name: xxd-panel-028 description: "Create XXD Panel 028 raster artwork from one image or a directory batch while treating the bundled original style brief as the sole aesthetic authority. Supports one or more top-bottom, left-right, design-only, and four-device wallpaper outputs; multiple ratios or exact sizes; prompt-generated, user-exact, or text-free typography; and inline parameters. Use when the user invokes xxd-panel-028 or requests the exact Panel 028 style."

XXD Panel 028

Create fresh PNG artwork with references/original-prompt/zh-CN.md as the sole creative and aesthetic authority. This Skill is a runtime shell: it resolves delivery variables, appends them to the untouched source-brief body, executes bitmap generation, and verifies the result. It must not write a second art direction.

Authority boundary

1. Read references/original-prompt/zh-CN.md completely immediately before building every generation request.
2. Its first Markdown heading is an administrative document label, not creative content. Omit that heading from the image-generation request, then copy the entire remaining source-brief body verbatim. Do not summarize, translate, polish, expand, reinterpret, or replace that body with this file, a README, a sample, either runtime-adapter reference, or one of the reading translations in references/original-prompt/. Those translations exist for international readers and sharing; they are never runtime prompt alternatives.
3. The source brief owns subject transformation, composition inside the design region, colour, palette derivation or fixed colour choices, materials, texture, whitespace, text amount, wording logic, typographic character, and prohibitions.
4. Runtime instructions may change only: selected output mode, final canvas ratio or pixels, placement or visibility of the reality source, device profile, wallpaper relationship, target text language, and user-exact text.
5. Append all runtime instructions after the complete source-brief body. Never insert them into, or rewrite, its aesthetic paragraphs.
6. When a source brief describes an old 3:4 top-bottom container, treat that as the original presentation container. The selected mode-specific block replaces that container completely and decisively; every remaining aesthetic and transformation instruction stays active.
7. Never create a generic palette, extract extra colours, lock a swatch set, invent an “aesthetic motive,” pre-compose a title, generate a copy package, or run an external semantic-reading framework. If the source brief itself asks for any of those things, let the image model perform them exactly there.
8. Samples show outcomes only. Do not copy their subject matter, colours, text, layout accidents, or aspect ratio.

Use these roles when the container changes:

REALITY SOURCE = the uploaded photograph or other user-supplied source image
TRANSFORMED DESIGN = the source brief's “lower half” aesthetic transformation

The mode name is not a loose hint. For every requested asset, append exactly one of the four mode-specific delivery blocks below. Never send unused mode alternatives to the image model. Within the selected contract, let the image model decide proportions, scale, crop or environmental extension, whitespace, overlap and internal layout unless the user explicitly requests exact geometry.

Resolve only runtime variables

Every invocation is a new job unless the user explicitly asks to inspect, edit, continue, or reuse a named result. Repeating the same request requires fresh generation and a new task folder; an old matching file never completes a new request. Remembered delivery preferences may reduce repeated questions, but never authorize reuse of an old source, result, prompt, or task directory.

Inline fast path

Parse parameters anywhere after the invocation and input:

/xxd-panel-028 <source-or-directory> --mode top-bottom,design-only \
  --size auto,3:4,9:16,2160x3840 \
  --text prompt --locale ja-JP

• --mode: one or more of top-bottom, left-right, design-only, wallpaper-pack. Repeated flags and comma-separated values accumulate; accept natural-language equivalents and all / 全部.
• --size: one or more of auto, source, 1:1, 3:4, 4:3, 4:5, 5:4, 2:3, 3:2, 9:16, 16:9, 21:9, 5:7, 7:5, any custom ratio, or exact WIDTHxHEIGHT. Repeated flags accumulate; accept ×.
• --text: prompt, exact, or none. Accept legacy aliases auto → prompt and custom → exact without presenting the old labels to users.
• --copy "...": exact user text; implies --text exact. Do not rewrite or translate it.
• --locale: target language, market, or locale for visible text.
• --wallpaper: linked or independent.
• --wallpaper-size: labelled device sizes, for example phone=1440x3200,ipad=2048x2732,desktop=3840x2160,watch=1024x1024.
• --out: explicit output root.
• --prefs: last, edit, new, off, or clear; controls whether this invocation reuses, edits, ignores, suppresses, or clears the family-wide remembered delivery preference.

Explicit parameters override ambiguous prose. Multi-value parameters accumulate and deduplicate in user order; single-value parameters use the last explicit value. If every required variable is resolved, skip preflight and generate. If values are partial, ask only for unresolved variables. Ask about a direct contradiction such as --text none with --copy.

