Codex 新能力实战:用五个 Agent 搭 AI 营销团队

这支 YouTube 视频把 Codex 的新版能力从“写代码工具”重新解释为“可编排的自主工作团队”:用一个 campaign manager 统筹 research、content strategy、creative copy、data analysis 四个专业 Agent,围绕 Flowers.co 母亲节活动交付竞品研究、内容日历、KPI 看板、社媒文案和最终 QA 报告。

基本信息

  • 来源类型:YouTube 视频字幕
  • 原文位置raw/articles/2026-07-19-youtube-codex-ai-marketing-team.md
  • 原始链接https://youtu.be/pwBPScxS1bg
  • Telegram 桩文件raw/articles/2026-07-19-205532-tg-1e7929.md
  • 消化日期:2026-07-19
  • 素材主题:Codex 新版项目工作区、权限与模型设置、五 Agent 营销团队搭建、插件/文件夹/任务清单式编排、交付物 QA

核心观点

  1. Codex 的定位被视频重定义为“自主工作团队”,不只是 coding assistant:视频开头列出新版 Codex 可并行运行多个 Agent、直接操作电脑、跨天/跨周排程任务,并连接 Microsoft Suite、Slack、Gmail、Notion 等 90+ 插件。作者据此认为 Codex 的产品心智应从“写代码”升级为“可执行、可调度、可连接业务工具的 autonomous workforce”。

  2. 五角色 + 一个编排者是最小可懂的营销团队结构:视频设计了 research、content strategy、creative copy、data analysis、campaign manager 五个 Agent。前四个分别负责调研、策略、文案和数据,campaign manager 负责读 brief、生成 task manifest、按依赖关系分派任务、读取各 Agent summary、做品牌 QA 并汇总最终 campaign summary report。这是典型的 Multi-Agent 系统 层次委托模式。

  3. agent.md / brand.md 是项目真相源,而不是临时 prompt:搭建第一步不是直接让 Agent 开干,而是在 Codex 项目里创建 agent.md 定义系统架构、五个 Agent 的职责、依赖库、输出目录、文件命名、日志和编排规则;再创建 brand.md 存品牌名、受众、语气、定位、主题和 KPI 基线。这个做法与 Agentic Workflow 的“共享状态 + 明确中间格式 + 可追踪输出”一致。

  4. 每个 Agent 都有专属输出文件夹和严格交付物格式,避免互相覆盖:视频强调 research、content strategy、creative copy、data analysis、campaign manager 各自写入 /output 下独立子目录,并采用 date-agent-task 命名。Agent 之间不靠“脑内记忆”传递,而是读取前序产物、写 summary 和报告,campaign manager 再按路径核验所有交付物。

  5. Research Agent 是最重工具栈节点:浏览器抓取 + 截图 + PPT/Doc 输出:research agent 被设定为五个 Agent 中技术能力最强的角色,使用 Playwright 浏览网页、截屏、抽取结构化数据,并通过 python-pptxpython-docx 生成竞品分析 deck 和 written brief。测试任务是为 Flowers.co 母亲节活动找美国鲜花配送市场前五竞争对手,截首页/定价页,分析 messaging 与 positioning,并输出至少 8 页 slide deck。

  6. Content Strategy Agent 把调研转成可执行内容计划:它先读 brand.md,再读 research agent 的竞品分析,输出两周 Instagram 母亲节内容日历,包含 reels、carousels、static posts 三种格式,每条给 content theme、hook direction、CTA、visual notes,并生成 spreadsheet + strategy doc。视频中特别指出表格自动按内容类型着色,说明 Agent 不只是填文本,也会主动增强可读性。

  7. Data Analysis Agent 负责把 KPI 转成可行动缺口,而不只是画图:它读取品牌 KPI baseline 与 target,清洗/分析数据,生成交互式 HTML dashboard 和一页 performance snapshot report。测试中它识别出三个最大缺口:subscription customers、repeat purchase rate、email list growth,并把这些缺口转成三条 data-driven priorities,再建议交给 content strategy agent 更新日历。

  8. Creative Copy Agent 的关键不是“写几条文案”,而是依赖前序 brief 且自检品牌一致性:视频要求 creative copy agent 读取内容策略产物,为前五条 Instagram 帖子写 hook-led captions,每条 150 词以内、带 CTA 和 8 个 hashtag。更关键的是规则:若没有 content strategy agent 的 creative brief,它必须停止请求 brief,不能凭空猜测。输出后还要对照品牌 voice、strategy、positioning 自审。

  9. Campaign Manager Agent 的价值在于 QA 和可追溯汇总:最终 orchestrator 会验证 12 个 deliverables 是否存在,读取各 summary 文件,对照 brand.md 做 QA,生成 campaign overview、deliverable status table、research findings、KPI summary、content strategy summary、copy samples、QA notes 和 next steps。它不仅确认 12/12 输出存在,还标出品牌命名不一致、banned words 候选和 FTD 网站 403 阻挡 Playwright 的已知 gap。

