Claude + Higgsfield 构建 AI 创意代理工作流:从手动生成到自动化内容工厂
用 Claude 作为统一界面连接 Higgsfield AI 图像/视频生成平台,通过 Skill 逆向工程 + Routines 定时任务 + Google Sheets 数据追踪,构建可扩展的 AI 创意代理系统
基本信息
- 来源类型:视频(YouTube 教程)
- 原文位置:raw/articles/2026-06-04-010944-tg-505479.md
- 原文 URL:https://www.youtube.com/watch?v=xn6Z5PYyAIE
- 消化日期:2026-06-04
核心观点
-
Claude 作为创意代理统一界面比手动操作效率提升 100 倍:通过 MCP 连接器让 Claude 直接调用 Higgsfield 的图像/视频生成能力,一条 Prompt 即可完成品牌研究→产品线规划→产品图→Instagram 广告→UGC 视频全链路——作者用一个 Prompt 在 5 分钟内为虚构耳机品牌 Murmur 生成了 3 个产品线的完整营销素材(产品照+广告+视频)
-
CLI 优于 MCP 连接器,Token 效率是核心差异:在 Claude Code(桌面版)场景下,Higgsfield CLI 比 MCP 连接器更优——MCP 暴露所有工具描述会占用大量 Token,CLI 直接调用更高效。操作上只需 3 条命令:安装 CLI → OAuth 认证 → 安装 Agent Skills
-
外部专家知识注入让 Claude 从”通用 AI”升级为”领域专家”:作者不直接让 Claude 写广告文案,而是先用 Deep Research 生成一份 617 行的 “Advertising Masterclass” 研究报告(覆盖 TikTok/Meta/X 各平台有机广告最佳实践),作为项目文件常驻。所有后续创意生成都参考这份研究文档——核心方法论是”利用别人的已被验证的专业知识,注入自己的系统”
-
GWS CLI + Google Sheets 构建创意资产数据库:通过 GWS CLI(Google Workspace CLI)让 Claude 读取 Higgsfield 账户中的所有 45 个历史生成记录(产品、风格、模型、提示词、结果 URL、Job ID),自动写入 Google Sheets 的”Generations”标签页。这个数据库后续用于:分析哪些创意表现最好、规划新变体(不同价值主张×不同标题×不同头像×不同风格=100 个测试矩阵)、追踪生成状态
-
Skill 逆向工程是构建一致性创意输出的核心方法:找到最满意的生成结果 → 复制其 Prompt → 让 Claude 将其逆向工程为
.claude/skills/下的 Skill 文件。Skill 的本质是”AI 的配方”——每次执行结果一致,且每次迭代都会自动优化(踩到的限制词、不满意的风格都会被记录并避免)。作者还把触发敏感内容限制的 Prompt 分析并加入 Skill 的”禁止词列表”,避免重复被拒 -
Routines 定时任务实现”睡觉时自动生成”:规划了两阶段 Routine 系统——周日晚上 Routine(分析上周数据+规划 50 个新创意变体写入 Sheet)+ 周一早上 Routine(从 Sheet 读取 30 个待生成项目→生成→回写状态)。可以从 50 扩展到 100-200,最终可连接 Meta Ads Manager 实现全自动投放
-
参考图一致性是产品广告的核心痛点和解法:首轮生成的广告中产品外观与参考图不匹配(AI 自行”发挥”了瓶子设计),作者强调必须明确告诉 Claude “使用此参考图片,不得做任何修改,保持相同的颜色和文字”。这是 AI 创意广告从”生成着玩”到”商业可用”的关键约束
-
敏感内容限制是可工程化解决的问题:Higgsfield 会因 Prompt 中特定词汇拒绝生成并退还 Credits。作者的解法不是避免使用,而是让 Claude 读取被拒的 Prompt、分析被拒原因、自动移除问题词汇后重新提交——整个过程对用户透明,且结果沉淀到 Skill 中避免下次再犯
实操内容保留
代码/配置
Higgsfield CLI 安装三步命令(在 Claude Code 中执行):
- 安装 Higgsfield CLI
- 运行
higgsfield auth login进行 OAuth 认证 - 安装 Higgsfield Agent Skills
(具体命令需从 higgsfield.ai 的 MCP & CLI 页面获取,因官方可能更新)
Prompt 模板
品牌从零构建 Prompt:
Build me a headphone brand from scratch. I want you to do research, build the branding, build the product catalog, and for each of them I want you to generate assets. So a product photo, an Instagram ad, and a UGC video, and I told it to use the Higgsfield MCP for all of these generations.
广告研究文档生成 Prompt:
I need you to do a deep research on the best strategies for advertising in 2026 when it comes to organic advertisements on platforms like TikTok, Meta, or X. And what captures people’s attention, what converts, and how it differs per platform. Create a full markdown file called advertising masterclass that would live in this project.
创意变体规划 Prompt:
Look at all of the different generations that we’ve done. Read that Advertising Masterclass doc, and help me figure out a bunch of different variations that we could create. Mix and match a bunch of different variables so we can ultimately test a bunch of these. Use your creativity, use your best practices, and help me get a bunch of different ideas for more creatives.