Remembered preference gate

After parsing current prose and inline parameters, read references/runtime-preferences.md completely and follow it before the ordinary preflight. When unresolved settings remain and a valid previous record exists, offer exactly three clear routes: reuse, reuse and edit, or fresh configuration. A fully specified current invocation remains a true fast path and skips this question.

The remembered record is shared across the XXD Panel family because these are common delivery variables, not Soldier aesthetics. Current explicit requirements always win. Never remember exact copy, source paths, Panel selection, generated content, model routes, credentials, or secrets. Once the current settings are fully resolved, save only the safe delivery fields through scripts/panel_preferences.py, unless the user selected --prefs off or asked not to remember the invocation.

Directory batch intake

A readable directory supplied as the source is explicit batch intent. Enter batch processing immediately; do not ask whether the user wants a batch and do not show an unrelated mode menu before inventory.

1. Recursively inventory supported raster files (.png, .jpg, .jpeg, .webp, .heic, .heif, .avif, .tif, .tiff, .bmp) case-insensitively. Respect an explicit top-level-only request or user exclusions.
2. Ignore hidden files and hidden directories, non-images, this Skill's generated task folders, and output roots inside the supplied tree. Do not follow a symbolic link whose resolved target escapes the supplied directory.
3. Use stable natural ordering by relative path. Report the discovered count and recursive scope before generation. Identify unreadable or undecodable candidates; never silently skip them. If no supported images exist, stop without creating an empty task directory.
4. Build one queue item per discovered source. All queue items use this same numbered Panel and its unchanged source brief; a Soldier never selects, blends, or switches to another Panel during a batch.
5. Resolve shared mode, size, text, locale, wallpaper, and output settings once for the whole batch. Parameters and clear prose still skip questions. Resolve auto and source per image because orientations may differ. Before execution, state the input count, shared settings, per-image exceptions, and total expected output count.
6. Treat every source as a logically isolated generation job: read the source brief afresh, construct its prompt independently, and never carry another image's content, inferred meaning, wording, prompt, anchor, or result into it. Wallpaper anchors belong only to their own source.
7. --text prompt generates source-grounded wording independently for each image. --text none applies to all images. A single --copy deliberately applies the same exact text to every image; when exact text differs, accept an explicit relative-path-or-filename → exact-text mapping and verify every intended source before generation. Never reuse one mapped caption for another source.
8. A failed source or asset must be named and reported without silently dropping later queue items. Apply the normal single-asset retry rule only to the failed output, continue with the remaining queue, and finish with succeeded/failed counts and paths.

One batch invocation is one physical task directory even though its queue items are logically isolated. Do not create a task directory per source.

Capability-adaptive preflight

Do not switch the user's session into Plan mode merely to obtain a question UI. A Skill describes behaviour but cannot create UI capabilities that the host has not exposed. Detect the actual question-tool schema and follow this order:

1. Real multi-select tool available — for example Claude Code AskUserQuestion with multiSelect: true: use genuine checkbox questions for modes and sizes. Use single-select questions for text mode and wallpaper relationship. Do not print a redundant numbered menu after showing the form.
   • First form: modes (top-bottom, left-right, design-only, wallpaper-pack) with multiSelect: true; size route (自动适配, 跟随原图, 常用比例, 自定义) with multi-select enabled when the host permits combined size sources; text mode as single-select.
   • When 常用比例 is selected, show real checkbox groups for all concrete ratios, split only as required by host option limits: square 1:1; portrait 3:4, 4:5, 2:3, 9:16, 5:7; landscape 4:3, 5:4, 3:2, 16:9, 21:9, 7:5. Accumulate selections across groups. When 自定义 is selected, collect one or more ratios or exact pixel targets through the host's free-input/Other path.
   • If wallpaper is selected, ask linked versus independent as a genuine single-select. Ask only unresolved questions and respect answers already present in prose or parameters.
2. Only a mutually exclusive question tool is available — for example Codex request_user_input: use it only for genuinely single-choice fields such as text mode, wallpaper relationship, or the size-entry route. Never represent modes or concrete sizes as a single-select when the user may choose several. Collect those multi-value fields with the two-round combination input below. Do not claim that a single-choice card is a checkbox.
3. No interactive question tool is available: use the same two-round combination input below. It is typed multi-value input, not a clickable form. Never draw Markdown - [ ] boxes or other fake controls.

Inline parameters and clear natural language always take priority. Skip every field already resolved. When all required values are present, skip every question and generate immediately.

