The original brief is authoritative

references/original-prompt/zh-CN.md is this project's sole creative and aesthetic authority. The Skill no longer summarizes or expands it, and it does not impose a shared palette, colour plan, aesthetic motive, title, or microcopy package. GPT Image 2 follows that brief's own rules for colour, material, composition, whitespace, wording, and typography.

Mode and size completely replace the legacy 3:4 top-bottom delivery container without rewriting the transformation aesthetic. Each asset sends GPT Image 2 one selected mode's final contract instead of asking it to interpret four alternatives inside a generic template.

Four combinable output modes

Select one or more of top-bottom, left-right, design-only, and wallpaper-pack. When several are selected, each is generated independently with its own prompt.

• top-bottom: one complete canvas with the reality view above and transformed design below.
• left-right: one complete canvas whose left-right structure runs from top edge to bottom edge, source left and design right. Typography stays inside that structure rather than creating a shared third footer; the outer regions are strictly 50:50; only each region’s internal crop, scale, whitespace and typography may vary.
• design-only: the source is a non-visible reference for identity, structure, colour logic, and facts; every visible element follows this Panel's transformation language.
• wallpaper-pack: each device receives an independently composed full-canvas transformed wallpaper, with no source-photo region.

There is no seam, midpoint-percentage, or pixel-coordinate test. Deterministic assembly is used only when the user explicitly requests exact panel geometry or pixel-identical source preservation.

Ordinary sizes are also multi-select: auto-fit, source aspect, 1:1, 3:4, 4:3, 4:5, 5:4, 2:3, 3:2, 9:16, 16:9, 21:9, 5:7, 7:5, or custom ratios/exact pixels. There is no silent default. Every distinct aspect is independently recomposed from the same verbatim source brief.

Wallpaper packs may be linked or independent. A linked pack creates one anchor image, then recomposes each remaining device from the original source plus that anchor; it never crops one image into four sizes.

Each invocation creates one task directory and writes every final PNG directly into it, with no source, mode, size, or device subfolders. Filenames include source order and a sanitized source name, for example source-001-street-left-right-3x2-2160x1440.png and source-001-street-wallpaper-linked-phone-1440x3200.png.

Pass an image directory directly to batch-process it with this same Soldier. Every discovered image keeps Panel 059's one original aesthetic. The Skill recursively inventories common raster formats, reports the count, resolves shared mode/size/text/locale settings once, and then generates each source independently without content or copy leaking between items. The whole batch still uses one flat task directory; an individual failure is recorded without silently dropping later images.

Text modes

Before generation, resolve one of three choices:

1. Model generates text from the original prompt: the user supplies only the language or locale; GPT Image 2 follows the source brief's wording, amount, tone, and typography logic. Every word arises from the current image's content, atmosphere, or implied meaning, and anything presented as factual or documentary information must be grounded in supplied, visible, or verified facts.
2. Use my exact text: pass it verbatim, without rewriting, translating, or adding a title; typography still follows the source brief.
3. No text: prohibit visible text and pseudo-text.

The outer Skill no longer pre-writes titles, microcopy, or copy packages. Output language is resolved separately from the interface language and is never guessed from a person, scene, or filename.

Complete-canvas first, raster-only delivery

The image model owns the aesthetics of the entire finished composition; paired layouts also default to one complete-canvas generation. scripts/compose_panel.py remains only for condition-based recovery, lossless pixel calibration, and read-only audit. It is not run pre-emptively and does not judge aesthetic success.

Every deliverable is a raster PNG and every invocation creates a fresh task under ~/Desktop/xxd/. The configured image route exposes sanitised status only—never providers, endpoints, credentials, headers, prompts, responses, or account details. SVG, HTML, Canvas, diagrams, and programmatic drawing are not substitutes for the final artwork.

Capability-adaptive questions and inline parameters

The same Skill adapts to the host's real interaction capabilities and never presents decorative symbols as clickable controls:

• When Claude Code exposes AskUserQuestion + multiSelect: true: modes and sizes use genuine checkboxes; text mode and wallpaper relationship use single-select. Common sizes are grouped into square, portrait, and landscape checkbox questions, selections accumulate across groups, and custom sizes use free input.
• When Codex exposes only request_user_input: use it only for mutually exclusive fields such as text mode and wallpaper relationship. Do not misrepresent modes or sizes as single-choice; collect them through clear combination input.
• With no interactive question tool: use two typed rounds—modes first, then sizes plus text. Never draw fake - [ ] boxes or ask the user to switch to Plan mode merely to obtain a form.

The second round initially shows only Smart recommendation, Source aspect, Common ratios, and Custom. Expand the full library only when requested: 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. Any ratios may be combined, and exact pixels are always accepted.

