深度技術分析:具體 Workflow、Commands、工具
呢份係補充分析,聚焦係每條影片嘅具體技術細節、真實 commands、prompt 例子
一、DLS / Design Language System(FwmhsetCbGY)
DLS.md 文件結構(具體 Sections):
1. Topography - 字體系統(字型、大小、weight)
2. Color Palettes - 色彩規範(精確 hex codes)
3. Spacing Rules - 間距規則(px 數值)
4. Component Standards - 按鈕、卡片、表格等組件標準
5. Facial Hierarchy - 視覺層級
6. Design Tokens - 所有精確設計值的定義
Skill.md 結構:
- Purpose - 技能做咩
- When to use - 點樣 trigger
- Step-by-step - 執行步驟
- Brand file checks - 執行前驗證:brand_identity.md、DLS.md、Claude.md 是否存在
- Guard rails - 安全規則(冇 brand context 唔可以 proceed)
完整 Setup Workflow:
步驟 1:準備 /knowledge folder
├── brand_identity.md ← 品牌名稱、audience、personality
├── DLS.md ← design tokens、color codes、typography
└── Claude.md ← project 描述、folder structure
步驟 2:生成 skill.md
- 一條 prompt 讓 CC 生成 skill.md
- skill 自動 check 3 個必要文件
步驟 3:使用
- /design-system "Create Instagram post"
- Claude 自動 apply DLS,無需每次重新解釋品牌
步驟 4:生成輸出
- Frontend design plugin → HTML artifact(免費)
- Nano Banana MCP → 圖片生成(需 API key)
- Remotion + remotion-best-practices skill → 影片生成
關鍵工具:
- Frontend design plugin(HTML/CSS 設計,CC 免費)
- Nano Banana MCP(AI 圖片生成)
- Remotion(JavaScript 影片生成)
- Playwright(網頁截圖/爬蟲)
二、Programmatic SEO / 大量網頁生成
搜尋結果:19 條片冇提及 headless CMS(Strapi/Webflow)做 1000 個頁面的技術細節。
但有相關嘅概念:
Landing Page 生成(G1YcfZKlpUU):
- Cory Haynes repo 有
/landingskill - 輸入:業務 URL
- 輸出:完整 landing page copy + structure(Markdown 格式)
- 但係:係生成文案,唔係自動上傳到 CMS
SEO Skills(atONpdivNBo):
/SEOskill → technical + on-page + keyword audit/SEO audit→ 返回修復優先清單
Content Scale(QcUjsmtT7qw):
- 51 Substack posts → 分析 winning patterns → 生成無限新 ideas
- 概念:scale content 唔係 pages
結論:呢批 YouTubers 聚焦係 AI 生成文案/內容,唔係 programmatic 建立大量網頁。如果需要 1000 個 landing pages,需要另外研究 CC + Strapi/Webflow API 嘅 workflow。
三、逐個影片詳細技術細節
eorc3jLBqIA — Marketing Audit Tool
15 個 Commands 完整清單(推斷自 transcript):
1. /market audit ← 主命令:launch 5 parallel agents
2. /market competitors ← 競爭對手分析
3. /market funnel ← 轉化漏斗分析
4. /market landing ← Landing page 審計
5. /market SEO ← SEO 審計
6. /market copy ← 文案重寫
7. /market email ← Email sequence
8. /market social ← Social media 策略
9. /market-report-PDF ← 生成 PDF 報告
10-15. 其他 sub-commands
5 個 Parallel Agents 功能:
Agent 1: Content & messaging analysis
Agent 2: Conversion optimization
Agent 3: SEO & discoverability
Agent 4: Competitive analysis
Agent 5: Strategy & growth opportunities
PDF 生成方式:
- Python 3(CC 自動 install)
- 生成有分數的報告(score out of 100)
- 例子輸出:Overall score: 64/100,有各 dimension 細分
GitHub Repo: 影片 description 有連結(免費開源)
JqXkPlX_4gQ — AI Sales Team(14 個 Commands)
完整 Command 清單:
1. /sales prospect ← 公司 research(5 parallel agents)
2. /sales contact ← 找 decision makers
3. /sales outreach ← Cold email(5 個 variations)
4. /sales qualification ← Lead scoring
5. /sales prep ← 會議準備 brief
6. /sales proposal ← 客戶 proposal 生成
7. /sales pipeline ← Sales PDF report
8. /sales research ← 公司深度研究
9. /sales follow-up ← Follow-up sequences
10. /sales objection ← Objection handling scripts
11. /sales ICP ← Ideal customer profile 建立
12. /sales competitive ← 競爭情報
13-14. 其他
完整 Workflow 演示:
輸入:/sales prospect loom.com
5 Parallel Agents 同時執行:
├── Company research
├── Decision maker finding(LinkedIn/web search)
├── Opportunity scoring
├── Competitor analysis
└── Outreach strategy
Web 搜尋工具:Perplexity API
PDF 輸出內容:
├── Score: 75/100(High Priority)
├── Decision makers(VP Revenue、Head of Sales + context)
├── Next steps:LinkedIn request + cold email
└── Methodology explanation
後續:/sales outreach
輸出 5 個 email variations:
├── Day 1 Hook:Subject line + personalized message
