Original video in Chinese.
Key Takeaway
- Kimi K2 Thinking advantages: excellent compatibility with Claude Code, handles complex workflows/Skills, costs far less than Claude; the author uses it as a daily Agent engine. In the past, domestic models were poorly adapted; now it’s possible to build an AI OS.
- AI OS requirements: improve information efficiency (text first, transcription/keyword extraction/summaries/structuring); strengthen thinking depth (built-in SOP and 12 frameworks in Super Analyst Skill for autonomous analysis) and breadth (custom commands to examine blind spots/switch perspectives/analogize/innovate).
- Practical expansion: phase one handles information/thinking (results saved in the Key Topics folder); phase two integrates notes/folders/new Skills; the Claude Code framework is powerful and has great potential beyond programming, and I encourage you to try it.
Kimi’s newly released K2 Thinking is really impressive! It works extremely well with Claude Code. Even Skills involving complex workflows and tool calls are no problem for it. And K2 Thinking is much cheaper than Claude.
This is exactly the kind of model I’ve been waiting for. Not to use it for programming, but as the driving engine for my everyday Agent.
Think about it: Claude Code can call so many tools and perform such complex retrieval and reasoning. If you move it from programming to other tasks, it should work just as well, and the results should be great!
On top of that, Claude Code has MCP, custom commands, Sub-agents, and now Skills too. It’s just so perfect for building a personal AI operating system.
But the reason I hadn’t started before was the cost.
Whether you subscribe or call it through the API, it’s troublesome and expensive.
So I kept thinking: if I could keep the Claude Code framework unchanged, but swap out the model inside it for one that’s also very capable, much cheaper, and easier to integrate, wouldn’t that be perfect?
Over the past six months, I’ve tried many models, especially domestic ones, but none of them quite cut it. There were reasons on both the model side and the compatibility side with the Claude Code framework. Kimi K2 Thinking, newly released, finally gave me hope. My AI OS has officially begun!
Hello everyone, welcome back to my channel. Modestly speaking, I’m one of the few creators in China who can explain both the Why and the How of AI clearly. What I offer is far more valuable than tutorials. Remember to follow me. If you want to connect with me, come to our newtype community. This community has been running for 600 days, and more than 1,800 friends have paid to join.
If you’re a domestic user, you can join via Knowledge Planet. If you’re overseas, you can join via Substack. My first course, daily newsletter, and exclusive videos are all available inside the community.
Back to today’s topic: building a personal AI OS based on Kimi K2 Thinking and Claude Code.
My basic requirements for this AI-powered operating system are twofold:
First, improve the efficiency with which I handle external information.
Second, strengthen my thinking ability from both depth and breadth.
Let’s start with the first.
Most of the information I need to consume every day is in text form. Even for YouTube videos, I’ll export the transcript. Because whether it’s me or AI, text is the most efficient medium to process.
A fair amount of this text content is clipped from web pages. I use a tool for this, one I recommended in a previous video: Obsidian Web Clipper. It’s just a browser extension. After clicking it, it saves the current webpage’s content to a designated folder.
So all the materials I need to digest later are stored in a folder called RAW, as in “raw meat.” And each week has its own subfolder.
The benefit of doing this is that I can process everything in one batch each week, without making a mess. And each time I process it, I just drag the whole subfolder into the chat box, and the AI knows it’s supposed to handle that folder.
Here, I use Claude Code’s built-in custom command feature. This feature is especially well-suited for repetitive work.
Whenever I’ve accumulated a week’s worth of materials, whether I’ve read the contents or not, I run this custom command. It instructs the Kimi K2 Thinking model to distill each article, including the Summary, Key Takeaway, and Something New.
The first two are fairly standard. Something New is something I specifically set up. I ask the AI to draw on its own knowledge to identify what in this content it personally finds “interesting” or “valuable,” and to express that idea with complete logic. If needed, it can even go online to search for some additional material.
The benefit of doing this is that it gives me a new perspective. Whether it’s useful or not, I welcome it. As long as it can give me some inspiration, even a tiny point, I’ll feel that it’s very valuable.
After the AI finishes processing the week’s materials, it generates a new Markdown document and saves it in a folder called Medium Rare. The folder I just mentioned is called RAW, or raw meat. Then after processing, it’s medium rare meat, which is already cooked enough for me to eat.
That’s the first part, improving the efficiency with which I handle external information.
Once information has been taken in, it has to be digested and thought through. In this stage, I need Kimi K2 Thinking to help me with two things.
One is to strengthen depth.
Whenever I have a question I want to explore thoroughly and deeply, I always have Kimi K2 Thinking start the Super Analyst Skill first.
I introduced this Skill in two previous videos. It has a built-in SOP I designed, including the processes of judgment, collection, and analysis. I also paired it with 12 analytical frameworks, and it can autonomously choose one to three of them depending on the situation.
Once this Skill runs, basically every aspect of the question I want to understand is covered. After that, I just need to pick out the things I’m interested in or still don’t understand, and discuss them further with Kimi.
And all of those discussions are saved in a folder called Key Topics. This includes both the results from running Super Analyst and my conversation with Kimi. This can all be done by the AI—you only need to create a custom command—but I still do it manually myself. It’s all just copy and paste, so there’s no need to waste tokens.
Two is to strengthen breadth.
It’s very easy for us to get trapped in our own thinking patterns. And AI may also just go along with whatever we say. So I’m especially cautious about this.
Once a discussion on a question is basically done, I want the AI to help me catch omissions and make up for gaps, and to see whether there are any new angles, and so on.
So I created this simple prompt and put it into a custom command. I ask the AI to examine the previous discussion from these four angles:
Blind spots and omissions. Are there any potential assumptions that haven’t been stated explicitly? Are there any questions or stakeholders that have been overlooked?
Perspective switching. If we look at this issue from a completely opposite position, what kind of outcome would there be?
Cross-domain analogy. For the issue we’re discussing now, such as a financial issue, are there similar examples in other fields, such as biology?
Innovation and inspiration. Based on all the information, if we really let our imagination run wild, what ideas might come up?
At the end of every round of thinking, I have the AI do this kind of scan, which helps me broaden the breadth of my thinking and find potential holes.
Likewise, I also copy this whole discussion process into a Markdown document and save it in the Key Topics folder.
What I just introduced is only phase one of my personal AI operating system. For phase two, I plan to bring over all my past notes and articles and build out the whole folder system. In addition, there are several more Skills to create.
So you can see, the Claude Code framework is really, really great. And Kimi’s new model can drive this framework too. I suggest everyone try it when you have time. It really isn’t just for programming.
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 next time!