Original video in Chinese.
Key Takeaway
- AI OS functions: built with Claude Code + Kimi K2 Thinking. Skills include Super Writer (judgment/search/style analysis/writing) and Super Analyst (searching/multi-framework analysis); custom commands include new angle (spotting blind spots), weekly report (distilling/linking articles), XXX (breaking down posts).
- Setup approach: Skills handle complex tasks (SOP/MCP/Python/resources); custom commands are for repetitive simple needs (Markdown docs); folder pipeline: INPUT stores raw material, PROCESSING stores AI processing results, OUTPUT stores user notes, forming a searchable reference system.
- System metaphor: Skills/commands are like software, folders are like memory/hard drives, Claude Code framework drives the Kimi CPU; customize it according to your own needs and open the door to a new world.
The AI operating system I built has already become the core of my daily work.
If I want AI to help me write something, I start my own Skill — Super Writer. Once this capability pack is installed, AI’s writing ability improves dramatically.
It will first judge how complex the request is. Then, based on the reference materials I provide, it will search the web for more information. Next, based on the reference texts I provide, it will use MCP to call a set of analytical methods to distill the writing style that should be referenced. Finally, it will use MCP to call the most suitable writing methodology and complete the article.
If I want AI to help me analyze something, I start my own Skill — Super Analyst. Once this capability pack is installed, AI’s analytical and thinking abilities improve dramatically.
It will develop a very detailed search strategy to gather the materials completely. It will also call one to three different analytical methodologies and combine them to produce an analysis report.
If I want AI to help me spot omissions and fill in the gaps, I run a custom command called new angle. AI will switch to a completely new perspective, revisit the discussion we just had, and see whether there are any blind spots.
If I want AI to help me go through all the articles I’ve saved, I run a custom command called weekly report. AI will distill each article and, based on its own understanding, write a short piece telling me what it thinks is worth paying attention to. Not only that, it will also try to connect all these articles together and see what else it can discover.
If I want AI to help me analyze why a post went viral, I run a custom command called XXX. AI will turn into a seasoned self-media teacher and break it down from the opening hook, the closing call to action, and other aspects, listing out all the techniques used in it.
Everything I just talked about is done inside Claude Code, and the model I use is Kimi K2 Thinking. In the last video, I gave a quick introduction to this operating system. I could tell that a lot of people in the community were interested in it and wanted to build one too. So in this video, I’m going to go into detail about my setup philosophy.
Hello everyone, welcome back to my channel. Modestly speaking, I’m one of the few creators in China who can explain the Why and How of AI clearly. What I offer is worth far more than a tutorial. Remember to follow me. If you want to connect with me, come join our newtype community. This community has been running for 600 days, and more than 1,800 people have already paid to join.
If you’re a user in China, you can join through Knowledge Planet. If you’re overseas, you can join through Substack. My first course, daily newsletter, and exclusive videos are all available inside the community.
Back to today’s topic: the AI OS setup philosophy.
This approach starts from two questions: what kind of model should I choose, and what kind of Agent framework should I choose?
As for the model, someone asked me in the comments: why not use Claude’s model?
Two reasons:
First, I can indeed subscribe to Claude membership, like the $100 tier I’ve been using before. But all of that quota goes into coding, and there’s basically none left for daily use.
Second, using Claude’s model requires some special network setup. I also don’t want to be constrained by that; it feels very unfree.
So Kimi’s model became my first choice. It’s pretty good in capability, relatively cheap, and the official side has also adapted it for Claude Code — which is much smarter than what other domestic vendors are doing.
As for why I chose Claude Code as the Agent framework, it’s simple: I think, overall, it’s the best solution available right now, period. It has everything it should have. On top of that, there’s already a fairly mature community providing a wide variety of tools.
Also, if you use it in VS Code or Cursor, those software tools already have a folder system built in, which can be used to store and manage all documents.
So the combination of model, framework, and software becomes the foundation of my AI OS.
So how do I implement my daily work needs?
Through the two built-in Claude Code features: Skills and custom commands.
Skills are specifically for handling complex tasks, because they have the highest level of integration. In the prompt, you can define an SOP, call MCP tools at certain steps, or run Python scripts. You can also place all kinds of resources in folders, such as reference cases.
That’s why abilities like deep analysis have been turned into Skills by me.
Custom commands, on the other hand, are specifically for those simple tasks that need to be repeated often. I only need to create a Markdown document and write the requirements into it. The document name is the command name. When I want to call it normally, I just type a slash and it appears.
So I turned all those more fragmented needs into custom commands. As I said in the Knowledge Planet community, it’s actually very simple: just write out a rough version of the requirement yourself, and then let AI refine it.
So even with Skills and custom commands, that’s still not enough — you also need a folder system, or rather, a logic. My needs are all centered around information processing. So I designed three layers:
INPUT is specifically for storing raw material. This includes clipped articles and posts I’ve saved manually. These are all materials waiting to be processed.
After I use a custom command to have AI process them, AI will create Markdown documents inside the PROCESSING folder and store the results there. For example, the weekly report.
The materials in the PROCESSING folder are content that has been processed by AI but not yet digested by me. Once I’ve read everything, I’ll pick out the parts I think are useful, turn them into notes, and store them in the third folder, OUTPUT.
So you can see that these three layers of folders are a pipeline, processed in sequence by AI and then by me. Everything produced in the process is saved in Markdown format, and all of it can be retrieved by AI and used as reference for future answers.
This is the basic philosophy behind the personal AI operating system I built.
Skills and custom commands are like different kinds of software. Once they’re running, they can deliver the functions I want.
The three layers of folders are like memory and a hard drive, responsible for storing files of all sizes.
And all of this runs on the Claude Code framework, powered by the Kimi model as the CPU.
Once you understand my approach, you can design your own system according to your own situation. Trust me, once you build it, trust me, a new world will open up.
OK, that’s it for this episode. If you want to learn AI, want to become a super individual, and want to find like-minded people, come join our newtype community. See you in the next episode!