├── Day 3 Value Add:Context from research
├── Day 7 Social Proof
└── Day 14/21 Follow-ups
qwZ9zZqZymU — CC Ad Uploader(Meta API)
完整步驟:
準備:
- Google Drive link(含 ad copy、images、videos)
Prompt 例子(完整):
"Hey, let's launch these ads in the Bboards account
in the testing campaign. Let's use the same settings
and headlines, descriptions, ad copies as the
February testing campaign testing ad set."
CC 自動執行:
1. Fetch account data(所有 settings、pixels、previous campaigns)
2. 提取 February campaign 設定
3. 建議 bidding strategy(lowest cost / cost bidding)
4. Auto-generate asset names
5. 設定 budget(例如 $20/day)
6. Dry run:確認 11 videos、2 videos、9 images
確認後執行:
- 設定命名規則
- 配置 UTM parameters
- 關閉不需要的 Meta AI features
- 設定 correct pixel
- 驗證 URL
特殊功能:
- 現有貼文轉廣告:"make an existing post ad and put it in the feed adset"
→ Claude fetch post ID → 保留原有 social engagement
- 廣告分析:"check this adset. What's the best ad right now based on amount spent?"
重點:全程不需要打開 Meta Ads Manager
zDvgZwuvt8o — 18 AI Marketing Agents
完整 18 個 Commands:
1. /audit ← 8 dimensions marketing audit(分數:45/100)
2. /report ← Complete marketing strategy(90-day plan)
3. /copy ← Before/after rewrites + 5 headline variants
4. /email ← Full email sequence
5. /social ← 30-day content calendar + 1 week posts
6. /SEO ← Technical + on-page + keyword report
7. /brand ← Brand identity analysis
8. /content ← Content strategy
9. /funnel ← Sales funnel analysis(~2 min)
10. /landing ← Landing page optimization
11. /influencer ← Influencer database search
12. /AB-test ← A/B testing framework
13. /ROI ← ROI calculation
14. /hashtag ← Hashtag research
15. /competitor ← Competitor benchmarking
16. /growth ← Growth strategy
17. /CRO ← Conversion rate optimization
18. /analytics ← Analytics review
安裝方式:
# 下載 GitHub repo zip
# 解壓到 VS Code folder
npm install # Mac/Linux
# 或對應 Windows 指令
# 結果:18 commands installed globally真實演示(/audit 例子):
輸入:/audit [business URL]
輸出(幾分鐘內):
- Overall score: 45/100(very poor)
- Issues breakdown by dimension
- Conversion optimization weight
- Findings(Critical/High priority)
- Revenue impact estimate
- Competitor benchmarks
- 90-day action plan
QcUjsmtT7qw — AI Content Team(Substack 分析)
具體 Workflow(51 篇 Substack 文章):
輸入:上傳 51 個 Substack posts(raw/messy data OK)
Lookalike Skill 分析:
- 識別 top 30% performing posts(如有 engagement data)
- 建立 Content Profile:
├── Winning formula clusters
├── Structural DNA(post length、sections、paragraph style)
├── Hook formula
└── Emotional playbook
Prompt 例子:
"Create talking points for marketers trying to adopt AI"
"Extract viral talking points from trend and topics"
輸出類型:
- Viral talking points(dated、platform-specific)
- Educational posts
- Data nuggets
- Spicy takes
- Story sparks
真實例子輸出:
"Your marketing team is building the tools you used to buy"
11 Skills 架構(完整 Content Team):
1. Orchestrator skill(master controller)
2. Audience profile builder
3. Writing style A(Persona 1)
4. Writing style B(Persona 2)
5. Talking point extractor
6. Lookalike content skill(免費提供)
7. Post enricher
8. LinkedIn content skill
9. Newsletter content skill
10. X/Twitter content skill
11. YouTube content skill
+ Feedback loops(月度自動更新)
自動改進系統:
- App 追蹤所有 generated posts
- 輸入 performance data
- 月度 review:skills 自動更新
G1YcfZKlpUU — 80% Marketing Automation(Cory Haynes Repo)
具體使用步驟:
步驟 1:Setup product_marketing_context.md
- Company name
- Website URL
- Company description
- Past work examples
(更多 context = 更好輸出)
步驟 2:下載 skill from GitHub(Cory Haynes repo)
例如:page-CRO skill.md
步驟 3:使用 /page-CRO
"I want to optimize for booking a discovery call.
