Back to Details

generate-youtube-thumbnail

Compare original and translation side by side

🇺🇸

Original

English
🇨🇳

Translation

Chinese

Generate YouTube Thumbnail

生成YouTube缩略图

A reusable workflow for creating YouTube thumbnails via Arcads' Nano Banana 2 image endpoint with proper character likeness and proven CTR formulas.
一个可复用的工作流，通过Arcads的Nano Banana 2图像端点创建YouTube缩略图，确保人物相似度匹配，并采用经验证的CTR公式。

When to use this skill

何时使用本技能

Trigger on phrases like:
  • "make me a YouTube thumbnail"
  • "create a thumbnail for this video"
  • "I need thumbnail variations / A/B tests"
  • "remake this thumbnail with my face"
  • "generate 10 thumbnail concepts"
  • "thumbnail with [me / my product / my brand]"
当出现以下表述时触发：
  • "帮我做一个YouTube缩略图"
  • "为这个视频创建缩略图"
  • "我需要缩略图变体/A/B测试"
  • "用我的脸重新制作这个缩略图"
  • "生成10个缩略图概念"
  • "包含[我/我的产品/我的品牌]的缩略图"

Read order

阅读顺序

  1. This file — workflow, decision tree, batch generation
  2. shared/skills/generate-youtube-thumbnail/prompting/guide.md — likeness alignment, expressions cheat sheet, prompt structure (shared across all generative-AI APIs in this portfolio)
  3. shared/skills/generate-youtube-thumbnail/prompting/formulas.md — 5 proven thumbnail formulas with templates (shared)
  4. scripts/generate-batch.sh — Arcads-specific batch script (presigned upload + S3 PUT pipeline)
  1. 当前文件 — 工作流、决策树、批量生成
  2. shared/skills/generate-youtube-thumbnail/prompting/guide.md — 相似度对齐、表情速查表、提示词结构（适用于本组合中的所有生成式AI API）
  3. shared/skills/generate-youtube-thumbnail/prompting/formulas.md — 5个经验证的缩略图公式及模板（共享内容）
  4. scripts/generate-batch.sh — Arcads专属批量脚本（预签名上传+S3 PUT流水线）

Prerequisites

前置条件

  • .env
    with 
    ARCADS_BASIC_AUTH
    and 
    ARCADS_API_KEY
  • A 
    productId
    (defaults to MASTER_CONTEXT.md's "My workspace" product)
  • Reference images on disk (NOT pasted in chat — chat-pasted images are NOT accessible to the API): 
    • face/
      — 5+ photos of the subject (headshot + 3/4 angles + close-ups + expressions)
    • logos/
      — brand logos as files (orange Claude Code starburst, black Arcads A, etc.)
    • products/
      — clean product shots
    • examples/
      — real ad screenshots, comparison material
    • style/
      — example thumbnails the user wants to match aesthetically
If references are missing or the user pastes images in chat instead of saving them, stop and ask the user to drop the actual files into a project folder (e.g. 
references/youtube thumbnail/
). Chat paste ≠ file on disk.
  • 包含
    ARCADS_BASIC_AUTH
    和
    ARCADS_API_KEY
    的
    .env
    文件
  • 一个
    productId
    （默认使用MASTER_CONTEXT.md中「我的工作区」产品）
  • 磁盘中的参考图片（不可粘贴在聊天中——聊天中粘贴的图片无法被API访问）： 
    • face/
      — 5张以上主体照片（头像+3/4角度照+特写+表情照）
    • logos/
      — 品牌Logo文件（如橙色Claude Code星爆标志、黑色Arcads A标志等）
    • products/
      — 清晰的产品照片
    • examples/
      — 真实广告截图、对比素材
    • style/
      — 用户希望匹配风格的示例缩略图
如果缺少参考图片，或用户将图片粘贴在聊天中而非保存为文件，请停止操作并要求用户将实际文件放入项目文件夹（如
references/youtube thumbnail/
）。聊天粘贴≠磁盘文件。

