Portrait Script Cover

---
name: "portrait-script-cover"
description: "Create or revise a Chinese short-video cover from a spoken script, headline, and the creator's own portrait photo. Use this skill whenever a user asks for a Douyin/TikTok China, Xiaohongshu, Reels, or vertical video cover that should feature their real face and vary with each topic. Require a clear portrait upload before generating; choose or recommend one of five built-in visual styles from the content, then derive the pose, expression, outfit, diagram, and composition instead of reusing one template."
compatibility: "Requires Codex image_gen and view_image. Canva is optional for a final editable import."
---

# Portrait Script Cover

Create a recognizable series without turning every episode into the same poster. The script determines the visual argument; the visual system makes the account recognizable.

## Required inputs

Extract inputs already present in the conversation. Do not ask for them twice.

- **Portrait photo — required.** Before generating, the user must upload a clear, front-facing photo of the person who should appear on the cover. If no such image is available, stop and ask for one. Do not use a stranger, stock portrait, or a previous user's photo.
- **Script or topic — required.** Ask for a spoken script, outline, or sufficiently concrete topic.
- **Headline — preferred.** Use the exact user-provided headline. If none is available, propose three concise Simplified Chinese headlines and wait for a choice.
- **Optional:** account positioning, prior covers, a style reference, a desired CTA, and a Canva editable version.

Treat the uploaded portrait only as a reference for the current task. Do not save it into a skill folder, repository, public location, or reuse it for another person.

Repository demo images, if present, are synthetic fictional examples only. Never use them as a substitute for the current user's required portrait photo.

## Content brief

Before prompting the image model, infer these five fields internally:

1. **Tension:** the misconception, problem, risk, or opportunity that makes someone stop scrolling.
2. **Claim:** the one idea the video wants the viewer to remember.
3. **Concrete scene:** the work situation, behavior, object, or process that makes the claim tangible.
4. **Visual argument:** a hand-drawn metaphor that explains the claim: broken path, loop, comparison, checklist, map, bottleneck, question cloud, ladder, funnel, or before/after.
5. **Human state:** the creator's appropriate expression and gesture: skeptical, concerned, curious, explaining, warning, demonstrating, or confident.

Do not place the whole script on the cover. A cover needs a headline and one visual argument, not a transcript.

## Five built-in visual styles

Use one style per cover. If the user names a style, follow it. Otherwise recommend one based on the script before generating. Do not blend styles unless the user explicitly asks: a clear style decision produces a more recognizable cover.

| Style | Best for | Visual direction | Avoid |
| --- | --- | --- | --- |
| **手绘旧纸水墨** | 概念解释、知识拆解、业务逻辑 | 暖米白旧纸、淡网格、深蓝手绘线稿、蓝橙水墨和粗手写标题 | 复制任何具体博主的画面、做成企业 PPT |
| **蓝图工程** | SOP、流程改造、自动化、系统落地 | 深普鲁士蓝图纸、白色技术线稿、工程网格、单一安全橙强调 | 羊皮纸、水彩、游戏 UI、无意义的复杂参数 |
| **杂志拼贴** | 反常识、踩坑、强观点、批评性内容 | 象牙新闻纸、撕纸边缘、复印颗粒、剪贴卡片、深蓝与朱红批注 | 杂乱的报纸正文、品牌 Logo、无关装饰 |
| **孔版印刷** | 个人观点、趋势判断、轻量方法论 | 黄米纸、钴蓝和珊瑚橙双色套印、网点、错位油墨和手绘箭头 | 伪复古宣传画、大量颜色、不可读标题 |
| **暗黑未来科技** | AI 趋势、产品发布、系统能力展示 | 近黑海军蓝、冷白和青蓝光、克制橙色信号、抽象网络与清晰路径 | 霓虹城市、黑客、机器人、游戏界面、赛博朋克杂乱感 |

All styles use a 3:4 vertical composition and preserve the creator's real face. This is a style library, not a request to imitate a particular creator, composition, phrase, or artwork.