Can you help with my hero section copy?"
CC 會問:
- Existing hero copy?(yes/no)
- Primary target visitor?
- Traffic source?(YouTube/Google/etc)
輸出:
- 3 headline variations:
① Outcome-led
② Problem-led
③ Specificity-led
- Thinking behind each
- Recommendations(proof metrics、testimonial、how-it-works)
- Design notes for page structure
步驟 4:使用 /paid-ads
"I want to start running ads for my agency
to get more consistent lead flow"
CC 會問:
- Monthly ad budget?
- Conversion goal?
- Platform?(LinkedIn/Meta/Google/multiple)
- Landing page?(yes/no - 需要建立嗎?)
輸出:
- LinkedIn ad strategy(2 campaigns + budget allocation)
- Ad copy variations(3x:outcome hook、pain hook、social proof)
- Targeting specifics
- Landing page brief(完整頁面結構)
- 30-day optimization plan
XPl6IKDADkU — AI Social Media Manager(1.5M Followers Creator)
完整設定:
工具: Blotato(creator API 工具,連接多平台)
- 連接 Instagram、TikTok、YouTube、LinkedIn、Twitter
Brand Voice Setup 問題清單:
1. What is your brand voice? (tone, topics, target audience)
2. Signature phrases or catchphrases?
3. Connected Blotato accounts? (account IDs、page IDs - 可自動 fetch)
4. API key storage? (environment variable)
5. Workflow preference? (draft first always / auto-publish)
6. Visual generation? (ask each time / always / never)
7. Post log format? (markdown table / CSV / JSON)
Skill 文件結構(social-media-post.md):
Platform-specific guidelines:
- Twitter: short、punchy、handle + hashtags
- LinkedIn: professional、longer form、no hashtags
- Instagram: emojis、storytelling、line breaks使用例子:
/social-media-post
Topic: "first-time home buyer tips"
Draft mode 輸出:
├── Twitter version (punchy, 280 chars)
├── LinkedIn version (professional, longer)
└── Instagram version (emojis, storytelling)
批准後:
→ Generate visual (Blotato template)
→ Choose: infographic / carousel / slideshow / video
→ Download visual-preview.jpeg
→ Publish all platforms simultaneously
→ Auto log: date、platform、URL
Quality Gates(防止低質輸出):
- 移除 m-dashes
- 限制 emoji 數量 < 2
- Brand voice samples(previous posts)
atONpdivNBo — Complete AI Marketing Team(32 個 Skills)
安裝方式:
Prompt: "Install 32 skills from GitHub repo globally"
結果: "All 32 skills installed globally"
32 個 Skills 組織架構:
SEO Category:
├── CRO(Conversion Rate Optimization)
└── SEO audit
Marketing Category:
├── Paid ads
├── Analytics
├── Growth strategy
└── Sales / RevOps
Content Category:
├── Planning
├── Email sequences
└── Social media
(共 32 個)
工作流演示:
/SEO-audit [website URL]
輸出:
- Critical: XML sitemap missing
- High: Canonical tags missing
- Quick summary + next steps
(無需再次提供 URL)
/report
輸出:完整 marketing strategy report
成本計算:
- Claude Pro:$20-200/month
- = 32 個 marketing team members 的工作量
A6RbawFHC80 — Claude Cowork:批量 Sub-agents
原理:
- Main agent spawn 多個 sub-agents
- 各自獨立、並行(同時)執行
- Sub-agents 返回 summary 給 main agent
真實 Lead Qualification 例子:
Input: 150 marketing agencies list(from Apollo)
Prompt:
"Qualify these 150 leads against ICP criteria.