Two-round combination-input fallback

First round — modes. This is a typed combination question, so label it honestly instead of calling it a checkbox:

请选择一个或多个成品类型：

1. 上下对照（原图在上，028 设计在下）
2. 左右对照（原图在左，028 设计在右）
3. 只要设计图（原图仅作参考，不出现在成品中）
4. 四端壁纸（手机、iPad、电脑、儿童手表）

可回复：1｜1+3｜2、4｜全部

Second round — when at least one ordinary mode is selected, ask size route and text mode together. If the selection contains only wallpaper-pack, skip the ordinary-size part and resolve text plus device sizes and wallpaper relationship.

请选择尺寸，可多选：

1. 智能推荐（显示本次计算出的比例与像素）
2. 跟随原图
3. 常用比例（可直接写 3:4、9:16 等一个或多个比例）
4. 自定义（输入一个或多个比例／准确像素）

请选择文字方式：

A. 模型根据原始提示词生成文字（注明语言或地区）
B. 使用我的准确文字（发送原文并注明语言或地区）
C. 不要文字

可回复：尺寸 1＋9:16；文字 A，日语
可回复：尺寸 3:4、16:9；文字 B，准确文字「……」，简体中文

If the user answers only 常用比例, show the grouped library once and accept any combination:

方形：1:1
竖版：3:4、4:5、2:3、9:16、5:7
横版：4:3、5:4、3:2、16:9、21:9、7:5
也可以直接输入准确像素。

Do not decide wording outside the source brief. For option A, provide only the target language to the image model; the image model follows the source brief's existing text logic. For option B, pass the user's characters verbatim and add no other copy. For option C, prohibit all visible text and pseudo-text.

Do not infer language or market from a person's appearance, name, clothing, scene, filename, metadata, signage, or the language used to operate the Skill. Resolve it explicitly whenever text is requested. Use natural target-language shaping, direction, punctuation, spacing, and line breaking. For Arabic, preserve connected forms and right-to-left text flow without indiscriminately mirroring the artwork.

If wallpaper-pack is selected, ask one additional single-choice question unless resolved:

四端壁纸采用哪种关系？

1. 连贯套装：先生成并确认一张定调图，其余三张参考原图＋定调图独立重构
2. 四张独立：四张都只参考原图，各自构图

Neither option crops or mechanically resizes one wallpaper into the other devices. Use the common device preset above unless the user supplies labelled device sizes. Ordinary-size choices do not multiply wallpaper outputs.

Size resolution

• No silent default ratio exists.
• For auto, recommend a whole-canvas ratio from the source orientation, selected mode, source brief and intended use, then show the resolved ratio and pixels. Do not derive panel boxes or crop coordinates; the image model composes the selected relationship inside that canvas.
• source means the source photograph's aspect for the whole requested ordinary output.
• Exact pixels override a ratio.
• Every distinct aspect ratio is a separate complete-canvas composition from the same source and the same verbatim style brief. Never crop one ratio from another. Multiple resolutions sharing one aspect may be exported from the highest-quality approved composition.
• A size set applies to every selected ordinary mode unless the user maps sizes per mode.
• Before generation, state the resolved modes, concrete sizes, text mode and locale, wallpaper relationship if relevant, and total output count. Do not ask for another confirmation when they are complete and consistent.

Build the generation prompt

For each distinct asset, concatenate exactly:

[VERBATIM SOURCE-BRIEF BODY FROM references/original-prompt/zh-CN.md,
EXCLUDING ONLY ITS ADMINISTRATIVE FIRST MARKDOWN HEADING]

[COMMON DELIVERY PREAMBLE]
[EXACTLY ONE SELECTED MODE BLOCK]
[EXACTLY ONE TEXT BLOCK]
[USER'S OTHER EXPLICIT REQUIREMENTS, IF ANY]

Use this common preamble:

MODE-SPECIFIC DELIVERY OVERRIDE — CURRENT ASSET

This block is the final authority only for the current presentation mode,
reality-source visibility, final canvas and device delivery. It completely
replaces the source brief's legacy statements about 3:4, upper/lower placement,
equal sections and the old top-bottom container.

Every source-brief instruction about the transformation's visual language,
subject identity, colour, material, texture, internal composition, whitespace,
text character and typography remains authoritative.

REALITY VIEW means the faithful photograph or factual scene defined by the source brief.
TRANSFORMED DESIGN means the source brief's designed reinterpretation of that reality view.

FINAL CANVAS: <resolved ratio and/or exact WIDTHxHEIGHT>
COMPOSITION METHOD: ONE COHERENT COMPLETE-CANVAS GENERATION
EXACT PANEL GEOMETRY: ONLY WHEN THE USER EXPLICITLY REQUESTS IT

Colour follows the original brief's existing colour instructions exactly.
Unless the user explicitly requests a colour change, do not add, replace,
summarize, or re-plan any palette.