Workflow

工作流

1. Gather requirements (in order)

1. 收集需求（按顺序）

Ask the user for any missing context, but only what you actually need:
  1. Concept — what's the video about? Single concept, A/B variations, or specific recreation of an existing thumbnail style?
  2. Subject — who is in the thumbnail (the user themselves, an AI character, no person)?
  3. Brand assets — which logos / products / brand colors should appear?
  4. Text — what should the title text say? Will text be baked in, or added in post (Canva/Photoshop)?
  5. Comparison material — for "real vs AI" thumbnails, what real ad and what AI-generated ad?
询问用户缺失的必要信息：
  1. 概念 — 视频主题是什么？单一概念、A/B变体，还是复刻现有缩略图风格？
  2. 主体 — 缩略图中的人物是谁（用户本人、AI角色、无人物）？
  3. 品牌资产 — 需要展示哪些Logo/产品/品牌颜色？
  4. 文字 — 标题文字内容是什么？文字是直接生成在图中，还是后期添加（Canva/Photoshop）？
  5. 对比素材 — 若制作「真实vs AI」缩略图，需提供哪些真实广告和AI生成广告？

2. Verify references exist on disk

2. 验证磁盘中存在参考文件

bash
ls "references/youtube thumbnail/"
If references are missing, ask the user to drop them. Do not proceed with text-only descriptions for brand-specific items (logos, branded products, branded apparel) — you'll get generic AI approximations that don't match the brand. Generic descriptions are OK for backgrounds, expressions, and clothing.
bash
ls "references/youtube thumbnail/"
如果缺少参考文件，请要求用户提供。对于品牌专属元素（Logo、品牌产品、品牌服饰），请勿仅使用文字描述进行操作——AI会生成与品牌不符的通用近似内容。背景、表情和服饰可使用通用描述。

3. Estimate cost and confirm

3. 估算成本并确认

Always present cost as an estimate before firing:
"Estimated cost: N variations × 24 credits = X credits. Tell me if you want to confirm exact pricing in the Arcads platform first."
在执行前务必提供估算成本：
"估算成本：N个变体 × 24积分 = X积分。如需确认精确定价，可前往Arcads平台查看。"

4. Pick a formula

4. 选择公式

See shared/skills/generate-youtube-thumbnail/prompting/formulas.md for the 5 proven formulas. Match the user's intent:
User says...Use formula
"Just me with my brand" / "branding thumbnail"Peace-sign / branding
"Real vs AI" / "compare" / "before/after"Real vs AI comparison
"Show the process" / "with the terminal"Terminal flow
"Surprised face" / "shocked reaction"Reaction shock
"Replace" / "alternative" / "swap out"Before/after split
查看**shared/skills/generate-youtube-thumbnail/prompting/formulas.md**中的5个经验证公式，匹配用户需求：
用户表述...使用公式
"只有我和我的品牌" / "品牌化缩略图"和平手势/品牌化
"真实vs AI" / "对比" / "前后对比"真实vs AI对比
"展示流程" / "包含终端界面"终端流程
"惊讶表情" / "震惊反应"反应震惊
"替换" / "替代方案" / "调换"前后分割对比

5. Compose prompts

5. 撰写提示词

Follow the template in shared/skills/generate-youtube-thumbnail/prompting/guide.md:
YouTube thumbnail, 16:9 landscape.
[SUBJECT — likeness block + clothing + framing + "no hands" if applicable]
Expression: [specific expression from expressions cheat sheet]
[LEFT visual element + reference]
[RIGHT visual element + reference]
Across the top in massive bold yellow block letters with thick black outline reads [TITLE].
Background: [color + glow]
Style: [aesthetic notes]
Avoid: distorted face, extra fingers, hands visible, blurry logos, generic face
Always include the CRITICAL CHARACTER LIKENESS block when the subject is a real person. See 
shared/skills/generate-youtube-thumbnail/prompting/guide.md
.
遵循**shared/skills/generate-youtube-thumbnail/prompting/guide.md**中的模板：
YouTube缩略图，16:9横版。
[主体 — 相似度模块+服饰+构图+如有需要添加「无手部」]
表情：[表情速查表中的具体表情]
[左侧视觉元素+参考]
[右侧视觉元素+参考]
顶部用巨大加粗黄色方块字体（带粗黑轮廓）显示[标题]。
背景：[颜色+光影效果]
风格：[美学说明]
避免：面部扭曲、多余手指、手部可见、Logo模糊、通用面部
当主体为真实人物时，务必添加关键的人物相似度模块，详见
shared/skills/generate-youtube-thumbnail/prompting/guide.md
。

