krea2-prompt-artisan

name	krea2-prompt-artisan
description	Use when the user asks for a Krea2, Krea 2, K2, Krea.ai, "k2提示词", "K2化", "生成一段k2提示词", or "将下面的提示词K2化" image-generation prompt, or asks Codex to turn a text idea or reference image into a Krea 2-ready prompt; always create a local HTML deliverable with copy controls by default.

Krea2 Prompt Artisan

Use this skill to turn a text idea, detailed visual brief, or reference image into a high-density Krea 2 / Krea.ai Turbo prompt. Produce a polished English prompt, a faithful Chinese translation, and always render a local HTML page for reading, copying, the one-line punctuated poem, and the Codex footer signature.

Workflow

1. Identify the input type:

   • Text-only idea or brief: work directly from the user's description.
   • Reference image: first analyze the image with Codex's available vision capability. For images pasted directly into the Codex chat, use the attached image itself; do not depend on QQ/WeChat/cache filesystem paths. Do not copy pasted images into the workspace unless the attachment is not visible, the resolution is insufficient, or file-based tooling is truly required. If a temporary image file is unavoidable, save it under the current workspace or the system temporary directory, use it, then delete it after verification.

2. Read references/krea2_system_prompt.md before generating the prompt unless the current conversation already contains those rules.

3. Generate one continuous English prompt for Krea 2:

   • Write natural prose, not a labeled checklist.
   • Default to at least 500 English words; prefer 650-850 words when the user did not ask for a shorter prompt.
   • Cover subject details, intended use and aspect ratio, environment, composition/camera, lighting, style/medium, materials/colors, and quality safeguards.
   • Include negative constraints naturally, such as no text, no logo, no watermark, no malformed anatomy, and no extra fingers when relevant.

4. Provide a faithful Chinese translation:

   • Match the English content closely.
   • Do not compress a long English prompt into a short Chinese summary.
   • Preserve every important visual detail.

5. Prepare display metadata:

   • title_cn or chinese_title: mandatory full Chinese title name. Use a concise Chinese title, ideally 6-18 Chinese characters and usually no more than 24. Do not abbreviate it in filename_title. The renderer uses this same Chinese title name in two places: the visible page title is Krea 2 提示词 - 标题名 with a relevant emoji prefix, such as 🎬 Krea 2 提示词 - 红妆庭院贵族仕女; the file name is Krea2_prompt_标题名.html, such as Krea2_prompt_红妆庭院贵族仕女.html. The 标题名 part must be identical in both places, fully Chinese, and contain no English letters. Also choose a suitable title_emoji from the prompt content when possible, such as flowers, night, ocean, city, forest, books, coffee, cinematic, or fantasy motifs.
   • scene_name: optional short display scene name; prefer Chinese when used, but do not use it as a shorter replacement for title_cn.
   • intro_text: one or two sentences summarizing the user's request and the chosen visual direction.
   • breakdown_rows: eight rows for subject details, canvas/use, environment, composition/camera, lighting/tone, style/medium, materials/colors, and quality safeguards. Each row must include non-empty design reasoning via rationale or a supported alias such as logic, design_logic, consideration, reasoning, why, note, or 设计考虑. Use the canonical English slot names or explicit slot_cn/slot_en; the HTML renderer displays the module column as Chinese plus English, with the Chinese label in a colorful accent and English in body-muted color. The renderer has a fallback, but agents should still provide real row-specific rationale.
   • optimization_notes: short notes explaining important prompt-design choices. When Advanced Stylization Mode is used, briefly state the cross-domain mapping source and which generic default aesthetics were removed.
   • poem: 七言绝句单行铁律：{{POEM}} 必须现场原创一首应景贴题、绝不重复的七言四句诗。必须恰好 28 个中文正文字符（标点和首尾 🌸 不计入字数），四句每句七字。为了排版精美，整首诗必须写在同一行（不使用 <br> 换行），每句加规范中文标点，推荐节奏为“七字，七字。七字，七字。”，首尾只用 🌸 符号包裹；诗句正文不得出现其他 emoji、图标或装饰符号。不满足 28 字或含有额外符号时渲染器会拒绝生成 HTML。
   • : use for the HTML footer signature.

6. Always render the HTML deliverable:

   • For every completed Krea2/K2 prompt task, run scripts/render_krea2_html.py with the generated data. Do not wait for the user to ask for HTML.
   • By default, the script writes .html files under the current Codex working directory:

<current working directory>\Krea2提示词

 This resolves to a project-local output folder named `Krea2提示词` under the current working directory. The renderer creates this directory automatically if it does not exist. For a custom location, pass `--output-dir <path>` or set the `KREA2_OUTPUT_DIR` environment variable.

• Use this JSON schema for the renderer payload: title_cn or chinese_title (mandatory full Chinese page/file title), optional matching filename_title, scene_name, optional title_emoji, intro_text, prompt_text for the English prompt, translation_text for the Chinese translation, breakdown_rows, optimization_notes, poem, and optional signature. The renderer also accepts scene_name_cn, english_prompt, and chinese_prompt as aliases, but prefer title_cn, prompt_text, and translation_text.
• Do not open the page during initial rendering. Render first without --open, inspect the generated HTML for title, prompt text, translation text, non-empty breakdown rationales, poem, signature, encoding, and absence of direct .json leftovers. Only after these checks pass, open the already-created file as the final step with scripts/render_krea2_html.py --open-existing <html path>.
• Because opening starts the user's default browser, run the final --open-existing command with escalated shell permissions when available. In the Codex Windows sandbox, a normal run can return opened: false or WinError 5. If that happens, rerun only the open-existing command with escalation instead of silently accepting no popup.
• Omit the final open step only when the user explicitly says not to open the browser.
• Prevent encoding damage: do not pipe JSON through a default Windows PowerShell native pipeline unless UTF-8 is set first. Prefer passing a UTF-8 JSON file with --input, or set $OutputEncoding=[Console]::OutputEncoding=[System.Text.UTF8Encoding]::new($false) and $env:PYTHONUTF8='1' before piping JSON to Python. The renderer writes HTML as UTF-8 with BOM and refuses to generate a page if Chinese/emoji fields have already turned into ?? question-mark mojibake.
• Do not create .md or any Markdown file as the final deliverable; Markdown may only be an internal draft.
• Do not leave payload .json files directly in the Krea2提示词 output directory. Prefer piping JSON through stdin. If a .json payload file is passed with , the renderer deletes only that input file after successful HTML generation, and only when it is a direct child of the output directory. It does not delete unrelated files and does not recurse into subfolders.

Output Expectations

In chat, keep the response concise:

• State that the Krea 2 prompt is ready.
• Include the created .html file path and whether it was opened in the browser.
• Treat the HTML file as the primary deliverable; do not stop after chat text or a Markdown document.
• Include the English prompt and Chinese translation directly only when the user asks for them in chat.

Boundaries

• Do not modify external application workspaces, model settings, Python runtimes, node_modules, or unrelated workspace files.
• Do not write output to another application's workspace by default.
• Do not call unrelated image-generation tools or hard-code provider-specific model IDs.
• Do not generate the final image unless the user separately asks for image generation.