As AI Agents gradually move from "chatting" to "doing things", an increasingly important question begins to emerge:In addition to generating text, can the Agent directly complete picture, video, audio and design tasks?
SamurAIGPT open source Generative-Media-Skills It is precisely solving this problem.
Generative-Media-Skills GitHub repository Encapsulation of a large number of generative media capabilities into Skills suitable for AI Agent calls allows Agents such as Claude Code, Cursor, Gemini CLI, and OpenCode to not only understand users 'creative needs, but also further transform needs into actual media production processes.
As of August 2026, the project's GitHub page shows that it has received more than 4,100 Stars and more than 470 Forks, and is still being updated.
1. It solves not "generating pictures", but "letting the Agent complete the creation"
Traditional AI picture or video tools usually follow this pattern:
User inputs Prompt → Model generation result → User downloads
This model is enough for a single creation, but it is far from enough for AI Agents.
What an Agent really needs is:
Understand requirements → select appropriate models → prepare materials → generate → edit → check results → continue to next step
For example, the user said:
"Make this product photo into a 10-second high-end advertising video."
This sentence actually contains multiple tasks:
- Understand product and brand positioning;
- Analyze original product pictures;
- Design visual style;
- Planning shots;
- Generate key pictures;
- Convert pictures into videos;
- Add appropriate exercise effects;
- Finally export advertising materials.
The core value of Generative-Media-Skills is to organize these complex tasks into Skill processes that agents can understand and execute。
Therefore, it is more like a "creative production layer" for AI agents, rather than just an API encapsulation.
2. Core + Library: The most noteworthy architecture of the project
This project uses a very clear Core / Library Two-layer architecture.
Core: Providing basic capabilities
/core The main underlying primitives in are:
- Upload media files;
- Photo editing;
- Platform certification;
- API call;
- Results polling.
These capabilities are not in themselves responsible for complex creative decisions, but rather provide stable infrastructure.
The project transfers the underlying call to muapi-cli, so developers don't need to rewrite curl, JSON parsing, and result polling logic for each model.
This is especially important for agents.
Agents prefer this interface:
生成图片
生成视频
编辑图片
上传素材
获取结果
Instead of facing a large number of API endpoints, HTTP headers, and JSON responses.
Library: Providing "expert knowledge"
what's really interesting is /library。
This is not simply a pile of API examples, but a code of professional creative knowledge into skills that AI Agents can execute.
For example, the project currently provides:
- Cinema Director
- Nano-Banana
- UI Designer
- Logo Creator
- Seedance 2
- AI Clipping
- YouTube Shorts
These skills are equivalent to adding a layer of "professional ability" to the Agent.
For example, a common model might only know:
"Generate a cinematic video."
Cinema Director type skills can further transform creativity into specific instructions such as camera movement, composition, lens language, light, scene and rhythm.
This is the very important design idea of the project:
Models are responsible for generation, and Skills are responsible for turning ideas into executable professional workflows.
3. 41 Recipes that turn complex creations into executable processes
The project currently provides 41 Workflow Recipe, covering Motion, Social, Visual and other directions.
The most noteworthy thing is that they are not a simple collection of Shell commands.
README describes Recipe as an end-to-end workflow that LLM is responsible for executing. Every Skill passes SKILL.md Describe the inputs and steps, and then an Agent such as Claude Code, Cursor, or MCP calls the underlying capabilities based on these steps.
This means:
Skill itself is more like an "AI creation SOP".
For example, a product advertising workflow can be abstracted as:
产品图片
↓
分析产品特征
↓
确定视觉方向
↓
设计广告画面
↓
生成关键帧
↓
关键帧 → 视频
↓
生成动态效果
↓
输出广告素材
For an Agent, this is much more powerful than calling the "Wensheng API" alone.
4. From video to social media: covering the entire content production chain
Another feature of Generative-Media-Skills is that it does not limit its focus to a particular media.
1. video generation
The project includes a large number of skills in video directions, such as:
- 3D Logo Animation
- AI Fight Scene Generator
- Animal Vlogger Video
- Cartoon Dance Animation
- Character Story Video
- Drone-Style Video
- Jewelry Product Video
- Music Video
- One-Shot Video
- Cinematic Product Ad
- UGC Video Factory
This means that the Agent can start from "one idea" and go all the way to the video.
In particular, a process like UGC Video Factory is essentially very close to commercial advertising production:
Character photos + product photos + scripts → vertical UGC advertising video.
2. Pictures and Design
Skills in visual direction are equally rich.
For example:
- Action Figure Generator
- Amazon Product Listing Pack
- Brand Kit
- Brochure Designer
- Fashion Try-On
- Floor Plan Rendering
- Interior Design
- Logo Generator
- Storyboard Generator
- YouTube Thumbnail
This further transformed the Agent's role from a "picture generator" into a small design studio.
For example, e-commerce sellers can propose:
"Help me make complete Amazon photo footage for this product."
Agents can generate main maps, lifestyle pictures, information maps, comparison maps and detail close-ups according to the established skills, without requiring users to design dozens of Prompt separately.