Append exactly one of these mode blocks; do not include the other three.

Top-bottom:

OUTPUT MODE: TOP_BOTTOM

Create one complete canvas whose dominant structure is two primary horizontal
parts: the REALITY VIEW above and the TRANSFORMED DESIGN below. Together they
organize the composition across its full width. Let the image model determine
their visual proportion and each part's internal crop, extension, whitespace
and typography from the source, source brief and final canvas.

Left-right:

OUTPUT MODE: LEFT_RIGHT

Create one complete canvas whose dominant structure is two primary vertical
parts: the REALITY VIEW on the left and the TRANSFORMED DESIGN on the right.
Together they organize the composition from the top edge to the bottom edge;
all visible material, including typography, belongs within this left-right
structure. Let the image model determine asymmetric widths and each part's
internal crop, extension, whitespace and typography from the source, source
brief and final canvas.

Design-only:

OUTPUT MODE: DESIGN_ONLY

Create one full-canvas artwork entirely in the TRANSFORMED DESIGN language.
Use the REALITY VIEW only as the non-visible source of identity, structure,
relationships, colour logic and facts. Every visible visual element must be
part of the source brief's designed reinterpretation rather than an
untransformed presentation of the source photograph.

Wallpaper pack — one block per device asset:

OUTPUT MODE: WALLPAPER_PACK
DEVICE PROFILE: <resolved PHONE, IPAD, DESKTOP or WATCH>
WALLPAPER RELATIONSHIP: <resolved INDEPENDENT or LINKED>

Create one full-canvas wallpaper for this device entirely in the TRANSFORMED
DESIGN language. Use the REALITY VIEW only as a non-visible reference. Recompose
the artwork for this device's canvas and usable screen space; every visible
element belongs to the designed result.

Append exactly one text block after the selected mode block. If the user has other explicit requirements, append those verbatim after the text block at the very end.

Prompt-generated text:

TEXT MODE: ORIGINAL_PROMPT_GENERATED
TEXT LANGUAGE: <resolved language or locale>

The image model must generate any wording by following the original brief's
existing text-generation logic. Every visible word must arise naturally from
the current source image's content, atmosphere or implied meaning as interpreted
through that logic. Anything presented as factual or documentary information
must come from user-supplied, visibly readable or otherwise verified source
facts; when those facts are unavailable, use poetic non-factual wording. The
runtime shell is never a source of visible copy.

User-exact text:

TEXT MODE: USER_EXACT
TEXT LANGUAGE: <resolved language or locale>
TEXT: 「<user's exact characters>」

Use the supplied text verbatim. Do not rewrite, translate, spell-correct, or add
any other wording. Typography and placement still follow the original brief.

Text-free:

TEXT MODE: NONE
Render no letters, characters, numbers, logos, captions, labels, or pseudo-text anywhere.

If the user explicitly changes colour or another style variable, append that exact request at the very end, after the text block, and identify it as a user override. Do not elaborate it.

Immediately before generation, verify that the constructed prompt contains exactly one OUTPUT MODE: line and no instructions from an unselected mode.

Bitmap execution

• Prefer GPT Image 2 whenever the host's built-in image capability or an already configured compatible route exposes it.
• Also support Seedance 5.0 Pro, Nano Banana Pro (Gemini Image Pro), Nano Banana 2 (Gemini Image Flash), or another compatible bitmap model only when it can preserve the reference, requested canvas, language, and multi-reference needs. Changing the model must not change this workflow or the prompt-authority boundary.
• Judge readiness by an actual image-capability check, not one missing tool or environment variable. Never name or expose a provider, endpoint, credential, header, account, route, or secret.
• If the built-in image tool is exposed, follow the installed imagegen Skill and make one generation/edit call per distinct asset. A wallpaper pack requires four outputs, not one contact sheet.
• If built-in generation is not exposed, run python3 scripts/configured_imagegen.py probe. Its sanitized result is the only allowed diagnostic surface. If ready, use edit with the source and a private prompt file; use generate only when no image reference genuinely exists.
• Do not echo a credential or full prompt in shell history or logs. Do not inspect configuration files manually. Do not modify provider, account, billing, credential, or global routing settings unless the user explicitly requests that separate change.
• If no compatible route is verified, ask the user to enable a suitable image tool or voluntarily provide an API key for the task. Never ask them to expose an existing secret, and never assert that a key is missing without a trusted sanitized result.
• Generate finished raster imagery. SVG, HTML, CSS, Canvas, diagrams, and programmatic vector drawing are not substitutes.

