Building an AI-Automated Topic Selection System: A Practical Guide to the Three-Dimension Method

Why Topic Selection Is the Hardest Part of Content Creation

For anyone running a content channel, finding the right topics is often the most painful yet most critical challenge. A strong topic largely determines content quality and final performance metrics. Bilibili UP creator JK shared that after growing 30,000 followers in six months as a side hustle, the question he got asked most in his DMs was: "How do you come up with your topics?"

His answer wasn't talent or luck — it was building a sustainable AI-automated topic selection system. Built on Codex, this system runs a topic pipeline that delivers a prioritized list of ideas to a Feishu spreadsheet every morning before he even gets out of bed. That "wake up and your topics are ready" experience is the core value of automation.

Background: Codex and Automated Topic Pipelines

Codex is OpenAI's code generation and automation model, capable of interpreting natural language instructions and turning them into executable scripts or workflows. Technically rooted in the GPT-3 family and fine-tuned for code tasks, Codex supports dozens of programming languages including Python and JavaScript. Architecturally, Codex functions as an "intent translator" — it can take a natural language request like "scrape the top 10 articles from 36Kr's AI channel every morning at 8am and write them into Feishu" and convert it directly into a deployable Python script, no programming background required. In the content creation space, Codex's core value lies in acting as a "glue layer" — connecting disparate data sources (RSS feeds, social media APIs, local file systems) into end-to-end automated pipelines. Around 2025, OpenAI further integrated Codex capabilities into the ChatGPT and API ecosystem, enabling non-technical users to describe their needs in plain language and have the model generate complete automation scripts. Feishu serves as the final dashboard layer, presenting AI-processed structured information in a human-readable format — completing the "AI processes, human decides" loop.

This article breaks down the core logic of this system — built on three dimensions: practice, trending topics, and benchmarking — all feeding into a single Feishu spreadsheet for unified prioritization.

Dimension One: Practice — Let What You've Done Naturally Become Content

The first principle JK emphasizes is: don't create content just for the sake of posting. Good content should grow organically from real practice; content creation is "the final step in a practice loop."

This is a genuinely insightful framing. If you treat content production as the starting point, creation becomes increasingly painful — you end up constantly forcing topics just to fill a schedule. Flip that around, and if content is the byproduct of real practice, topic ideas flow naturally.

He gives two examples: building a personal site with AI is itself a topic; building a topic selection system is also a topic. More cleverly, he has an Agent scan his conversation history with AI every day, pull out the most worth-sharing insights, and write them directly into the Feishu topic pool. The result: "You don't need to brainstorm topics separately — the things you've already done automatically become content."

Background: Memory-Augmented Agents and Workflow Datafication

Having an Agent scan conversation history touches on the concept of "Memory-Augmented Agents" in AI. These agents continuously read a user's historical interaction data — including chat logs, notes, and browsing behavior — to identify recurring interests and potential content opportunities. Technically, this typically involves vectorizing raw conversation data (Embedding), converting text into high-dimensional numerical vectors, then using cosine similarity retrieval or keyword extraction to surface segments with high "information density and narrative potential." Common implementation paths include using the OpenAI Embeddings API to vectorize conversation logs, storing them in vector databases like Pinecone or Chroma, and having an Agent trigger retrieval tasks on a schedule. The technical advantage of this architecture is incremental updating — each day's new conversations are simply appended as vectors without affecting historical data, and as the corpus grows, the system's accuracy in identifying the creator's interest patterns improves continuously. More importantly, personal workflows are knowledge validated through real practice — unlike topics conjured from thin air, they carry an inherent authenticity and credibility that resonates more deeply with audiences.

The essence of this approach is datafying your personal workflow — turning the byproducts of daily practice directly into content assets.

Dimension Two: Trending Topics — Use AI to Filter Industry News

The second dimension is staying on top of industry developments. But JK is clear: chasing trends must be done "from the foundation of growth within your own niche" — not as blind trend-following.

His approach: have AI automatically push a curated "AI Circle Daily Digest" every morning, filtered from 20+ articles based on his account positioning and audience profile, then written to Feishu and tagged as "Pending Review" by default.

That "Pending Review" design is key — AI handles the filtering and recommending; the final decision stays with the creator. If JK feels he can run with it, he does; if not, he skips it. His framing is clear: "Trending topics are a safety net for me, not a pressure to chase."

