One person doing content, daily updates without fail. So far, I've posted over a dozen long articles, averaging 120,000 views each, with a cumulative follower growth of 9,000+ and a stable bookmark rate of 0.5-1%—meaning one out of every hundred people who see it thinks the content is worth saving for later.
It's not that my writing is exceptional. It's because there's an AI content production system running behind the scenes—from topic selection, material sourcing, drafting, and illustration to data review, the entire process is executed by AI; I only make the judgments.
The prototype of this system comes from the Claude Code workflow shared publicly by @dontbesilent. He uses this approach to post 13,000 pieces of content a year, operating 7 platforms simultaneously, and gaining 700,000 followers annually. I took it and modified it significantly based on my needs for X long-form articles. This post shares the version I am currently using after those modifications.
Core Concept
@dontbesilent once mentioned a very fundamental issue: most people use AI for content in a fragmented way—they ask AI when they have an idea, post the answer, and then forget about it. Next time they have an idea, they start from scratch again.
His solution is to turn the entire process into a closed loop: ideas go into a topic library → AI searches the material library for reusable elements → write using a validated framework → publish → data review → distill effective patterns back into the methodology. Every creation adds something to the system, rather than reinventing the wheel every time.
I adopted this logic directly. Below is the version I modified.
Four-Layer Knowledge Base
I use Obsidian to manage content and Claude Code to execute. The system is split into four layers.
First Layer: Corpus.
The biggest problem with AI writing isn't that it writes poorly, but that it doesn't sound like you. Long-form readers read word by word; if the "AI flavor" is too strong, it feels off.
So I save everything I've said—tweets, viewpoints discussed in WeChat chat records, recordings, and fragmented thoughts jotted down. Then I extract a writing style guide from them: I like to state conclusions before giving reasons, I prefer numbers over adjectives, I like to use logic from other industries to explain the current matter, and I don't use "chicken soup" for closure.
AI reads this guide before every draft, so the first draft is at least 70-80% like me. After writing, I run a "de-AI flavor" check to highlight expressions that feel too mechanical for me to change.
What does it detect? Here are a few common pitfalls:
- Marketing buzzwords: empowerment, closed-loop, connecting, underlying logic—delete on sight.
- Speaking for the reader: "You might think..." "Many people will ask..."—how do you know what others think?
- Instructional tone: "Remember," "You must," "The core is just one sentence"—I'm chatting, not teaching a class.
- Fictional data: "90% of people don't know"—where did you get that 90%?
- Independent short sentences for dramatic effect: One sentence. One word. Paragraph. —This is the most "AI-flavored."
- Bolded slogans/golden quotes: Truly powerful people are all... —Delete.
These rules are stored in the system. AI automatically runs them after the first draft and marks hits in red. With these two steps, the "human touch" in long articles improves significantly.
Second Layer: Material Library.
Deconstructions of 47 similar accounts, data from over 1,100 pieces of content, structural analysis of viral articles, and reusable concepts and quotes.
Before writing a new article, AI first flips through the material library: who has written about similar topics, what angle got the data, and what structure readers are willing to save. It's not copying; it's choosing a path based on others' data.
After deconstructing 47 accounts, several findings directly influenced my topic strategy:
- Content with 1M+ views only falls into 5 categories: essential tool tutorials, medical/health science, AI + making money, persona analysis, and resource collections. Other types almost never exceed a million.
- Bookmark rates and exposure aren't necessarily positively correlated. Some articles have average exposure but high bookmark rates, indicating long-term value—these are worth writing repeatedly.
- Follower growth and exposure aren't necessarily positively correlated either. A persona post with 119K exposure gained 156 followers, while a tutorial with 77K exposure only gained 25. Personas make people want to follow the individual; tutorials make people save and leave.
Third Layer: Content Pipeline.
Topic Pool → To Be Deepened → In Progress → Published. The pool constantly holds a dozen topics ready to write and a dozen candidates needing more material. I don't just write whatever I feel like—I pick from the pool based on strategy.
Topics rotate through several tracks: project practice, AI money-making track deconstruction, low-threshold grassroots business, and new AI paradigm trends. Each track has different intensity—hardcore tool tutorials get the highest exposure, persona introductions grow followers fastest, and data reviews have a narrow audience but good bookmark rates. I choose topics based on current goals: tutorials for exposure, personas for followers, and reviews for long-term value.
Fourth Layer: Methodology.
What titles are effective, what topics go viral, what structures have high bookmark rates—all distilled from my own publishing data.
Titles are the easiest part to quantify. After a dozen long articles, the titles that perform well basically fall into four patterns:
Check before posting: Are there specific numbers? Is there an identity tag? Is there a contrast? Does the reader know what they will get after reading the title? The more hits, the better the data.
Illustrations
Illustrations for X long articles are crucial. In the feed, the user's attention order is HERO image > Title > Body. If the image is bad, no one clicks regardless of the title.
My principle: The HERO image, title, and hook trio should not repeat information. The HERO image tells you at a glance "what type of content this is," the title provides a data anchor to make people stop, and the first paragraph of the body expands on the details. Three things convey three different layers of information.
There are two styles of illustrations, automatically selected based on content type:
Tutorials use infographics—white background, light-colored decorative bubbles, rounded cards, flat icons, and large Chinese titles, like a clean hero banner on a SaaS website. Opinion pieces use conceptual posters—large text as the frame, with characters and text interlocking, like an exhibition poster rather than a PPT.
Each long article gets one cover plus two or three internal infographics. AI generates prompts based on the article content, calls the GPT Image 2 API to produce images, and then I download and crop them to the required ratio. What used to take half an hour in Canva now takes 10 minutes for three images.
Long-form Data
Here are a few representative ones:
Average exposure is around 120,000, with a bookmark rate of 0.5-1%. The AI fortune-telling post had the highest bookmark rate at 1.01%—the combination of AI + making money + information asymmetry makes readers save most actively.
Patterns Grown from Data
"Growing rules from data" is dontbesilent's core methodology. Here are specific patterns derived from my own X long-form data:
Titles must have specific numbers. "100k monetization in 4 months," "$155 vs $15," "452% ROI"—all successful long articles carry hard numbers. Numbers are the easiest thing to make people stop in a feed.
AI must be the protagonist. AI tutorial articles consistently stay above 100,000 views, while pure investment content rarely exceeds 50,000. People come to this account to see "how to use AI," not "how to trade stocks."
"Helping you save time" is the underlying logic of virality. Collecting public accounts, Codex introductions, illustration practice—the commonality of all viral long articles is "I've tried it, hit the pitfalls, and organized it for you; just follow along."
Viral Formula: Hardcore tutorial or real experience + Specific data anchor + Reproducible path. No viral title is an abstract concept. They all follow the structure of "I did X, and the result was Y"—sharing experiences plus data, not lecturing.
These rules are updated with every new article posted. The system is self-correcting.
You Can Use It Directly
dontbesilent's dbskill (4000+ stars on GitHub) is a great starting point. You can also do what I did: take his core ideas and modify them according to your own needs.
You don't have to get it right in one step. Start by building your topic pool and material library, run it for two weeks, and let the data tell you which direction to adjust.
