agentsclimarketplace

Laohan fengmianqiuzhi

Skill hanzhcn/laohan-skills/laohan-fengmianqiuzhi

基于秋芝2046实战模板和29种场景风格生成真人口播封面;固定输出3种视觉风格,每种生成3:4、4:3、16:9三个尺寸,共9张真实候选,最终选择1个风格并保留其3个尺寸,同时绑定真实头像、provider 与选择证据。Use when 用户说"生成封面提示词""做封面""封面""封面图""封面词"或工作流进入⑥。From its SKILL.md

Install
npx -y skills add hanzhcn/laohan-skills --skill laohan-fengmianqiuzhi

Assembled from the repository path, not quoted from the project. Check it against their README if it does not work.

One thing to look at

  • 11 stars11 stars. Stars are a popularity signal and not a quality one, but at this level it is likely that nobody has read this closely except its author, and you would be relying on your own review.

SKILL.md

14.5 KB, ~4.8k tokens by cl100k_base, as published. Nobody here has run it

封面 brief 与提示词执行器

以 v1.0 经实战使用的秋芝2046模板为默认创作基线,完整恢复标题拆解、六层英文 prompt、CENTER人物写法、头部比例、白底大字、角色/动作/道具/场景差异和29种风格库;在此基础上继续扩充,不以“自由发挥”替代已有方法。v1.2 的头像 reference、provider 与选图证据合同继续保留。

边界

  • 输入真源:01-口播稿.mdepisode-config.jsoncanvas 与已锁定 distribution_contract
  • 输出:05-封面/cover-prompts.md,固定3个策略真正不同的视觉风格;每种分别提供3:4、4:3、16:9完整 prompt,共9个候选。
  • 不输出:假图片、selected-cover.json、CTR 结论或“某风格一定爆”。
  • REVIEW_GATED:最终选图留给 Jeffrey。
  • AUTONOMOUS_RUN:可由 agent proxy 依据缩略图可读性与稿件一致性选择;这仍不是发布效果证明。

Episode 前置 gate

  1. episode-config.schema_version 为2。
  2. distribution_contract.primary_platform 已包含在 platforms
  3. master_aspect_ratio 与 canvas 一致,locked_at 是有效 ISO-8601。
  4. 01-口播稿.md 标题已锁定。

任一项失败就停在⑥,不自行改画布或猜发布平台。

原版创作流程

先完整读取:

  • references/qiuzhi_analysis.md:人物比例、平台差异和已观察到的视觉机制;
  • references/styles.md:29种角色/场景方法库。它们可以直接使用、混合或扩充,不得压缩成少量固定标签。
  1. 读取口播稿,提取核心主题、品牌/产品名、关键词和视觉意象。
  2. 设计3种视觉上完全不同的画面故事。差异必须来自角色、环境、动作和夸张主道具,不是只换服装或背景色。
  3. 把口播稿标题拆成三层封面大字:
    • TOP-CENTER横式句子:优先使用完整标题,红描边白字;标题过长时只做不改变含义的断句,不自造抽象概念。
    • MID-LEFT关键词:2—5字,突出主题、对象或问题。
    • MID-RIGHT关键词:2—5字,突出结果、冲突或情绪。
    • 具体品牌/产品名不能丢;陌生人只看封面也要知道视频讲什么。
  4. 推荐与本期最匹配的V1,V2/V3保留为视觉反差备选。创新方向必须在完成这3个原版方向后再增加,不能用创新名义删掉原版输出。

好的大字来自标题并能独立看懂;”贴墙站””拆标签””反哺飞轮”等脱离标题的自造概念不能替代主题。

好的拆分示例(从标题拆出,层次清晰):

  • 标题”每次立flag两天就倒?真不是你懒” -> 横式”每次立flag两天就倒?真不是你懒” | 左”改变失败” | 右”不是你的错”
  • 标题”月薪两万五被AI取代,法院判了” -> 横式”月薪两万五被AI取代,法院判了” | 左”被AI开除” | 右”真话扎心”
  • 标题”花大钱补课,考场都没了” -> 横式”花大钱补课,考场都没了” | 左”AI拆考场” | 右”还在补课?”