Use one complete-canvas generation per output. Give the image model the source, complete source-brief body, common preamble, one selected mode block, one text block and final canvas in one request; do not pre-crop the source or generate separate panels. The model decides how to preserve the important subject and context inside the selected mode contract. Retry a failed complete canvas once by restating only a meaningful failed content or delivery constraint. Use scripts/compose_panel.py only when the user explicitly requests pixel-exact panel geometry or pixel-identical source preservation. The script may crop/paste/size/audit raster files; it must never invent the artwork or judge style.

For a linked wallpaper pack, generate one device image first as the visual anchor, then supply both the original source and that approved anchor to each remaining device generation. Recompose each canvas; never crop an earlier wallpaper or chain derivatives. For an independent pack, every device sees only the original source.

Source and output isolation

Use only sources attached to the current invocation, explicit image paths, an explicitly supplied input directory, or a prior source explicitly identified by the user as “the same image.” Scan a directory only under the directory-batch rules above; never scan Desktop, workspace roots, output folders, or unrelated directories for a substitute. Historical outputs and sample assets are not inputs unless explicitly named.

Write every selected final PNG directly inside the fresh task directory. The task directory is the only grouping layer:

~/Desktop/xxd/xxd-panel-028/<fresh-task>/
├── source-001-photo-name-top-bottom-3x4-1536x2048.png
├── source-001-photo-name-left-right-3x2-2160x1440.png
├── source-001-photo-name-design-only-3x4-1536x2048.png
├── source-001-photo-name-wallpaper-linked-phone-1440x3200.png
├── source-001-photo-name-wallpaper-linked-ipad-2048x2732.png
├── source-001-photo-name-wallpaper-linked-desktop-3840x2160.png
└── source-001-photo-name-wallpaper-linked-watch-1024x1024.png

Use zero-padded source-order prefixes plus a sanitized source stem for multiple inputs, for example source-001-street and source-002-flower. Follow with the mode, then a collision-safe size label; include wallpaper relationship and device for wallpaper files. If different relative paths share a stem, retain enough path context or a short collision suffix to keep names unique. Normalize ratio separators to x in filenames and include exact output pixels whenever they help distinguish requested variants. Do not create source, mode, size, or device subdirectories, and do not create empty directories for unselected modes.

--out replaces the root but not fresh-task isolation. Reserve a collision-safe task name before generation. Do not create an automatic collage, overview, mockup, or combined preview. Return absolute PNG paths in source order, then mode order 1→4; wallpaper order is phone, iPad, desktop, watch.

Acceptance gate

Inspect every final PNG at full size and thumbnail size. Accept only when:

• it is a fresh result from the correct current source or theme;
• mode, source visibility, whole-canvas ratio or exact pixels, count, and PNG format match the resolved runtime variables;
• the source remains recognisable and its important subject and context are not needlessly truncated, stretched or replaced;
• top-bottom reads as two primary horizontal parts, with the reality view above and transformed design below, while leaving their exact proportions to the image model;
• left-right reads as two primary vertical parts that organize the canvas from top to bottom, with the reality view left and transformed design right; typography is integrated into that structure rather than becoming a shared third band;
• design-only and every wallpaper contain a full-canvas transformed design: the reality source remains reference material, and no visible area presents it as an untransformed source photograph;
• no seam detection, midpoint percentage, coordinate measurement, or fixed-crop audit is used unless the user explicitly requested exact geometry;
• the result follows references/original-prompt/zh-CN.md, especially its own colour, material, composition, whitespace, and typography requirements, without an outer Skill palette or added art direction;
• prompt-generated text uses the requested language, follows the source brief's own text logic and is meaningfully rooted in the current source image; anything that reads as factual or documentary information is traceable to supplied, visible or verified facts; user-exact text is verbatim with no additions; text-free output contains no text or pseudo-text;
• linked and independent wallpaper rules are respected;
• for a directory batch, every discovered source appears exactly once in the queue, all requested outputs stay inside the one batch task directory, and the final report accounts for every success and failure;
• no SVG/code-rendered substitute, watermark, UI, route information, or secret appears.

When a result fails, retry only the failed source-brief or runtime requirement. Do not “improve” it by introducing a new aesthetic theory.

Runtime adapters

• references/original-prompt/README.md — language index for the canonical original and four faithful reading translations
• references/runtime-preferences.md — safe cross-invocation reuse, edit, fresh-start, and opt-out contract
• references/xxd-panel-028-prompt.zh-CN.md
• references/xxd-panel-028-prompt.en.md

These files document the same minimal adapter. They never replace references/original-prompt/zh-CN.md.