All settings can also be passed inline:

/xxd-panel-059 photo.jpg --mode top-bottom,design-only --size auto,3:4,9:16 --text prompt --locale ja-JP

Supported parameters are --mode, repeatable or comma-separated --size, --text prompt|exact|none, --locale, --copy, --wallpaper linked|independent, --wallpaper-size, --out, and --prefs. Complete parameters skip preflight; partial parameters trigger only missing questions.

Reuse the last preference or start fresh

When a valid preference exists and this invocation still has unresolved delivery settings, the Skill summarises the previous modes, sizes, text and locale, wallpaper relationship, and output location, then offers three mutually exclusive choices: Reuse, Reuse and edit, or Fresh configuration. Explicit current requirements always win, and complete parameters still run without redundant questions.

Use --prefs last|edit|new|off|clear to choose the route inline:

/xxd-panel-059 photo.jpg --prefs last
/xxd-panel-059 photo.jpg --prefs edit
/xxd-panel-059 photo.jpg --prefs new

The record stores delivery settings only. It never stores source images, exact copy, generated results, Panel choices, model routes, API credentials, or other sensitive information.

Replace photo.jpg with an image-directory path to enter batch processing automatically; no separate --batch switch is required.

Parameter quick reference

Replace photo.jpg with a local path, uploaded image, image directory, or another source explicitly supplied by the user. A directory enters batch processing automatically.

Copyable commands by use case

# Top-bottom comparison, auto canvas, prompt-generated British English
/xxd-panel-059 photo.jpg --mode top-bottom --size auto --text prompt --locale en-GB

# Left-right comparison, 3:2 landscape, no text
/xxd-panel-059 photo.jpg --mode left-right --size 3:2 --text none

# Design-only, 9:16 phone portrait, Japanese text
/xxd-panel-059 photo.jpg --mode design-only --size 9:16 --text prompt --locale ja-JP

# Linked wallpaper pack: establish one anchor, then recompose every device
/xxd-panel-059 photo.jpg --mode wallpaper-pack --wallpaper linked --text none

# Independent wallpapers with exact device resolutions
/xxd-panel-059 photo.jpg --mode wallpaper-pack --wallpaper independent --wallpaper-size phone=1440x3200,ipad=2048x2732,desktop=3840x2160,watch=1024x1024 --text none

# Follow the source image's aspect ratio
/xxd-panel-059 photo.jpg --mode design-only --size source --text none

# Create square, portrait, and landscape compositions
/xxd-panel-059 photo.jpg --mode design-only --size 1:1,3:4,16:9 --text none

# Combine a custom ratio with exact pixel targets
/xxd-panel-059 photo.jpg --mode design-only --size 11:14,2160x3840,3840x2160 --text none

# Let the original prompt generate wording; constrain only the locale
/xxd-panel-059 photo.jpg --mode design-only --size 3:4 --text prompt --locale en-GB

# Use exact copy without rewriting, translating, or adding a title
/xxd-panel-059 photo.jpg --mode design-only --size 3:4 --copy "Let the light stay" --locale en-GB

# Two modes × three sizes produce six independently composed assets
/xxd-panel-059 photo.jpg --mode top-bottom,design-only --size 1:1,3:4,9:16 --text prompt --locale en-GB

# Apply Panel 059 to a whole directory; shared settings resolve once
/xxd-panel-059 "/path/to/photos" --mode design-only --size auto,9:16 --text prompt --locale en-GB

# Repeated parameters accumulate; place the task under a chosen root
/xxd-panel-059 photo.jpg --mode top-bottom --mode left-right --size 3:4 --size 16:9 --text none --out ./deliveries

Image-model priority

GPT Image 2 is the default first choice. It keeps this project's established workflow: high-fidelity source reference, explicit whole-canvas selection before generation, one complete-canvas generation for paired modes, and scripted composition only as a conditional fallback.

Seedance 5.0 Pro, Nano Banana Pro (Gemini Image Pro), Nano Banana 2 (Gemini Image Flash), or another compatible bitmap model may also be used when it is actually available through the current tools or configured routes and can satisfy source fidelity, whole-canvas ratio, target-language text, and linked-wallpaper multi-reference requirements. An alternative changes only the generation route; it must not change modes, canvas, copy, locale, wallpaper relationship, or the complete-canvas-first strategy.

If no suitable route is available, the Skill asks the user to enable an image-generation tool or provide an API key. User-provided credentials may be used for the current task without being echoed, displayed, logged, or exposed. They are not persisted, and provider, account, billing, or global route configuration is not modified, unless the user explicitly requests that configuration change.