REFERENCE_REQUIRED 是执行合同,不是提示语。frontmatter 必须同时写 image_providerreference_mode: REQUIRED、本期内 reference_assetreference_sha256;reference 必须复制到 05-封面/reference/。provider 调用时必须把该文件作为 reference image 参数传入,不能只在 prompt 里说“参考头像”。

头像只锁定人物身份和真实面部特征,不锁原照片的服装、姿势、表情、背景或光线;这些按每个视觉世界重新设计。

原版英文提示词模板

每条 prompt 必须保留以下六层结构,只替换方括号内容:

A [aspect ratio and orientation] video thumbnail cover image.

LAYOUT: Large close-up head shot with symmetrical text framing.
- TOP-CENTER: Horizontal sentence text "[完整标题]" in white (#FFFFFF) with thick bright red (#FF3333) outline, spanning the width above the person's head
- MID-LEFT: Large bold cartoon-style Chinese text "[左关键词]" in bright yellow (#FFD700) with thick black outline
- MID-RIGHT: Large bold cartoon-style Chinese text "[右关键词]" in bright yellow (#FFD700) with thick black outline, mirroring the left side
- CENTER: [完整场景描述]
- BOTTOM: Tag text "寒武纪说AI" in white with black outline

BACKGROUND: Clean white (#FFFFFF) or very light gray (#F5F5F5), minimal and flat. No gradient, unless this candidate deliberately uses one of the expanded scene worlds and clearly preserves thumbnail readability.

COLOR SCHEME: White/light background + white text with red outline + bright yellow text with black outline + colorful costume and exaggerated props.

TYPOGRAPHY: Extra bold cartoon-style Chinese font, very thick stroke outline on all text, rounded and playful, highly readable at thumbnail size.

STYLE: Photorealistic person with cosplay costume and exaggerated props. High-energy storytelling vibe. The person must look like a real photograph, NOT cartoon, NOT anime, NOT illustration.

DO NOT include: Watermarks, logos, QR codes, social media UI. Do not alter the person's identity or make the person cartoonish.

模板第一行 A [aspect ratio and orientation] video thumbnail cover image. 是占位符,按候选画布替换为具体写法:

  • 3:4 竖版 -> A vertical (3:4 aspect ratio) video thumbnail cover image.
  • 4:3 横版 -> A horizontal (4:3 aspect ratio) video thumbnail cover image.
  • 16:9 宽屏 -> A widescreen (16:9 aspect ratio) video thumbnail cover image.

这套模板是默认基线,不是待删除的”模板味”。若某个扩展方向需要深色背景、非对称构图或完整场景,可以有意识突破其中一项,但必须保留同等强度的标题可读性、人物识别和故事冲击,并在方向说明里写明突破理由。

CENTER写法

CENTER必须包含:

  • The person in the reference photo, large close-up head and upper body dominating the frame, head takes up about 25-30% of image height
  • 与本期主题直接相关的表情、姿势、服装、动作、夸张主道具和场景互动
  • Keep exact facial appearance from the reference photo.

三种方向必须是三个不同故事。同一方向适配不同平台比例时,CENTER也要改变姿势、道具互动和场景细节,不能只替换比例字符串。

完整 CENTER 示例(饺子馆 - 侦探风):

The person in the reference photo, large close-up head and upper body dominating the frame, head takes up about 25-30% of image height, leaning forward with one hand on the table, squinting with curiosity. wearing a detective trench coat with a name tag that reads "AI探店员", using a magnifying glass to inspect a giant steaming bowl of dumplings where tiny glowing circuit board patterns emerge from the steam. A phone showing a GitHub skill page floats nearby. Keep exact facial appearance from the reference photo.

完整 CENTER 示例(三文件 - 侦探风):

The person in the reference photo, large close-up head and upper body dominating the frame, head takes up about 25-30% of image height, standing with arms crossed and one eyebrow raised, confident smirk. wearing a detective trench coat with glowing blue tech trim, three floating holographic file cards orbit around the person - each card pulses with different colors (blue, green, purple) showing "CLAUDE.md" "claude.json" "ECC hooks". Keep exact facial appearance from the reference photo.