5. Real killer abilities: combination, not a single point model
One of the biggest problems with generative AI at present is that there are so many models.
Pictures include Midjourney, Flux, Gemini, etc.;
Videos include Kling, Seedance, Veo, etc.;
Music also includes Suno and others.
The real headache for users is often not:
"Is there a model?"
Instead:
“How exactly should I put these models together?”
The answer to Generative-Media-Skills is to solve it through Skills.
README shows that projects can access more than 100 AI models, and covers different types of generation capabilities such as Midjourney, Flux Kontext, Seed, Kling, and Veo.
As a result, models are no longer the core object that users directly face.
Users are faced with:
"Make a product advertisement."
The Agent then decides what model, what parameters, and what processes should be used behind it.
This actually comes from Model-centric to Task-centric transformation.
6. Why is MCP important?
The project also provides MCP Server.
The approach given in the README is:
muapi mcp serve
Through MCP, you can expose these media capabilities directly to Claude Desktop, Cursor, or other MCP-compatible Agents.
The changes this has brought are very obvious.
Traditional way:
ChatGPT / Claude
↓
告诉用户怎么调用 API
↓
用户自己运行脚本
MCP + Skills：
用户
↓
AI Agent
↓
理解创意
↓
选择 Skill
↓
调用媒体工具
↓
生成素材
↓
返回结果
In other words, AI has changed from "guiding users to using tools" to "Use your own tools to get the job done”。
This may be the direction that is really worth paying attention to in generative media in the Agent era.
7. Quick experience
The installation process of the project is also relatively straightforward.
first install muapi-cli：
npm install -g muapi-cli
Then configure the API Key:
muapi auth configure
Then install all Skills:
npx skills add SamurAIGPT/Generative-Media-Skills --all
You can also install only specific skills:
npx skills add SamurAIGPT/Generative-Media-Skills \
--skill muapi-media-generation
The project also supports installation for specific agents, such as Claude Code and Cursor.
After completion, you can directly generate basic pictures:
muapi image generate \
"a cyberpunk city at night" \
--model flux-dev
Or enter a more advanced Skill workflow.
For example, Cinema Director can be used for movie scenes:
bash scripts/generate-film.sh \
--subject "a cybernetic dragon over Tokyo" \
--intent "epic" \
--model "kling-v3.0-pro" \
--duration 10 \
--view
These examples illustrate that the project is trying to build a project based onNatural Language Creativity → Agent Skill → CLI/API → Media ResultsThe complete link.
8. What does it really represent?
If you only regard Generative-Media-Skills as a "collection of AI image/video generation tools", you will actually underestimate it.
What deserves more attention is:
First, Prompt is becoming Workflow
Past:
Write a good Prompt.
Now:
Design a complete Workflow.
Future:
Let the Agent choose and execute the Workflow itself.
Second, Skill is becoming a "professional knowledge plug-in" for Agents
A universal model knows many things, but that doesn't mean it's naturally a photographer, director, designer, or advertising creative director.
Skills can structure knowledge in these areas.
For example:
用户意图
↓
专业 Skill
↓
创作策略
↓
模型选择
↓
参数与 Prompt
↓
生成
Therefore, the Skill layer actually assumes the role of "professional experience".
Third, AI creation began to move from single models to model choreography
The truly powerful AI creation system in the future may not be:
"The strongest model."
Instead:
"Agents that can automatically combine the most suitable model based on tasks."
An advertising project may also require:
LLM
↓
脚本
Image Model
↓
视觉素材
Video Model
↓
动态镜头
Audio Model
↓
配乐 / 音效 / 配音
Editing
↓
最终成片
Generative-Media-Skills is already trying to standardize this model.
9. Who is the most suitable for use?
This project is particularly suitable for the following categories of people.
AI developer
If you are developing Claude Code, Cursor, or other Agents, you can directly learn from its Skill organization.
content creators
If you need to batch produce pictures, Short Video, advertisements, and social media footage, Recipe model can significantly reduce duplication of Prompt work.
e-commerce team
Workflows such as product advertising, product maps, Amazon Listing, and UGC advertising are all very close to actual business needs.
designer
Skills such as Logo, Brand Kit, UI, interior design, and brochures can be used as creative exploration tools.
Agent developer
More importantly, this project provides a paradigm worth studying:
How to transform the "creative process of human experts" into skills that agents can execute.
10. Conclusion: The next generation of AI creation tools may not be an App
The most noteworthy thing about Generative-Media-Skills is not how many models it gathers, but how it demonstrates a new product form:
AI Agent + Skills + Models + Workflow。
Models provide capabilities, Skills provide expertise, Workflow is responsible for organizing steps, and Agents are responsible for understanding user intentions and executing the entire process.
This means that when we use AI to create in the future, we may no longer need to:
Open the picture tool → write Prompt → download → open the video tool → write Prompt again → edit → export.
You may just need to tell the Agent:
"Make my product into a complete social media advertising campaign."
The rest is done by the Agent himself.
From this perspective, Generative-Media-Skills is not just a generative media toolbox, but is exploring a more important direction:
Let AI move from "generating content" to "completing creative tasks."