  10. 视频把 Claude Code 与 Codex 做了成本/额度叙事对比,但本质是“任务路由与工具心智”之争:作者提到 Claude Opus 新 tokenizer 可能让同样输入产生更多 token、用户遇到 usage limit;随后用 Codex 的并行 Agent、90+ 插件、自动排程、intelligence slider 作为替代叙事。对知识库而言,关键不是站队某模型,而是看到 Agent 工作台竞争正在从模型质量转向:并行执行、插件连接、排程、权限、输出路径、QA 证据和团队级工作流。

实操内容保留

1. Codex 项目初始化

视频中的初始化步骤:

  1. 在 Codex 左侧 Projects 选择 Start a new one
  2. 将项目重命名为 marketing team
  3. 通过聊天框下方的 project dropdown 确认当前 active project,可在项目间切换或选择不在项目中工作。
  4. 权限保持 default,让 Codex 在执行重大动作前请求 approval;视频不建议一开始就打开 full access。
  5. 模型示例为 GPT 5.5,intelligence level 日常任务选择 medium,以在输出质量和 token efficiency 之间取得平衡。

2. agent.md 项目真相源

视频描述的 agent.md 应包含:

# Marketing Team Agent System
 
## System Architecture
| Agent | Skill File | Primary Responsibility |
|---|---|---|
| Campaign Manager | skills/campaign-manager.md | Read brief, build task manifest, delegate, QA, final report |
| Research Agent | skills/research.md | Live web research, screenshots, competitor analysis, decks/docs |
| Content Strategy Agent | skills/content-strategy.md | Calendar, messaging framework, creative brief |
| Creative Copy Agent | skills/creative-copy.md | Channel copy, self-review against brand voice |
| Data Analysis Agent | skills/data-analysis.md | KPI analysis, charts, dashboards, reports |
 
## Dependencies
- Node
- Python
- Playwright
- python-pptx
- python-docx
- openpyxl
- pandas
- matplotlib / plotly
 
## Output Structure
/output/research/
/output/content-strategy/
/output/creative-copy/
/output/data-analysis/
/output/campaign-manager/
 
## Orchestration Rules
- Campaign manager reads campaign brief and task manifest.
- Each specialist writes outputs and summary files.
- Campaign manager verifies output paths, QA notes, and gaps.
- All actions are logged.
 
## File Naming
YYYY-MM-DD_agent_task.ext

视频没有逐字展示完整文件,上述为它口述结构的忠实整理,核心是:职责表、依赖、输出目录、编排规则、命名规范和日志。

3. brand.md 品牌上下文

brand.md 存放所有 Agent 开工前必须共享的品牌信息:

  • brand name
  • audience
  • tone
  • positioning
  • campaign themes
  • KPI baselines and targets
  • channel preferences
  • banned words / brand-safe wording candidates

视频示例品牌是 Flowers.co,作者提醒:示例可以简单,但真实项目里给 Agent 的上下文越完整,输出越稳定。

4. Research Agent workflow

Research Agent 的六阶段工作流可整理为:

  1. Understand task:读取任务、品牌和目标市场。
  2. Live web research:用 Playwright 浏览竞品官网。
  3. Capture screenshots:截首页、定价页等关键页面。
  4. Structure findings:整理定位、价格、信息架构、主视觉、CTA。
  5. Generate deliverables:输出 slide deck 和 written research brief。
  6. Save and log:保存到 research 文件夹,并记录失败/阻挡页面。

测试任务:

For Flowers.co's Mother's Day campaign, find the top five competitors in the US flower delivery market. Screenshot their home pages and pricing pages. Analyze messaging and positioning. Deliver two outputs: a competitor analysis slide deck of at least eight slides, and a written research brief with all screenshots embedded.

视频中发现的问题也值得保留:

  • 截图模糊:需要在 skill file 中增加固定 resolution / 不压缩长图的规则。
  • FTD 页面 403 或 asset blocked:Agent 不能静默跳过,应记录 fallback 和 known gap。

5. Content Strategy Agent workflow

Content Strategy Agent 接收 research 输出后,生成两周 Instagram 内容日历:

Build a two-week Instagram content calendar for Flowers.co's Mother's Day campaign. Read the brand file, pull the competitor analysis from the research agent, and deliver a mix of reels, carousels, and static posts. Each post should include a content theme, hook direction, call to action, and visual notes. Output a spreadsheet and a short strategy document explaining the rationale.

输出表格字段:

DateFormatContent ThemeHook DirectionCTAVisual Notes

视频亮点:Codex 自动按 reels / carousels / static posts 给内容类型上色,提升排程表可读性。

6. Data Analysis Agent workflow

Data Analysis Agent 的工具栈和输出:

  • pandas:加载、清洗数据
  • matplotlib / plotly:图表或交互式 dashboard
  • python-docx:报告
  • openpyxl:表格导出

测试任务:

Read KPI baselines and targets from the brand file. Produce a grouped bar chart showing current metrics versus 12-month targets, and a one-page performance snapshot report with the biggest gaps and three data-driven priorities.