注意:同样是"侦探风",但姿势、道具互动、场景细节完全不同。

29种风格与扩充

references/styles.md 中29种风格是可直接调用的成熟库,不是只读灵感。每期先从库中挑选、混合或变体出3个强方向;若主题适合库外新场景,再把新方向追加进候选和风格库评估,不以新方向替代原有探索数量。

完整 prompt 合同

每条 prompt 必须可直接交给 image provider,并明确画布比例、三层文字、人物、场景、动作、服装、主道具、构图、背景、配色、字体、photorealistic约束和禁止项。若任一 prompt 只替换标题就能用于另一选题,必须重写CENTER故事。

输出格式

Episode mode 固定写 05-封面/cover-prompts.md

---
schema_version: 1
script_hash: <01-口播稿.md sha256>
distribution_locked_at: <ISO-8601>
prompt_executor: cover-prompt-strategy
image_provider: <executor-lock 中⑥登记的 image provider>
reference_mode: REQUIRED # 或 NONE
reference_asset: reference/jeffrey-reference.jpg # NONE 时为 null
reference_sha256: <sha256> # NONE 时为 null
style_count: 3
size_count_per_style: 3
candidate_count: 9
sizes:
  - {aspect_ratio: "3:4", width: 1080, height: 1440}
  - {aspect_ratio: "4:3", width: 1440, height: 1080}
  - {aspect_ratio: "16:9", width: 1920, height: 1080}
---

## V1 <风格名>

<为什么适合本期;使用了哪种原版/扩展风格;角色、动作、主道具和场景是什么>

### V1-3x4 · 1080×1440

```text
<完整3:4 prompt>

V1-4x3 · 1440×1080

<完整4:3 prompt>

V1-16x9 · 1920×1080

<完整16:9 prompt>

V2、V3使用同一结构。Episode 与独立任务都固定输出3种风格×3种比例,共9个完整提示词。三个尺寸不是裁切版:同一风格内也要重新安排人物姿势、道具互动和场景细节。

## Image provider 与选择证据

prompt 写完后,编排器调用 `executor-lock.json` 已登记的 image provider,把 V1/V2/V3 的3:4、4:3、16:9各真实生成一张,共9张并全部参加比较。原图必须放入本期 `05-封面/`,不得用旧期图片、同图裁切或只写 prompt 冒充。provider 若支持 reference edit,使用已登记 reference 参数;OpenAI ImageGen 对人物身份保持使用 reference-image edit 路径,不得改走 brand-new generation。

每次真实调用追加写入 `05-封面/provider-requests.json`:`schema_version: 1`、`script_hash`、`image_provider`、`reference_mode`、reference 路径/SHA,以及每个候选的 `candidate_id`、`style_id`、`aspect_ratio`、`canvas`、`source_prompt`、`reference_asset`、`reference_sha256`、`output_asset`、`output_sha256`、`requested_at`。9个候选 ID 固定为 `V1|V2|V3` 与 `3x4|4x3|16x9` 的组合。声明 REQUIRED 却缺 reference,或 output SHA/尺寸与文件不一致,立即停在⑥。

provider 单次失败可用同一候选重试一次;仅当 `executor-lock.json` 已登记允许的 fallback 时才可切换 provider。任一风格缺任一尺寸、任一候选不可解码、9张不是独立生成、候选违反事实/可读性约束,或 default 与 fallback 均失败时,留在⑥并报告 `BLOCKED`,不得创建 `selected-cover.json` 或用旧图补位。

最终 `selected-cover.json` 至少包含:

