Hi everyone,
It has been a bit boring recently. No big model or product launches, so I started digging into Anthropic skills.
I was traveling in the southern hemisphere when they launched it.
Then I saw that Tencent codebuddy now supports skills.
Before doing anything, I wanted to clarify the difference between Skills and MCP. They seem similar. So what is the real difference
After comparing and reading, the core difference is positioning.
Skills are skills.
Think of a person who can draw, browse the web, or write reports. Those are skills. They are Markdown docs stored in folders, and they still stay inside the Claude sandbox.
MCP is a protocol.
That protocol lets models access external services.
This diagram was everywhere when MCP got popular.
Once I understood the difference, I decided to define a few skills in codebuddy code.
Since I am new to skills, and I believe AI should do the work when possible, I cloned the official Anthropic skills repo first.
mkdir - p ~/.codebuddy & cd ~/.codebuddy
git clone https://github.com/anthropics/skills.git
Now you have the skills configured locally.
You can verify the setup like this:
If it lists available skills, the setup worked.
There is another setup method that looks cooler. Use the /plugin command.
Select Add marketplace, enter the repo name,
and load the skills library.
But these skills are just examples. My goal is to define my own skill.
So I found a doc template and asked codebuddy code how it was designed.
Then I let codebuddy code generate a new skills doc based on the template and my needs.
For example, there are many popular infographics. I do not know why they are so hot on X, but people share them all the time.
So I joined the trend and made something like this:
But the workflow for making such infographics has a problem. Even in the easiest case, I need to build a Gem in the Gemini app, then upload docs and open the web UI. It is not convenient.
This is where skills help. The terminal is always open, files are local, and if something is wrong, codebuddy code can fix it.
If this becomes a skill, it is easy to reuse every day.
The creation process is simple. Just say:
Follow the skills document format and create an infographic extraction skill. Prompt below...
After a bit of work, you can see the new skill in ~/.codebuddy/skills/.
Now I tried it.
When I reviewed the run, I noticed codebuddy code used Anthropic’s PDF skill to read the PDF file.
So you can also combine multiple skills together.
The final result:
Compared to the earlier sample image,
the overall style is consistent. That means the rewritten skill worked well.
Now lets see how it wrote the skill:
In a skill, the most critical piece is skill.md.
assets and reference are the output files and reference documents.
Inside skill.md,
You must follow the format and define the skill name and description at the top. The rest is flexible.
Beyond infographics, I realized another class of tasks is also perfect for skills.
Every model vendor has API definitions. I keep reading those docs, but I forget parameters over time. Tools like Postman are not very friendly to beginners.
So writing a skill makes sense.
For example, I wrapped the Ji Meng 4.0 API into a skill. The process is simple: find the Ji Meng 4.0 docs on Volcano Engine, copy the curl commands for text to image and image to image, and ask codebuddy code to turn them into a skill.
This time it even wrote the execution scripts under scripts and included prompt guides for image generation. It is definitely better than what I would write.
I do not think codebuddy code is just a coding agent. If it can write code, it can also write text. If it can search the web, it can read the local file system.
With skills, it can evolve further. If you want, it can do anything: audio, video, images, and more. It feels like everything can become a skill. We move from asking it to do tasks to teaching it reusable capabilities.
That is pretty cool, right?
Thanks for reading.
If this helped, please like, share, and follow. Do not miss updates and consider starring the account.