Skill 逆向工程 Prompt:
This prompt that you’re looking at above is my favorite output we’ve gotten from Higgsfield Marketing Studio. This was a hypermotion fast-paced kind of like launch video for our product, and I loved it. Turn this into a skill that lives locally inside of this project in the dot Claude/skills directory, so that anytime I ask for a hypermotion style video, you will utilize this, and they’re always consistent.
Routine 配置 Prompt:
Every Sunday I want you to look at this Google sheet, and pull in data from Instagram, and analyze what’s working, what’s not, and then ideate, and add 50 new generations to the sheet.
Monday morning you’re going to go to the sheet and pick 30 videos with a blank status. Create the prompts for all of them, generate all of them, and mark them off as complete.
操作步骤
- 在 higgsfield.ai 注册账号
- 在 Claude Web 中:Settings → Connectors → Add Custom Connector → 粘贴 MCP 命令 → OAuth 认证
- 在 Claude Code 中:创建空项目文件夹 → 粘贴 3 条 CLI 命令安装+认证+Skills
- 先用 Deep Research 生成领域研究报告作为项目常驻知识
- 用 GWS CLI 创建 Google Sheets 作为创意资产数据库
- 用自然语言让 Claude 读取历史生成记录并写入 Sheets
- 规划创意变体矩阵(变量:价值主张×标题×头像×风格)
- 批量生成→Sheets 自动回写状态和结果 URL
- 找到最佳输出 → 逆向工程为 Skill 文件
- 配置 Routines 实现定时自动生成
关键概念
- Higgsfield — AI 图像/视频生成平台,支持 MCP 和 CLI 两种接入方式,内含 Marketing Studio(hypermotion/UGC/开箱等多种视频风格)
- Claude Code — 作为创意代理的执行平台,通过 CLI 接入 Higgsfield + GWS CLI + Skill + Routines 构建完整自动化工作流
- Skill — AI 的”配方”——将满意的生成结果逆向工程为可复用的 Skill 文件,每次迭代自动优化
- MCP 模型上下文协议 — Claude 连接外部工具的协议,但在 Agent 场景下 CLI 更高效(Token 成本更低)
- GWS CLI — Google Workspace CLI,让 Agent 高效操作 Sheets/Docs/Gmail/Calendar/Drive 的命令行工具
- Routines — Claude Code 的定时任务功能,按设定周期自动注入 Prompt 执行工作流
与其他素材的关联
- 与 2026-05-10-codex-canva-operations-assets 的关系:两者都涉及”AI 批量生成营销素材”,但路径不同——Codex+Canva 走模板填充路线(选题库→Codex 生成数据→Canva 套图),Claude+Higgsfield 走 AI 原生生成路线(Prompt→Higgsfield 直出图像/视频)。本素材更适合视频和动态素材场景,Codex+Canva 更适合图文海报场景
- 与 2026-05-10-gpt-image-2-prompt-templates 的关系:GPT Image 2 的 Prompt-as-Code 强调结构化 JSON/YAML 提示词,本素材强调”逆向工程最佳输出为 Skill”——两种方法论互补,前者适合批量标准化,后者适合个性化高质量
- 与 2026-05-11-claude-code-6-skills 的关系:掘金作者的 6 个精选 Skill 中有 3 个是创作类(SEO Blog Writer / Newsletter Automation / Content Repurposer),本素材展示的是视觉创意类 Skill 的构建方法——从满意输出逆向工程而非从零设计,是 Skill 构建”逆向工程”路径的实战案例
原文精彩摘录
This stuff isn’t magic. This stuff just lets you automate things and ideate. So, my point being, if you’re not a master copywriter or advertiser, it might be really tough for you to build amazing tier-one advertisement copy and creatives. And that’s why what you can do with Claude is you can utilize other people’s expertise, and you can bring that in to make Claude code the subject matter expert here.
A skill is essentially a recipe for an AI agent. So, if someone said, “Hey, can you make me some chocolate chip pancakes?” you would pull up a recipe of chocolate chip pancakes, and you would make it, and the next time you would pull up that same recipe, and you would make the pancakes, and they’d be the exact same. But, if you didn’t have a recipe, and you were kind of guessing the measurements and guessing the order and the temperature, your pancakes would come out different every single time. So, when we give our agent a skill, it basically means, “Okay, whenever I want an Instagram ad, you do it exactly like this, and now everything feels on brand, everything feels consistent.”
We could have a Monday morning routine that picks 30 videos with a blank status, creates the prompts for all of them, generates all of them, and then we’d wake up on Monday morning with all of these completed with URLs and with job IDs. And then we could scale it up. Maybe we’re doing planning every Sunday and Thursday, and then maybe we’re doing generating every Monday and Friday, and maybe we were up from 50 to 100 or 200. And that’s how you sort of push the system to the point where it’s scaling way faster than you could as a human or even maybe multiple humans.