Please SPIN UP 15 PARALLEL SUB-AGENTS.
Each agent research 10 leads."
執行:
- 15 sub-agents 同時 launch
- 每個 verify:SEO services? US-based?
- Web search 驗證
- Return summary
Output: 82/150 qualified(約 2 分鐘)
Lead Enrichment(第二步):
/enrichment command
"Spin up 10 PARALLEL sub-agents that enrich 7 leads each"
每個 agent:
- Find last LinkedIn post
- Company size
- Company description
(用 Apify LinkedIn scrapers)
Output: Enriched CSV
Email Personalization(第三步):
17 sub-agents → 37 personalized outreach messages
添加入 CSV
⚠️ 重要提示:
- 必須講「spin up X parallel sub-agents」 - 唔講會變 sequential,慢好多
- 每個 sub-agent 處理 5-15 個任務最優
- Token 消耗極高:建議 Claude Max plan
- 100-200 leads 最理想,超過 200 會複雜
xRdlmFBtGn0 — Claude Cowork 商業應用
Claude Cowork vs Claude Code:
| Claude Cowork | Claude Code | |
|---|---|---|
| 介面 | Web app(GUI) | CLI terminal |
| 目標用戶 | 非技術人員 | 技術人員 |
| 運行方式 | 本機後台 | 本機前台 |
| 文件訪問 | 完整本機訪問 | 完整本機訪問 |
| 後台執行 | ✅ | ❌(需等候) |
真實商業案例(Lenny Podcast):
Input:
- 320 podcast transcripts(folder access)
- YouTube analytics CSV
Prompt:
"Generate new format for podcast
that will generate more subscribers"
CC 自動:
1. Read all 320 transcripts locally
2. Analyze YouTube analytics data
3. 識別 patterns:
- Best performing: 15-25 min tutorials
- Highest CTR topics
- Best hooks
Output(15-20 分鐘):
- Presentation deck with slides
- JavaScript charts embedded
- Recommendations: 18 min sweet spot
- Formula: time saved + specific outcome + method
- Winning format pillars
價值:策略值數千元
另一個案例(Vibhu):
2 小時內完成:
- 14 個 job descriptions
- 1 個 marketing strategy
- Partner emails
- Website copy
- 23 個 LinkedIn DM replies
= 相當於 2 個月工作量
四、關鍵技術洞察(跨影片總結)
速度比較表
| 任務 | 工具 | 時間 | 傳統時間 |
|---|---|---|---|
| Marketing Audit | 5 parallel agents | 2 min | 1-2 週(Agency) |
| Sales Prospect | 5 parallel agents | 1-2 min | 半天手動 |
| Lead Qualification | 15 parallel agents | 2 min | 幾天人手 |
| Podcast Strategy | Co-work | 15 min | 數千元顧問費 |
| Social Content | Skill + Blotato | 5 min | 1 小時 copywriter |
| Ad Launch | Meta API | 即時 | 半天 media buyer |
最重要嘅概念
- Skills = 可重用的 system prompts(冇編碼,.md 文件)
- Sub-agents = 批量並行處理(必須主動說「parallel」)
- Context memory = 唔使重複提供資料(CC 記得)
- Local execution (Co-work) = 後台處理 + 完整文件訪問
- GitHub repos = Open-source skills libraries(免費)
Sub-agents 使用模板
"Please spin up [N] parallel sub-agents.
Each agent should [task] for [X] items."
例子:
Spin up 15 parallel sub-agents, each research 10 leadsUse 5 parallel agents to audit this website
分析基於完整 transcript 提取,2026-03-27