Original video in Chinese.
Key Takeaway
- Skills capability packs: Anthropic’s new feature packages SOPs/tools/resources into folder-based packs; an agent can tell what to do at a glance after opening them (for example, canvas-design generates PDF/PNG); scalable to different needs, with open-source GitHub examples.
- Layered loading: metadata stays resident to decide whether it’s needed, SKILL.md loads the steps, and scripts/docs are read when work begins; context grows step by step, avoiding overhead, and multiple packs can be mounted neatly.
- Development potential: modular design makes sharing/a marketplace easy (lower barrier than MCP), balances prompts and scripts; Anthropic is heading in the right direction, pushing native agent capabilities, and the author decided not to unsubscribe.
I had already been thinking about unsubscribing from Claude, but then Anthropic released Claude Skills, and I changed my mind.
I have a feeling that this thing is very important for the development of agents. Because at its core, it’s a kind of “format for packaging capabilities,” or simply put, a “capability pack.” In this pack, all the SOPs, tools, and resources needed for the agent to complete a specific task are bundled together. Open it once, and you know how to do it.
And Anthropic also designed a layered loading method that can greatly save an agent’s context window, allowing the agent to mount a large number of capability packs and choose freely without pressure.
Honestly, you can question Anthropic’s character as a company, but you can’t question their ability or their determination to push agent development forward.
Hello everyone, welcome back to my channel. Modestly speaking, I’m one of the few creators in China who can clearly explain both the Why and the How of AI. What I offer is far more valuable than tutorials. Remember to follow. If you want to connect with me, come join our newtype community. This community has been running for 600 days, and more than 1,700 friends have already paid to join.
If you’re a user in China, you can join through Knowledge Planet (Zhishi Xingqiu). If you’re overseas, you can join through Substack. My first course, daily newsletter, and exclusive videos are all available in the community.
Back to today’s topic: an agent capability pack — Claude Skills.
To experience this new feature inside Claude, you first need to turn it on in the backend.
Go to the Capabilities page, turn on code execution and file creation, and Skills will be available.
Anthropic has preset a bunch of Skills, and we can just check all of them. Then pick one to try, like canvas-design.
This Skill will guide the agent to output two files: one is a PDF version of the design concept, and the other is a PNG visual design.
Without any specific requirements, the agent chose to express the concept of time using mathematics and geometry. The whole process was smooth, and there wasn’t anything particularly special about it — just reading, thinking, and executing.
If you want to understand what exactly this capability pack called Skills looks like and how it works, you need to go to the official GitHub repository. Anthropic has open-sourced all of them.
Take the mcp-builder Skill as an example. You can already see from its folder structure:
First is a Markdown file called SKILL, which records in detail how to create a high-quality MCP server. It corresponds to the SOP.
Second is the scripts folder, which contains several scripts, such as two Python scripts, connections and evaluation. It corresponds to the tools.
Finally, there is the reference folder, which contains some Guidelines documents and reference examples. It corresponds to the resources.
SOPs, tools, and resources — these three are the classic components of a Skills capability pack. Tools can be all kinds of scripts. Resources can be reference examples, fonts, themes, and so on.
Of course, some simple capability packs won’t include everything. For example, the canvas-design I just demonstrated is SOP plus resources; brand-guidelines is even simpler, with just an SOP.
Anyway, different capabilities have different requirements — it all depends. With this capability pack, the agent can complete all kinds of specific tasks.
But then the question comes up again:
Model context windows are limited. Once there are too many Skills, will it become too heavy and affect the model’s decision-making?
Context is a key issue that must be considered when designing agents. In this regard, Anthropic uses a layered design.
The first layer is metadata.
Its content is very short — just the name plus a description. Because it’s so short, the metadata can stay resident in the agent’s mind. That way the agent knows exactly how many specific skills it has, and can load them when a relevant situation arises.
The second layer is SKILL.md.
When the agent determines that it needs to use this capability, it loads this Markdown document. Then it learns the full SOP, including the specific steps, how to use the tools and resources, and so on.
The third layer is scripts and reference documents.
Once the work officially begins, the agent reads the scripts in the scripts folder or the documents in the reference folder according to the SOP’s requirements.
You see, context increases progressively across these three layers. That avoids consuming too many context resources right from the start.
Claude Skills may look very simple — aren’t they just some folders and documents? But what I really like and admire is the design philosophy behind them.
One capability is one folder, one capability pack. All the SOPs, tools, and resources related to that capability are bundled together.
This modular design allows one agent to mount multiple capability packs. Then everything is loaded on demand, clean and tidy.
Claude Skills can find a balance between simple prompting and complex scripts — that’s impressive.
Also, folders are easy to understand, easy to share, and have almost no deployment barrier. So it’s very possible to form a Skills Market. The barrier to entry is much lower than an MCP Market.
Finally, apps and tools were actually not things native to the AI era; they were just expedient stopgaps. Capabilities, especially modular capability packs, are the things native to this era.
From MCP at the beginning of the year to Claude Code later on, and now Claude Skills, after using them all the way through, I have a very strong feeling: Anthropic is on the right path.
OK, that’s it for this episode. If you want to understand AI, want to become a super individual, and want to find like-minded people, come join our newtype community. See you in the next one!