视频中临时调整:将 dashboard 输出从 PNG 改为 HTML,并要求 interactive dashboard,减少“先生成再转换”的额外步骤。

7. Creative Copy Agent workflow

Creative Copy Agent 规则:

  • 支持 Instagram、TikTok、Facebook、email、SMS、Google Ads、product descriptions 等渠道。
  • 每个渠道有独立长度、结构、tone、CTA 规则。
  • 如果没有 creative brief,应停止并请求 brief,不做猜测。
  • 输出前做 self-review,确保不违背 brand voice、strategy、positioning。

测试任务:

Write Instagram captions for the first five posts from the Mother's Day calendar. For each post, include a hook-led caption under 150 words, a clear call to action, and eight relevant hashtags. Output one structured markdown file.

视频提到的优质 hook 示例:

Not every mom is called mom.

作者认为这比 “Mother’s Day is coming” 更能 earn the scroll。

8. Campaign Manager workflow

Campaign Manager 是最终编排与 QA 节点:

  1. Read campaign brief。
  2. Build task manifest。
  3. Delegate tasks with strict structure:
    • Agent code name
    • Task description
    • Input path
    • Expected output
    • Deadline
  4. Monitor completion by reading each agent’s summary file。
  5. Run brand QA against brand.md
  6. Compile final campaign summary report。

最终报告结构:

  • Campaign overview
  • Deliverable status table
  • Research findings
  • KPI summary
  • Content strategy summary
  • Copy samples
  • QA notes
  • Next steps

视频中的 QA 结果:

  • 12 outputs / 12 confirmed。
  • 发现 flowers.co / Flowers.co 命名不一致。
  • 标出 luxurydiscountexclusive 等 banned words candidates 供人工复核。
  • 记录 FTD 被 Playwright 403 阻挡,不隐藏失败。

关键概念

  • Codex — 视频的承载平台,从 coding tool 扩展为营销团队工作台。
  • Multi-Agent 系统 — 五 Agent 分工、campaign manager 层次委托和 summary 文件协同。
  • Agentic Workflowagent.mdbrand.md、task manifest、输出目录、QA 报告构成完整业务流水线。
  • 营销智能体 — 本视频把营销智能体从投放后台扩展到调研、策略、文案、数据和活动管理全链路。
  • AI Agent 智能体 — 各专业角色以明确输入输出和权限边界执行任务。
  • AI营销趋势 — AI 营销从“生成内容”升级到“编排团队与工作流”。
  • AI办公自动化 — 用业务文件、表格、报告、看板和任务清单组织可复用工作。
  • AI编程开发 — Codex 的平台能力仍来自项目文件、依赖、工具调用、Playwright/Python 输出链路等工程化基础。

与其他素材的关联

  • 2026-06-27-claude-code-ai-marketing-team 高度同构:那篇用 Claude Code 搭 5 Agent + 12 Skill 营销团队,本篇用 Codex 搭 5 Agent 营销团队。差异在于本篇更强调 Codex 项目 UI、90+ 插件、排程、campaign manager QA 和 agent.md/brand.md 文件化编排。
  • 2026-07-07-woshipm-cannes-marketing-ai-agent-creator-economy 的关系:戛纳素材强调营销智能体在素材管理、投放和异常检测后台的确定价值;本篇展示前台营销生产链也可以 Agent 化,但仍需要 QA、品牌词审核和失败记录。
  • 2026-07-11-playpcesor-codex-knowledge-workflows 的关系:电脑玩物素材强调 Codex 进入知识工作中间流程;本篇把同一能力扩展到营销团队协作:从读资料、写报告、生成表格到交付 campaign summary。
  • 2026-07-17-woshipm-claude-cowork-agent-workflow 的关系:两者都强调独立审计/QA 与失败记录。本篇 campaign manager 记录 403、命名不一致、banned words candidates,符合“不能静默跳过 gap”的信任工程原则。

原文精彩摘录

Codex can now run multiple agents in parallel, operates your computer directly, schedule work for itself across days or weeks, and connect to over 90 new plugins including Microsoft Suite, Slack, Gmail, and Notion. And this is not just a coding assistant. This is actually an autonomous workforce.

Five specialized agents covering research, content strategy, creative copy, data analysis, and campaign management all coordinated by one single orchestrator that runs the whole operation.

The first thing we need, of course, is an agent.md file. You can think of it as the source of truth for the entire project, like the same way a Claude.md file works if you have watched any of our Claude Code tutorial videos before.

Every single agent gets its own subfolder, so nothing overlaps here. And finally, we set the orchestration rules here, like how the campaign manager delegates tasks, how agents share outputs through a central task manifest, and how every single action gets logged.

The agent identified the three biggest gaps to close, subscription customers, repeat purchase rate, and email list growth. And then mapped those gaps to three concrete data-driven priorities for the marketing team.

If there is no creative brief from the content strategy agent, then the agent will stop and request one. It will not do any guessing work.

Every output from every agent is referenced with its exact file path. And this can be the document that you hand to a client or a campaign owner at the end of a sprint.

The agent did not silently skip it. It just locked the block and documented the fallback and flagged it as a known gap with a clear recognition. Accept it or capture it manually. This is exactly how the system is supposed to work.

相关页面