Original video in Chinese.
Key Takeaway
- Core architecture of newtype OS: a multi-Agent system customized for content production, made up of Chief (the main Agent, responsible for overall direction and final review) + Deputy (the assistant, responsible for task scheduling and coordinating sub-Agents) + multiple specialized sub-Agents (Writer, Editor, etc.) + custom Skills, creating an efficient and controllable production pipeline.
- Actual operating mechanism: Chief first sets clear acceptance criteria → creates an outline → assigns tasks → sub-Agents execute in parallel (writing + refinement) → Deputy summarizes → Chief reviews and iterates a second time → final output. The entire process is highly structured, avoiding context explosion and quality loss of control.
- Design philosophy: humans always hold the big picture and final decision-making power, while AI is only responsible for execution and amplification (rather than replacement); by presetting standards, constraining with Skills, and dividing labor among multiple Agents, the author upgrades AI from an “assistant” to a “professional content team,” significantly improving production efficiency.
I’ve finally optimized my content production system — newtype OS — to a pretty ideal state. Next, I’ll use one example to show you how it works and why I designed it this way. After watching this, you’ll understand why I’ve put so much effort into developing it and recommending it.
In newtype OS, I gave a very simple request: help me write a Newsletter, roughly about how intellectual deflation in the AI era will become more and more severe. Those “good students” who used to get jobs through knowledge and experience may go through a rough patch.
After receiving the request, Chief, the main Agent, did two things:
First, it retrieved my past documents. By analyzing the content, it got a clear understanding of my writing style.
Second, it launched a Skill called super-workflow. This Skill was specially customized by me for Chief. I’ve mentioned in the Knowledge Planet that the purpose of this Skill is to strengthen management of the entire production process.
Just like human management, if there are no institutional constraints, it’s really easy to slack off. This super-workflow Skill is a clear constraint on Chief.
You can see that under the requirements of super-workflow, Chief made a to-do list. The first step was to set the acceptance criteria. Why is this necessary?
Because today’s bottleneck for Agents isn’t whether the model is strong enough, but that it doesn’t have a clear standard to measure the result against.
Setting the acceptance criteria before starting means that when the sub-Agents return their work, the main Agent has a way to evaluate it. This is extremely, extremely important!
After defining the conditions for success and failure, Chief then provided a content outline. Why let it do that?
For work like content production, there can only be one core. Just like our human content team, it can only listen to the editor-in-chief. The editor-in-chief has the final decision-making power.
It’s the same in newtype OS. I designed many sub-Agents, such as Writer, Editor, and so on. They are all specialized Agents, each good at a particular area. But overall direction can only be handled by Chief. Chief communicates with the user at the front end, makes judgments, and gives instructions. That’s also why Chief has to be assigned the strongest model.
Once the outline was done, Chief handed the work over to Deputy.
If Chief is compared to the editor-in-chief, then Deputy is the deputy editor-in-chief, its assistant. Chief only communicates with the user, while Deputy only communicates with those sub-Agents. In this way, Chief’s context window won’t explode. When the mind isn’t chaotic, the work won’t be chaotic either.
Among the assigned tasks, in addition to having the Writer sub-Agent draft the article, the Editor is also required to do one round of refinement to make sure the writing style is fully aligned with my previous content.
The whole drafting and revising process takes about two or three minutes. After Deputy returns the result, Chief starts reviewing it.
Because the acceptance criteria were set at the beginning, the first step is to review according to the success criteria and failure conditions. After the standards-compliance review comes the quality-improvement review. Chief will directly point out a few areas that need adjustment.
All of this feedback is then sent back to Deputy, which calls the sub-Agents again for optimization.
Finally, Chief does a final check. It fixes some minor issues on the spot, instead of making human managers leave everything to their subordinates to change. Then, an article is finished.
What you just saw is the process of generating an article with newtype OS. It is a system fully customized for content production. From multi-Agent orchestration to Skill reinforcement, everything is aimed at making the whole process more efficient and more controllable.
In actual use, I don’t actually let Chief handle everything by itself the way I just demonstrated. I always first discuss it thoroughly with it, and I provide the logical framework. Then I let Chief write a draft according to that framework and make a sample version. If it’s okay, I’ll then write it myself.
This is actually my approach to all AI tools:
First, the big picture must absolutely be controlled and decided by me. AI can only act as a Copilot.
Second, I use AI not because I can’t do something — in fact, I’m very good at it. I use AI to amplify what I’m good at, and amplify it to the point where others can’t catch up.
That’s also the original intention behind why I made newtype OS.
As I said in the Knowledge Planet, I’ve already brought this system to the level of around 60 or 70 points. Next, I’ll add a self-summarization module to it. By then, newtype OS will be even more powerful. After watching the video, remember to try it out.
OK, that’s all 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!