```json
{
  "selected_style_id": "V1",
  "selected_asset": "05-封面/V1-16x9.png",
  "selected_assets": [
    {"candidate_id": "V1-3x4", "aspect_ratio": "3:4", "canvas": {"width": 1080, "height": 1440}, "asset": "05-封面/V1-3x4.png", "source_prompt": "05-封面/cover-prompts.md#V1-3x4"},
    {"candidate_id": "V1-4x3", "aspect_ratio": "4:3", "canvas": {"width": 1440, "height": 1080}, "asset": "05-封面/V1-4x3.png", "source_prompt": "05-封面/cover-prompts.md#V1-4x3"},
    {"candidate_id": "V1-16x9", "aspect_ratio": "16:9", "canvas": {"width": 1920, "height": 1080}, "asset": "05-封面/V1-16x9.png", "source_prompt": "05-封面/cover-prompts.md#V1-16x9"}
  ],
  "title": "与当前稿一致的标题",
  "canvas": {"width": 1920, "height": 1080},
  "large_text": ["实际画面中的大字"],
  "source_prompt": "05-封面/cover-prompts.md#V1-16x9",
  "prompt_executor": "cover-prompt-strategy",
  "image_provider": "openai-imagegen",
  "reference_mode": "REQUIRED",
  "reference_asset": "05-封面/reference/jeffrey-reference.jpg",
  "reference_sha256": "<sha256>",
  "selection_mode": "AGENT_PROXY",
  "script_hash": "<sha256>",
  "selected_at": "<ISO-8601>"
}

cover-review.json 必须复制 script_hashselected_style_idselected_assetselected_assetsprompt_executorimage_providerselection_mode,写 thumbnail_readability: "PASS"、3个风格结论、9个真实 candidates、非空 expected_metric、非空 reviewer 和有效 reviewed_at。风格层只能一个 SELECTED;候选层记录各自尺寸、asset/thumbnail、缩略图可读性和稿件一致性。最小结构:

{
  "schema_version": 1,
  "script_hash": "<sha256>",
  "selected_style_id": "V1",
  "selected_asset": "05-封面/V1-16x9.png",
  "selected_assets": ["05-封面/V1-3x4.png", "05-封面/V1-4x3.png", "05-封面/V1-16x9.png"],
  "prompt_executor": "cover-prompt-strategy",
  "image_provider": "openai-imagegen",
  "reference_mode": "REQUIRED",
  "reference_asset": "05-封面/reference/jeffrey-reference.jpg",
  "reference_sha256": "<sha256>",
  "selection_mode": "AGENT_PROXY",
  "thumbnail_readability": "PASS",
  "styles": [
    {"style_id": "V1", "verdict": "SELECTED", "reason": "具体理由"},
    {"style_id": "V2", "verdict": "NOT_SELECTED", "reason": "具体理由"},
    {"style_id": "V3", "verdict": "NOT_SELECTED", "reason": "具体理由"}
  ],
  "candidates": [
    {"candidate_id": "V1-3x4", "style_id": "V1", "aspect_ratio": "3:4", "canvas": {"width": 1080, "height": 1440}, "asset": "05-封面/V1-3x4.png", "thumbnail": "05-封面/thumbnails/V1-3x4.png", "readability": "PASS", "script_consistency": "PASS"}
  ],
  "expected_metric": "发布后固定窗口验证的指标",
  "reviewer": "审阅者",
  "reviewed_at": "<ISO-8601>"
}

REVIEW_GATED 的 selection_mode 必须是 JEFFREY;AUTONOMOUS_RUN 可用 AGENT_PROXY

选择标准

先按风格分组匿名看9张缩略图,比较3个完整风格组:

  • 真实像素/比例正确;
  • 缩到实际信息流尺寸后核心文字仍可辨;
  • 标题、对象、动作和画面不冲突;
  • 无水印、二维码、伪 UI、错误汉字或未核验事实;
  • 三个风格确实是不同场景、动作和主道具,不是同一模板换皮;每个风格的3个尺寸都独立构图且完整可用。

发布后才能验证:点击/首屏代理、5秒留存及与同账号基线的关系。没有固定窗口数据时只能写“当前可读性选择”,不能升级某模板为默认。

What ships with it: 2 files

13.0 KB alongside SKILL.md

references/

Keep looking

Skills are one crate of 325,949. Ordering is by how many stacks a row turns up in, so the top of any crate is what has actually been picked rather than what has the most stars.