🦁 XXD Panel 028｜等距微景观

把照片安放成一件源图配色的等距纸上微缩景观

简体中文 · English · 日本語 · 한국어 · العربية

原始 X 样张

16:9 左右双联样张

3:4 上下双联样张

以下四张复用上方 16:9 样张对应的四张独立素材，重新生成完整 3:4 上下双联画布；上部保留现实摄影，下部遵循 028 原始提示词重构。英文配字只从当前照片的内容、情绪或隐喻中生成，并已逐张复核。

小小东（@xiaoxiaodong01） · 2026-08-20 GPT2 x 立体 x 模型 x 微缩 x 美学提示词 x VOL.028

这些样张用于展示 028 的美学动机，不会把样张中的主体、构图、配色、文案或旧画幅变成生成参考或当前默认值。

正交等距 · 微缩基座 · 源图限定色盘 · 精细墨线 · 纸上编辑插画

XXD Panel 028 是一个面向 Codex 与兼容 Agent 的图像生成 Skill。它把照片中最可辨认的身份、轮廓、姿态与关系概括为一件正交等距微缩模型或建筑沙盘，安放在一小块源图有依据的纸面基座上。

清晰体块、边缘裁切、轻微投影和极少量微型细节建立尺度感、孤独感与收藏感，周围保留大面积开放纸面。色彩完全从源图提取、归纳、提纯、柔化；精细墨线、扁平色面、轻影与极浅纸粒形成安静、克制的编辑插画质感。

适用场景与解决的问题

当照片需要用于海报、封面、社交内容或壁纸时，真正需要解决的通常不是简单换一种画风，而是如何在重新设计之后，仍然保留主体身份、关键关系、情绪与辨识度。

Panel 028 适合这类任务：需要保留原图核心内容，同时希望最终视觉达到以下方向：把照片安放成一件源图配色的等距纸上微缩景观。它先从用途与期望结果出发，再决定构图、尺寸、文字和具体输出形式。

适合这些情况

• 保留内容： 原图中的人物、物件、动作和关系不能被无关模板替换。
• 提升表达： 希望照片获得更明确的编辑设计、艺术语言或叙事重点。
• 灵活交付： 同一素材需要对照图、纯设计图、多种比例或一套壁纸。

快速开始

无需先阅读全部参数。发送一张图片给 Agent，并说明：

用 XXD Panel 028 帮我处理这张图，先推荐最合适的构图和尺寸。

上方样张可用于确认视觉方向；准备开始时，直接前往开始使用。完整模式、尺寸和参数收录在后面的折叠资料中，可按需查阅。

原始提示词 · 五种语言

打开统一的多语言目录： 简体中文原文 · English · 日本語 · 한국어 · العربية

简体中文文件保存小小东提供的逐字原文，并且是运行时唯一审美权威；其他四个版本是忠实的阅读译文，方便国际读者理解和转发，不会反过来改写生图提示词。

快速判断：XXD Panel 028 适合你吗？

它如何把你的输入变成独特成品

普通“等距微缩”很容易退化成繁忙玩具小镇、现成游戏资产或一套与照片无关的薄荷绿和珊瑚红。

028 的顺序完全相反：

锁定身份／轮廓／姿态／关系 → 概括为正交等距体块与有限层级 → 放在一块源图有依据的纸面基座 → 以边缘裁切、轻影与极少细节建立尺度 → 从源图归纳并柔化有限色盘 → 用精细墨线、扁平色面和浅纸粒完成

如果换成一张无关照片，微缩轮廓、体块层级、基座形状、尺度线索、综合色盘和文字关系仍然成立，这张图就不属于 028。

成品中最容易识别的风格特点

• 源图专属微缩： 至少三个身份、轮廓、姿态、动作、功能、开口、尺度或关系线索进入同一个微缩模型。
• 一块纸面基座： 主体落在源图有依据的纸片、低基座或微型地块上，周围保留大面积开放纸面。
• 克制尺度细节： 体量对比、边缘裁切、轻影和极少人物、栏杆、树、轨道或小物件建立尺度，不形成第二焦点。
• 正交等距： 相机干净、体块清楚、层级简洁，不做摄影透视、不可能空间、繁忙城市或游戏关卡。
• 源图限定色盘： 从照片中提取最有识别度的颜色，归纳、提纯、柔化并保留原有冷暖性格，绝不套固定粉彩。
• 纸上材质： 精细墨线、扁平哑光色面、同色阶、邻近色、轻微投影和极浅纸粒共同建立结构。
• 纤细编辑文字： 一个短标题和极少微字沿留白边缘、主体基座、水平轴线或等距结构排列。

原始提示词优先，而不是二次导演

references/original-prompt/zh-CN.md 是本项目唯一的创作与审美权威。Skill 不再额外总结或扩写它，也不会统一规划颜色、色板、美学动机、标题或微文案。原始提示词要求怎样处理颜色、材料、构图、留白与文字，GPT Image 2 就按那套逻辑执行。