6. Generate (use the batch script)

6. 生成（使用批量脚本）

Copy 
scripts/generate-batch.sh
to a new versioned script (
scripts/generate-thumbnails-vN.sh
) and modify:
  1. Update 
    REF_BASE
    and 
    COMMON_REFS
    array with your reference file paths
  2. Replace the 
    PROMPTS
    array entries with your composed prompts
  3. Run with 
    bash scripts/generate-thumbnails-vN.sh > output/run.log 2>&1 &
  4. Monitor with 
    tail -F output/run.log | grep -E "DONE|FAILED|Asset"
The script handles:
  • Image upscaling (Lanczos to 1080px longest side, RGB JPEG conversion)
  • Presigned upload + S3 PUT
  • Fresh upload per generation (critical — re-using filePaths causes 500 errors)
  • Parallel firing (10 in flight ≈ 1.5 min total)
  • Retry on failure
  • Asset polling and download
将
scripts/generate-batch.sh
复制为新版本脚本（
scripts/generate-thumbnails-vN.sh
）并修改：
  1. 更新
    REF_BASE
    和
    COMMON_REFS
    数组，填入参考文件路径
  2. 将
    PROMPTS
    数组内容替换为撰写好的提示词
  3. 执行命令：
    bash scripts/generate-thumbnails-vN.sh > output/run.log 2>&1 &
  4. 通过
    tail -F output/run.log | grep -E "DONE|FAILED|Asset"
    监控进度
脚本功能：
  • 图片放大（使用Lanczos算法将最长边放大至1080px，转换为RGB JPEG格式）
  • 预签名上传+S3 PUT
  • 每次生成重新上传（至关重要——重复使用filePaths会导致500错误）
  • 并行执行（10个任务并行≈总耗时1.5分钟）
  • 失败重试
  • 资产轮询与下载

7. Review and present

7. 审核与展示

After all generations complete, read each thumbnail with the Read tool and present:
  • Brief verdict per thumbnail (likeness, readability, emotional impact)
  • Top 3 picks ranked by CTR potential
  • Specific reasons for the picks (which expression, which color contrast, which formula)
  • Offer next-step refinements (different expression, background color, copy variation)
所有生成完成后，使用读取工具查看每个缩略图并展示：
  • 每个缩略图的简短评价（相似度、可读性、情感冲击力）
  • 按CTR潜力排名的前3个选择
  • 选择的具体理由（表情、色彩对比、公式类型）
  • 提供后续优化选项（更换表情、背景颜色、文案变体）

8. Mandatory disclosures

8. 强制披露

  • Always label credit totals as estimates and tell the user to confirm exact pricing in Arcads
  • Cost data: Nano Banana 2 image = 24 credits per generation (post-April-2026 800x credit multiplier)
  • Generation time: ~30–60 seconds typical
  • Parallel budget: 10 in parallel finishes in ~1.5–2 min total
  • 始终将积分总额标注为估算值，并告知用户前往Arcads平台确认精确定价
  • 成本数据：Nano Banana 2图像生成 = 每生成一次24积分（2026年4月后800x积分乘数）
  • 生成时间：通常约30–60秒
  • 并行预算：10个任务并行总耗时≈1.5–2分钟

Quirks and pitfalls

注意事项与陷阱

Reference images are effectively one-shot per upload

参考图片每次上传仅可使用一次

Reusing the same uploaded 
filePath
across multiple generation calls causes 
HTTP 500 UNKNOWN_ERROR
. Always upload fresh references for each generation. The batch script handles this automatically with 
upload_all_fresh()
per 
generate_one()
call.
在多次生成调用中重复使用相同的已上传
filePath
会导致
HTTP 500 UNKNOWN_ERROR
。每次生成务必重新上传参考图片。批量脚本通过每个
generate_one()
调用中的
upload_all_fresh()
自动处理此问题。

