the rendered pixels are correct. OCR may assist inspection but cannot replace inspection of the actual artifact.
   • Report exactly one generated-text status per generated image:
     • generated text: passed only when the actual final artifact was reliably inspected and every approved string matches exactly with no cropping or unapproved text.
     • generated text: failed when any character differs, any approved text is missing, duplicated unexpectedly, unreadable, obscured, or cropped, or any unapproved visible text appears.
     • generated text: not tested when delivery is prompt-only or the actual pixels cannot be inspected reliably, including inaccessible or insufficient-resolution previews.
   • Do not describe the cover or its text as successful when the status is failed or not tested. Report image generation, artifact saving, and generated-text validation as separate outcomes.
   • Bound retries for one approved copy and layout: at most two image-generation attempts per target (the initial attempt plus one correction) and, for deterministic composition, at most one overlay correction per target after inspection. After the bound is reached, stop with generated text: failed; offer a next step instead of retrying indefinitely.
   • Retrying does not authorize new copy, extra deliverable files, a different style, a different ratio, or broader filesystem searches.
   • Skip image generation only when the user explicitly asks for prompt-only output or the current environment has no image-generation tool. If image generation is unavailable, return the prompt file path and the full prompt content with generated text: not tested.

First Response Format

For article-only input with no platform and no style, ask concisely:

1. Which platform/aspect ratio should this cover target?
   • Xiaohongshu: 3:4
   • WeChat public account: 2.35:1
   • X: 5:2
   • Custom: provide the ratio
2. Recommended styles:
   • Style A: reason tied to the article.
   • Style B: reason tied to the article.
   • Style C: reason tied to the article.
3. Visible copy:
   • Show user-authored cover copy verbatim when present.
   • Show every derived title or subtitle under a Suggested copy label and ask the user to approve or reject it.

End by asking the user to choose a platform and one style, or to say "auto" if they want the skill to decide those settings, and separately approve or reject any suggested visible copy.

Style Selection Heuristics

• Use styles whose outputs metadata contains cover or poster.
• Use OSB 工业蓝线条隐喻 when the visual direction requires a full-bleed authentic OSB board, upper-left matte industrial-blue signage, one lower-right continuous-line metaphor, and more negative space than text and graphics combined.
• Use Godot 2D 像素隐喻海报 when a theme should become one playable-feeling pixel-art level, one game mechanic, one character action, and one symbolic goal or obstacle rather than a literal illustration or pixel-filtered image.
• Use 立体纸雕概念海报 when one relationship, tension, or transformation should become a physically believable layered-paper metaphor with generous negative space, soft studio shadows, and typography integrated into the paper structure.
• Use 纸面击凸压凹封面 when the cover should feel like an art-book or independent-magazine surface: authentic paper fiber, letterpress-like emboss and deboss, editorial type as composition, one restrained relief metaphor, and large negative space.
• Use 超大标题图文穿插 when the title should become part of the composition through a single central subject, explicit front/back typography layers, controlled occlusion, and strong editorial-poster impact across custom ratios.
• Use business/report styles for strategy, product, AI, startup, industry, consulting, or analysis content.
• Use 复古时代错位编辑封面 for AI, coding, digital work, future tools, or contemporary topics that benefit from a human-centered mid-century illustration and one restrained era-displacement metaphor.
• Use journal/concept styles for science, research, medicine, engineering, or mechanism-heavy content.
• Use collage, giant-title, block, brick, or diffuse styles for social posts that need stronger shareability.
• Use black-white minimal or avant-geometry styles for abstract, philosophical, critical, or high-contrast editorial themes.
• Do not recommend styles whose metadata is avatar-only, portrait-only, pet-only, polaroid-only, or image-only unless they also declare cover or poster.

Cover Prompt Blueprint

Use references/cover-prompt-blueprint.md as the full prompt structure. The block below is retained only as the required task-field content that must be represented inside the integrated final prompt.

# punk-cover cover task instructions

Create one single cover image for {platform}. Aspect ratio: {ratio}.

Use the following derived content only:
- Title/topic: {title_or_topic}
- Optional subtitle: {subtitle}
- Short context summary: {summary}
- Visual subject: {visual_subject}
- Audience: {audience}
- Mood: {mood}
- Visual metaphor: {metaphor}
- Banned elements: {banned_elements}

The main title must use approved visible copy and be complete, accurate, and clearly readable. If a source title is long, do not derive a shorter visible title unless it was presented as a suggestion and explicitly approved.

For long articles, use only derived fields such as title, summary, visual subject, metaphor, and supplemental context. Do not paste the original article body into the image, prompt, metadata, or small text system.

The cover must work for sharing: first glance identifies the topic, second glance reveals the visual metaphor. The image should feel like a deliberate editorial cover, not a generic illustration.

Avoid universal cover failures: PPT cover feel, course-cover feel, generic information-graphic template, e-commerce advertisement, unrelated decoration, misspelled title, missing title, title cropped beyond recognition, or title severely blocked by visual elements.

Generate only one final image per target ratio. For an explicitly requested multi-size suite, generate separate independently composed images rather than a grid or contact sheet. Do not output explanations, alternatives, or multi-option compositions.

Output Discipline

• The final generated prompt must be one integrated cover-specific prompt compiled from the cover task fields, references/cover-prompt-blueprint.md, and exactly one selected style atom.
• Do not paste the selected style STYLE.md as a standalone second section. Adapt its style atoms into the cover blueprint.
• The final prompt must preserve the selected style's non-negotiable anchors from META.md and STYLE.md.
• Do not copy long source text into the final prompt; use summaries and extracted fields.
• Do not include style-selection rationale inside the prompt file.
• Do not combine multiple styles.
• Do not add a second custom style section beyond the selected style visual layer.
• Do not expose non-cover styles in the style menu.
• Do not treat generated-image completion as generated-text validation.
• For prompt-only delivery, always report generated text: not tested.
• Default to one output. Only generate multiple images when the user explicitly requests a multi-size suite, and keep each ratio as a separate composition and artifact.