Original video in Chinese.
Key Takeaway
- Humans are the editors-in-chief, and AI is only an assistant: I strongly oppose “AI one-click generation.” By default, AI follows the highest-probability, safest path, so the content is correct but mediocre, with no real feelings or value judgment. I can let AI suggest topic directions, but I must make the final call on what to do, what core materials to use, what I feel, and what stance to take.
- Build the system first, then talk about production: write an instruction manual for the account (especially make clear “what not to do”), then build a topic bank, a materials bank, and a visual materials bank. Distill “content motifs” from the materials (which can be adapted across WeChat Official Accounts/Xiaohongshu/video), instead of one-off topics. Humans are responsible for judgment, the system is responsible for memory, and AI is responsible for execution.
- The goal is compounding, not one-off tasks: most AI tutorials only teach you to write one article or change one title. What this book aims to solve is organizing positioning, topics, materials, creation, and review into a reusable workflow, and even packaging it into Skills and automated tasks, so that every creation becomes the starting point for the next one.
To use Codex for creator media, first make your account positioning and content boundaries clear, then build three material libraries: topics, materials, and visual materials. Determine the content motif from your own experiences and judgment, organize the framework, first draft, and revisions by platform, and finally solidify the steps that already work into a workflow. Humans handle editorial judgment, AI handles execution, and batch generation must not be treated as a content system.
My new book, “Using Codex for Creator Media,” is now on sale. In addition to the book itself, there is also a companion video course and Skill. If you buy the book and join the group, you can claim them.
Modestly speaking, this book is probably the best tutorial on the market for teaching you how to use AI for creator media.
I’m so confident because the mainstream view and practice both advocate “AI one-click generation.” I am firmly against this approach.
I can tell you very clearly: anything generated by AI with one click is extremely mediocre. If you don’t believe me, go look at the 10,000-word long reads on WeChat Official Accounts—that’s a disaster zone.
I no longer read any so-called “10,000-word long reads.” It feels like 90% are AI-generated, with no nutrition at all—just the same words repeated over and over, a complete waste of time.
All this content is made by those AI creator-media people and AI boot camps. They have made money. But the entire content environment and content ecosystem has been polluted.
Why is AI one-click generation so mediocre?
Because AI defaults to generating along the highest-probability, safest path, so it is easy for it to be correct, but mediocre. The more you read, the more you’ll find there is nothing new there, and it is completely not worth reading.
Also, AI has no feelings and no value judgment—it can only pretend to be a person. But real feelings and unique value judgments are precisely the most precious things in content.
So one principle I consistently follow in this book is: humans are the editors-in-chief, and AI is only an assistant.
In other fields, I would let AI take over completely, such as programming. But in content, I absolutely will not let go. AI can propose topic directions, but what content to ultimately make, and what content can represent me, must be decided by me. AI can write drafts, but the core materials, real feelings, and value judgments in the draft must come from me.
Based on this principle, I teach you step by step in the book how to use Codex to build an AI-native content production system.
Most AI tutorials solve one-off tasks, such as writing one article or changing one title. What I want to solve in the book is how to organize account positioning, topics, materials, creation, review, and so on into a system that can be run repeatedly. So that you do not have to start from scratch every time.
In this system, humans are responsible for judgment, the system is responsible for memory, and AI is responsible for execution.
First, you need to write an instruction manual for your account.
For example: who is this account? Who is it for? What will it focus on over the long term? What positions will it hold? What content will it absolutely not do? What tone must never appear? And so on.
If these things are not clearly written, AI can only guess.
In my own account manual, what I will definitely emphasize is what not to do. For example, I do not make What-type content—that is, content that simply says who released a new AI tool, a new model, and so on. This kind of “what not to do” actually sets a very clear, strict boundary. AI then knows exactly which direction it cannot go. This works very well in practice, and everyone can try it.
With the account manual in place, the next step is to build three material libraries: a topic bank, a materials bank, and a visual materials bank.
Put sudden ideas into the topic bank. Put things you have personally experienced, opinions you noted while reading, and questions raised by users into the materials bank. Put cover and visual references you like in daily life into the visual materials bank.
The one most easily overlooked among these three material libraries is the visual materials bank. Because everyone thinks scripts come first, text comes first. But today, video covers are very important, and the shooting methods for B-roll inside videos are very important. If you do not deliberately save these visual elements, it is very hard to make them just from fuzzy memory and feeling.
After the materials have accumulated, we distill a content motif from them.
Why a content motif, rather than a topic?
For me, the definition of a topic is narrower, and it only fits a certain content format. For example, a topic may be suitable for a WeChat Official Account article, but not for a Xiaohongshu post or an episode of video.
Remember what I’m saying: on WeChat Official Accounts, you need to explain thoroughly; on Xiaohongshu, you need to be concrete; in video, you need to say it out loud. So a WeChat Official Account article cannot be used directly as a video script.
Compared with a topic, a content motif is broader. It can become an article, a post, or a video. You only need to make some adjustments according to the content format, and the conversion is complete.
In the video course that comes with the new book, I demonstrate a “one motif, three-platform conversion table.” When the time comes, do one according to the example, and you’ll understand the difference between explaining thoroughly, being concrete, and saying it out loud.
Only after the motif is set can you choose a specific content format, create the framework, write the first draft, revise, and finalize it. Once a workflow is running smoothly, it can be fixed into a workflow, and even packaged as a Skill. For work that needs to be repeated and done on a schedule, you can further turn it into automated tasks. I won’t say more about these here; once everyone gets the book, you can read it slowly.
So what this book really wants to teach you is not some magical prompt, but how to lock in your judgment, organize your materials, hand off repetitive work, and make every creation and review the starting point for the next one.
This is what I mean by an AI-native content production system: humans are responsible for judgment, the system is responsible for memory, and AI is responsible for execution.
So, with the book plus the course plus the tools, I hope that together they will help readers of this book build such a system, so that content production can compound.
OK, that’s it for this episode. If you want to understand AI, if you want to reclaim individual sovereignty, if you want to find like-minded people, come join our newtype community. See you next time!
From Account Manual to a Publishable Piece of Content
You can use the worksheet below to check whether the process in this article has truly been implemented. It is a method summary, not an added efficiency test:
| Stage | Deliverable Left Behind | Judgment the Person Needs to Make |
|---|---|---|
| Account positioning | Audience, theme, stance, and a do-not-do list | Which content can represent oneself |
| Material accumulation | Three material libraries: topics, materials, and visual references | Which experiences are real, which materials can be cited |
| Determining the motif | A one-sentence question and core judgment | Why it is worth discussing, and which platform it suits |
| Framework and first draft | Outline, source locations, and items to verify | Whether the argument holds, and whether there is any fabricated experience |
| Revision and publishing | Final draft and publishing checklist | Final expression, facts, and scope of disclosure |
| Review | Problem log and reusable steps | Which step is worth solidifying into a Skill |
Keep the first round to one piece of content. Do not have AI expand it into dozens of pieces all at once before the process has even run through once. Separate the facts in the materials, the author’s judgment, and the suggestions added by the model; if there is no original basis for a case, mark it as pending verification first, and do not write it as your own personal experience.
When using project execution tools, first confirm the working directory, file access, and approval settings. Being able to generate a draft locally does not mean you have publishing authorization. You can check the permission mechanism in OpenAI official safety documentation. This checklist was compiled on 2026-09-27, and no on-site testing of a specific version was added.
Next: Wiki: How creators build a Wiki, Wiki: Bringing PDF materials into a knowledge base, Wiki: Claude Code and Codex, Using AI to automatically edit videos.