Image preprocessing is mandatory

图片预处理是必需的

Images smaller than 1080px longest side return 
422 — The provided image is too small.
The skill's 
prepare_image()
function:
  1. Opens the image
  2. Converts to RGB (strips alpha which trips some endpoints)
  3. Upscales to 1080px longest side with Lanczos resampling if needed
  4. Saves as JPEG quality 92
最长边小于1080px的图片会返回
422 — 提供的图片过小。
本技能的
prepare_image()
功能：
  1. 打开图片
  2. 转换为RGB格式（移除透明通道，避免触发部分端点错误）
  3. 若需要，使用Lanczos重采样将最长边放大至1080px
  4. 保存为质量92的JPEG格式

referenceImages
is array of plain strings

referenceImages
是纯字符串数组

Not objects. Sending 
[{filePath: "..."}]
returns 
400 — each value in referenceImages must be a string
. Send 
["external-api-temp-uploads/abc.jpg", ...]
instead.
不可为对象。发送
[{filePath: "..."}]
会返回
400 — referenceImages中的每个值必须为字符串
，应发送
["external-api-temp-uploads/abc.jpg", ...]
。

Chat-pasted images are NOT files

聊天中粘贴的图片并非文件

If the user pastes an image directly in chat, you cannot pass it to the API. Ask them to save the actual file into a project folder.
若用户直接在聊天中粘贴图片，无法将其传递给API。请要求用户将实际文件保存到项目文件夹中。

Likeness drift without enough references

参考图片不足会导致相似度偏差

With 1-2 face references the AI generalizes to "generic bearded man with glasses." With 5+ face references from different angles it locks in the specific person. Always use 5+ face references for character work.
仅1-2张面部参考图时，AI会生成「戴眼镜的普通胡子男性」这类通用形象。使用5张以上不同角度的面部参考图才能锁定特定人物。人物相关工作务必使用5张以上面部参考图。

macOS bash 3.2

macOS bash 3.2

Default macOS bash doesn't support 
declare -A
(associative arrays). The batch script uses indexed arrays + temp files instead.
默认macOS bash不支持
declare -A
（关联数组）。批量脚本使用索引数组+临时文件替代。

Stale presigned URLs

预签名URL过期

Presigned URLs expire after the 
expiresIn
window (~10 min). Don't reuse URLs across long-running jobs — upload fresh.
预签名URL会在
expiresIn
窗口（约10分钟）后过期。长时间任务中请勿重复使用URL——重新上传即可。

Brand-specific items need actual reference files

品牌专属元素需要实际参考文件

Text descriptions of brand-specific items (logos, branded apparel, custom merchandise) will produce generic approximations. The Mr. Paid Social hat from text alone reads as "MR PAID SOCIAL" but won't match the real patch typography. For pixel-accurate brand reproduction, save the actual brand asset to disk and pass it as a reference.
品牌专属元素（Logo、品牌服饰、定制周边）的文字描述会生成通用近似内容。仅用文字描述的Mr. Paid Social帽子会显示为「MR PAID SOCIAL」，但无法匹配真实补丁的排版。如需像素级精准还原品牌资产，请将实际品牌文件保存到磁盘并作为参考传递。

Cost reference

成本参考

OperationCreditsNotes
Nano Banana 2 image (1 generation)24post-800x multiplier
6-variation batch144typical for first explorations
10-variation batch240typical for refinements
20-variation batch480typical for broad concept exploration
Always present as estimates, confirm exact in the Arcads platform.
操作积分说明
Nano Banana 2图像生成（1次）24800x乘数后
6个变体批量生成144首次探索的典型数量
10个变体批量生成240优化阶段的典型数量
20个变体批量生成480广泛概念探索的典型数量
始终以估算值呈现，精确价格请前往Arcads平台确认。

See also

相关链接

  • shared/skills/generate-youtube-thumbnail/prompting/guide.md — likeness alignment, expressions, prompt structure (shared across APIs)
  • shared/skills/generate-youtube-thumbnail/prompting/formulas.md — 5 proven CTR formulas with prompt templates (shared)
  • scripts/generate-batch.sh — Arcads-specific bash batch generator (presigned upload pipeline)
  • arcads-external-api skill — underlying API reference for Nano Banana 2 endpoint and reference upload pipeline
  • shared/skills/generate-youtube-thumbnail/prompting/guide.md — 相似度对齐、表情、提示词结构（适用于所有API）
  • shared/skills/generate-youtube-thumbnail/prompting/formulas.md — 5个经验证的CTR公式及提示词模板（共享内容）
  • scripts/generate-batch.sh — Arcads专属bash批量生成器（预签名上传流水线）
  • arcads-external-api skill — Nano Banana 2端点及参考上传流水线的底层API参考