模式与尺寸会完整替换原始提示词旧有的 3:4 上下交付容器，但不改写它的转译美学。每张成品只向 GPT Image 2 发送一个已选模式的最终契约，不再把四种模式放进同一个通用模板让模型自行猜测。

四种可组合输出模式

模式可以单选或多选：top-bottom、left-right、design-only、wallpaper-pack；多选时，每个模式分别生成、分别拥有自己的提示词。

• top-bottom：整张画面由上方现实画面与下方设计转译两个主要部分组成。
• left-right：整张画面从顶部到底部保持完整左右结构，原图在左、设计在右；文字融入这两个区域，不再生成横跨全宽的第三块底栏。左右宽度可以不对称，由模型自主构图。
• design-only：原图只作为不可见的身份、结构、颜色与事实依据；整张成品的每个可见元素都必须属于该 Skill 自己的设计转译语言。
• wallpaper-pack：每台设备分别生成完整画布的设计转译壁纸，原图不作为照片区域出现。

上下或左右都不做分界线、中线百分比或像素坐标检测。只有用户明确要求像素级分区或原片逐像素不变时，才使用确定性拼合。

普通成品尺寸同样可以多选：自动适配、跟随原图、1:1、3:4、4:3、4:5、5:4、2:3、3:2、9:16、16:9、21:9、5:7、7:5，或自定义比例／准确像素。没有静默默认尺寸；每个不同比例都会基于同一份原始提示词独立重构。

壁纸套装可选“连贯”或“四张独立”。连贯模式先生成一张定调图，其余设备同时参考原图与定调图重新构图；不会把一张图机械裁成四种尺寸。

每次调用只建立一个任务目录，所有最终 PNG 直接放在该目录中，不再按来源、模式、尺寸或设备创建子文件夹。文件名使用 source-序号-原文件名-模式-比例-像素.png；壁纸额外包含连贯关系与设备，例如 source-001-street-left-right-3x2-2160x1440.png、source-001-street-wallpaper-linked-phone-1440x3200.png。

直接传入一个图片目录时，这个士兵 Skill 会自动批量处理目录中的图片，并让每张图片都使用 Panel 028 的同一套原始审美。默认递归扫描常见位图，先报告数量，再统一确认一次模式、尺寸、文字和语言；每张图片独立生成，互不借用内容或文案。整批仍然只创建一个扁平任务目录，单张失败会被记录，但不会静默跳过后续图片。

文字方式

正式生图前只确认三种选择：

1. 模型根据原始提示词生成文字：用户只指定语言或地区，文字内容、数量、气质与排版由 GPT Image 2 按原始提示词生成；所有文字都从当前图片的内容、气质或隐喻中自然生长，任何事实或资料式信息都必须有用户输入、图片可见内容或已核实信息作为依据。
2. 使用我的准确文字：逐字传给图像模型，不改写、不翻译、不补标题；排版仍遵循原始提示词。
3. 不要文字：严格禁止文字与伪文字。

外层 Skill 不再预编标题、微文案或文案包。文字语言与操作语言分开确认，不根据人物、场景或文件名猜测国家与受众。

完整画布优先与位图边界

图像模型负责整张成品的审美重构，双联也默认一次直出完整画布。scripts/compose_panel.py 只保留为条件明确的兜底、无创尺寸校准和只读审计工具，不再预先规划每次任务，也不评价审美是否成功。

全部交付为 PNG 位图。每次调用都在 ~/Desktop/xxd/ 下创建新任务；已配置图像通道只返回脱敏状态，不公开供应商、端点、凭据、请求头、提示词、响应或账户信息。SVG、HTML、Canvas、图表和程序绘图不能替代最终作品。

能力自适应问询与快捷参数

同一个 Skill 会根据宿主真正提供的交互能力选择界面，不会把文本符号伪装成可点击控件：

• Claude Code 提供 AskUserQuestion + multiSelect: true 时：模式和尺寸使用真正的 checkbox；文字方式与壁纸关系使用单选。常用尺寸会按方形、竖版、横版分组展示，并累计多组选项；自定义尺寸进入自由输入。
• Codex 只提供 request_user_input 时：它只用于文字方式、壁纸关系等互斥单选，不拿来伪装模式或尺寸多选。模式与尺寸改用清楚的组合输入。
• 没有交互工具时：使用两轮文字问询。第一轮选择一个或多个模式；第二轮填写尺寸与文字方式。Skill 不显示假的 - [ ]，也不会为了获得表单要求用户切换 Plan mode。