## Make every episode visibly different

The title is never the only thing that changes. Choose a combination that reflects the content, and avoid matching the most recent covers visible in the conversation.

| Content intent | Expression and action | Outfit direction | Diagram/metaphor |
| --- | --- | --- | --- |
| Call out a mistake or risk | Concerned, inspecting, looking down, warning gesture | Charcoal jacket, dark overshirt, gray-blue layer | Broken path, warning note, hole, faulty loop |
| Explain a concept | Thoughtful, looking upward, chin-rest, pointing at a concept | Khaki shirt, dark green knit, soft gray blazer | Question cloud, puzzle, causal map, magnifier |
| Teach a method or SOP | Demonstrating, pointing at steps, holding notes, drawing | White shirt with dark layer, olive work shirt | Checklist, three-step map, route, workbench |
| Show a result or case | Confident, light smile, presenting, upward gesture | Navy knit, warm brown jacket, simple blazer | Upward curve, before/after, dashboard, result card |
| Challenge conventional wisdom | Surprised, questioning, looking toward a contradiction | Denim jacket, caramel shirt, dark blue outerwear | Fork in the road, collision arrows, wave, question mark |

Vary at least two of: outfit, pose, expression, portrait position, title layout, and main diagram. These changes should have a reason in the script; random variation weakens the series.

## Workflow

1. Inspect the local portrait with `view_image`. Label it as an identity reference.
2. If a style reference exists, inspect it and label it as a style reference only. Do not copy its exact artwork or layout.
3. Build the content brief and choose the visual combination from the variation table.
4. Select a built-in style. If the script strongly maps to one style, state the recommendation in one line; otherwise ask the user to choose before generating.
5. Use `image_gen` to create a complete cover. Preserve the creator's facial identity and generate all headline text in Simplified Chinese.
6. Inspect the output. Check headline accuracy, facial identity, mobile readability, relevance of the diagram, and visual difference from recent covers.
7. If the result fails one check, iterate once with a prompt focused only on that defect.
8. Save the selected PNG in the current task's `outputs/` folder. If the user asks for editable output, import the final image to Canva; if import fails, still deliver the PNG and report the connection failure plainly.

## Image generation prompt template

Use this compact structure. Replace every bracketed value; do not show brackets in the final prompt.

```text
Use case: style-transfer
Asset type: 3:4 vertical Chinese short-video cover.
Input images: Image 1 is the creator's identity reference. Preserve the person's recognizable face, hairstyle, glasses or other stable features. Image 2, if present, is visual-language reference only.
Primary request: an original cover about [tension and claim].
Selected style: [one of the five built-in visual styles above]. Follow its visual direction and avoid list.
Subject: the creator wears [outfit], has a [expression] expression, and is [gesture/action] in [portrait position].
Scene/backdrop: warm aged off-white squared notebook paper, tactile grain, navy ink sketches, blue and burnt-orange watercolor dry-brush accents.
Hand-drawn visual argument: [concrete scene and diagram/metaphor].
Style/medium: Chinese editorial poster, ink sketch plus watercolor wash plus semi-realistic portrait, handmade and tactile; not a glossy corporate UI or flat vector infographic.
Composition: reserve [title area] for the headline. Portrait and diagram must not cover the title.
Text (verbatim, large Simplified Chinese only):
"[headline with intended line breaks]"
Render "[key phrase]" in burnt-orange rough brush lettering. Render all other headline text in deep navy brush lettering.
Constraints: exact Chinese headline, preserved facial identity, original composition, clear phone-grid readability, visibly different pose/outfit/diagram from recent covers.
Avoid: English, logos, watermark, generic stock portrait, duplicated title, wrong Chinese characters, copied creator-specific artwork.
```

## Quality gate

Do not deliver until all applicable checks pass:

- The Chinese headline is character-for-character correct and readable at thumbnail size.
- The portrait clearly resembles the uploaded person.
- The selected style is visually clear and the main diagram explains this script, not a generic AI topic.
- The cover differs from the user's recent covers in at least two visual dimensions.
- There is no extra English, watermark, logo, or copied artwork.

## Delivery

Return the PNG link and one short line describing the chosen visual argument. Include a Canva link only if the editable import succeeded.