Image Video Generation Workflow

基于 ComfyUI / MCP image server / 浏览器审看回路的图片与视频生成工作流技能卡，纳入 YouTube、Bilibili、Flickr 与 GitHub 的 source pack

Security Scan

This is a high-level, instruction-only workflow card for local image/video generation that does not request credentials, install code, or perform unexpected actions — it appears to do what it says.

Purpose & Capability

The skill's name and description (ComfyUI / MCP image server / browser review loop + source pack) align with the SKILL.md content: a human/agent-guided workflow for producing and curating images/videos. Minor mismatch: the frontmatter lists capability as "浏览器与执行" (browser and execution) but the instructions are high-level and don't declare any automated browser tooling, binaries, or commands; this is plausible for a manual workflow card but worth noting.

Instruction Scope

SKILL.md stays within the stated purpose: selecting a local generator, reviewing outputs in a browser, capturing stable prompts/params, and consulting GitHub/YouTube/Bilibili/Flickr for context. It does not instruct reading unrelated files, exfiltrating secrets, or calling unexpected endpoints.

Install Mechanism

No install spec and no code files are present (instruction-only). This minimizes on-disk risk — nothing will be downloaded or installed by the skill itself.

Credentials

The skill requires no environment variables, credentials, or config paths. That matches the workflow's purpose of guiding local/manual generation and curation.

Persistence & Privilege

always is false and there is no install step that modifies agent config or requests persistent privileges. The skill does not request elevated or persistent presence.

Assessment

This is a benign, high-level workflow card — it doesn't ask for credentials or install code. Before relying on it: (1) treat the SKILL.md as guidance, not an automated script — if you expect automation (browser control, headless runners), ask for explicit tooling and safe install instructions; (2) review any third-party GitHub repos linked before running their code (check licenses, README, and installation steps); (3) when using online source material (YouTube/Bilibili/Flickr), ensure you respect copyright and usage rights; (4) if you plan to automate browsing or remote downloads, require an explicit install spec and vetted release sources first.

When to Use

用户要快速搭一条本地可操作的图片/视频生成工作流，需要把生成引擎、浏览器审看、灵感素材收集和可复用提示词整理成一条稳定回路时使用。

What It Does

给出一条更接近当前 OpenClaw 采集现实的最小落地路径：

1. 用本地生成引擎产出第一轮图片或视频结果
2. 用浏览器统一审看生成结果
3. 用 YouTube / Bilibili 补工作流演示，用 Flickr 补视觉灵感，用 GitHub 补可执行入口
4. 只把稳定提示词、稳定工作流、稳定素材来源沉淀成可复用资产

How to Use

1. 先选主生成入口：
   • ComfyUI / Stable Diffusion：适合节点式图片和视频工作流
   • mcp-image：适合把图片生成与编辑接进 MCP 工具链
   • MeiGen-AI-Design-MCP：适合本地多方向并行生成与设计探索
2. 再建立浏览器审看回路：
   • 生成后统一在浏览器里看结果
   • 标记保留、废弃、待重跑
   • 把最稳定的一组提示词和参数回写到工作目录
3. 再按 source pack 补上下文，但不要混用角色：
   • GitHub：找真正可接入的仓库与工作流入口
   • YouTube / Bilibili：看实际演示、参数思路、回路编排
   • Flickr：补视觉灵感与素材方向，不把它当成可执行 workflow 文档
4. 只保留高价值沉淀：
   • 稳定提示词模板
   • 稳定工作流 JSON
   • 稳定素材来源清单
   • 明确的人工筛选规则

Source Pack Guidance

当前这张技能卡对应的已验证 source pack 是：

• youtube
• bilibili
• flickr
• github-search

使用顺序建议：

1. 先从 GitHub 选可执行入口
2. 再用 YouTube / Bilibili 补实操和审看回路
3. 最后用 Flickr 只补视觉参考，不让灵感源反客为主

Minimal Workflow

1. 选一个主生成入口并固定参数范围
2. 用同一主题跑 3 组图片或视频结果
3. 在浏览器统一审看并标记保留/废弃/待重跑
4. 查一个 GitHub 入口和一个视频演示，补最小执行路径
5. 只把保留下来的提示词、参数、审看结论沉淀成可重复工作流

Guardrails

• 不把 Flickr 这类灵感源当成实施说明书
• 不把单条视频教程当成唯一事实来源
• 先跑最小闭环，再增加模型、风格或后处理复杂度
• 每次只调整一组提示词或参数，避免失去对比性

References

• https://github.com/shinpr/mcp-image
• https://github.com/iconben/z-image-studio
• https://github.com/jau123/MeiGen-AI-Design-MCP
• https://www.youtube.com/watch?v=7aEQnTsI6zs
• https://www.bilibili.com/video/BV14mrfBSEdq/
• https://www.flickr.com/photos/dianeworland/55055024432/