默认第二轮只展示“智能推荐／跟随原图／常用比例／自定义”四个入口；只有选择常用比例时，才展开完整比例库：方形 1:1，竖版 3:4、4:5、2:3、9:16、5:7，横版 4:3、5:4、3:2、16:9、21:9、7:5。所有比例都可组合，也可直接输入准确像素。

全部设置都可以直接作为参数传入：

/xxd-panel-028 photo.jpg --mode top-bottom,design-only --size auto,3:4,9:16 --text prompt --locale ja-JP

支持 --mode、可重复或逗号分隔的 --size、--text prompt|exact|none、--locale、--copy、--wallpaper linked|independent、--wallpaper-size、--out 和 --prefs。参数齐全时跳过全部问询；参数不完整时只询问缺失项。

继承上次偏好，或全新配置

检测到历史偏好时，Skill 会先展示上次的模式、尺寸、文字与语言、壁纸关系和输出位置，再提供三个互斥选项：直接沿用、沿用并修改、全新配置。本次明确说出的要求始终优先；完整参数仍然直接执行。

可使用 --prefs last|edit|new|off|clear 直接指定行为：

/xxd-panel-028 photo.jpg --prefs last
/xxd-panel-028 photo.jpg --prefs edit
/xxd-panel-028 photo.jpg --prefs new

偏好只保存交付设置，不保存原图、准确文案、生成结果、Panel 选择、模型通道、API 凭据或其他敏感信息。

photo.jpg 也可以直接替换为图片目录路径；目录输入会自动进入批量，不需要额外的 --batch 开关。

参数速查

photo.jpg 可以替换为本地图片路径、已上传图片、图片目录或用户明确给出的其他来源。目录会自动进入批量处理。

按用途复制命令

# 上下对照：智能推荐画幅，模型按原始提示词生成简体中文文字
/xxd-panel-028 photo.jpg --mode top-bottom --size auto --text prompt --locale zh-CN

# 左右对照：3:2 横版，无文字
/xxd-panel-028 photo.jpg --mode left-right --size 3:2 --text none

# 只要设计图：9:16 手机竖版，日语文字
/xxd-panel-028 photo.jpg --mode design-only --size 9:16 --text prompt --locale ja-JP

# 四端连贯壁纸：先建立一张定调图，再按设备分别重构
/xxd-panel-028 photo.jpg --mode wallpaper-pack --wallpaper linked --text none

# 四张独立壁纸：同时指定四台设备的准确分辨率
/xxd-panel-028 photo.jpg --mode wallpaper-pack --wallpaper independent --wallpaper-size phone=1440x3200,ipad=2048x2732,desktop=3840x2160,watch=1024x1024 --text none

# 跟随原图比例
/xxd-panel-028 photo.jpg --mode design-only --size source --text none

# 同一用途生成方形、竖版和横版三种完整构图
/xxd-panel-028 photo.jpg --mode design-only --size 1:1,3:4,16:9 --text none

# 自定义比例与准确像素可以一起多选
/xxd-panel-028 photo.jpg --mode design-only --size 11:14,2160x3840,3840x2160 --text none

# 模型依照原始提示词决定文字，只限定语言
/xxd-panel-028 photo.jpg --mode design-only --size 3:4 --text prompt --locale en-GB

# 使用准确文案，不改写、不翻译、不补标题
/xxd-panel-028 photo.jpg --mode design-only --size 3:4 --copy "让风经过这里" --locale zh-CN

# 两种模式 × 三种尺寸，共生成六张独立完整构图
/xxd-panel-028 photo.jpg --mode top-bottom,design-only --size 1:1,3:4,9:16 --text prompt --locale zh-CN

# 整个目录统一使用 Panel 028；公共设置只解析一次
/xxd-panel-028 "/path/to/photos" --mode design-only --size auto,9:16 --text prompt --locale zh-CN

# 重复参数也会累积，并把本次任务放到指定根目录
/xxd-panel-028 photo.jpg --mode top-bottom --mode left-right --size 3:4 --size 16:9 --text none --out ./deliveries

生图模型优先级

GPT Image 2 是默认首选，并继续执行本项目现有的高保真垫图、生成前确认整张画幅、双联一次生成完整画布、脚本仅作条件式兜底等逻辑。

当当前工具或已配置兼容通道确实可用，并能满足原图保真、整张成品比例、目标语言文字和连贯壁纸多图参考等要求时，也支持 Seedance 5.0 Pro、Nano Banana Pro（Gemini Image Pro）、Nano Banana 2（Gemini Image Flash）或其他兼容位图模型。备用模型只替换生成通道，不得改变模式、画幅、文案