On the technical side, he suggests turning the news sites you follow into Skills connected to an Agent, enabling automated scraping and filtering.

Background: Tool Use and the Technical Principles Behind Skill Integration

Connecting news sites as Skills to an Agent is a textbook example of the "Tool Use / Function Calling" paradigm in modern large language model application development. Function Calling was formally introduced by OpenAI into the GPT-4 API in 2023, allowing developers to pre-define the input/output specifications of tool functions, with the model autonomously deciding when and how to call them during inference. Major AI Agent frameworks (LangChain, AutoGen, Dify, etc.) have built higher-level orchestration on top of this, supporting the wrapping of external APIs, web scrapers, and local scripts as callable tools. At the scheduling layer, Cron jobs or built-in platform "timed triggers" wake the Agent at a fixed time each day — the Agent then calls content-scraping tools, runs semantic matching and scoring, and writes results to Feishu via the API, all without human intervention. For content topic selection, this means an Agent can automatically visit platforms like Weibo Hot Search, 36Kr, and Sspai at a fixed daily time, extract titles and summaries, and perform semantic matching against the creator's pre-set "account positioning keywords" to output relevance scores. This transforms manual news-scrolling into a parameter-driven automated information funnel, dramatically reducing the time cost of information gathering.

The core idea: let AI handle the heavy lifting of information filtering; let humans do the value judgment.

Dimension Three: Benchmarking — Topic Signals Already Validated by Peers

The third dimension is benchmarking against peers — what quality creators are posting, what low-follower accounts are going viral with. These are ready-made topic signals.

JK has an Agent automatically scrape the previous day's content from the quality creators he follows on Xiaohongshu, generating a "Benchmark Content Daily." When he's short on ideas, he focuses specifically on content with high likes but low follower counts.

His logic is precise: if a low-follower account goes viral, the data wasn't built on an existing audience — the topic itself has genuine breakthrough power. This type of content has the highest replication and reference value.

Background: The Algorithm Logic Behind Low-Follower Viral Content

When a low-follower account achieves high engagement, the platform's recommendation algorithm is the mechanism behind it. Content platforms like Xiaohongshu and Bilibili universally employ a "decentralized distribution" strategy: new content is first pushed to a small cold-start traffic pool (typically hundreds to thousands of impressions), and the system uses early engagement metrics — like rate, save rate, completion rate, comment rate — to decide through multiple rounds of traffic competition whether to expand distribution. Xiaohongshu's CES (Content Engagement Score) system and Bilibili's "completion rate × engagement depth" composite model are both fundamentally using early sample data to predict a piece of content's long-tail distribution potential. Importantly, this competition mechanism treats large and small accounts equally — the system doesn't know the account's follower count at the initial distribution stage and relies purely on the content's early engagement data. This means a piece of content's reach depends on "content quality × user demand match," not the creator's follower base. A low-follower viral post is therefore a "pure signal" validated by the algorithm — it proves the topic had sufficient appeal in the cold-start phase, and the topic's scarcity or resonance has been backed by platform data. That's why JK treats low-follower viral content as the highest-priority reference signal in the benchmarking dimension.

When he spots a good topic, a browser plugin lets him sync it to the Feishu topic pool in one click.

Three Sources, One Table: Unified Priority Management

All three sources — practice, trending topics, and benchmarking — feed into a single Feishu spreadsheet, where the system automatically ranks them by priority. Every morning, the Agent checks the board: if there's a topic already in progress, it prompts continuation; if not, it recommends one.

JK specifically notes that topic weighting should be adjusted based on account type:

• He personally is a side-hustle exploration "time-for-money" creator, so practice and benchmarking are his primary sources, with trending topics and low-follower viral content as supplements;
• If you're a news/information creator, trending topics should be your main source;
• If you want to benchmark toward monetization, look for accounts with "10x the business scale you're aiming for, and a style you can actually learn from";
• If you just want to grow quickly in the short term, the low-follower viral content strategy is likely the better fit.

The same system framework, different weighting configurations for different creator types — this layered thinking is well worth borrowing.

Summary: Let AI Handle the Information Grunt Work, Save Your Judgment for What Matters

The essence of JK's topic selection system is an automated pipeline of information aggregation + AI filtering + human decision-making. Its value isn't in technical sophistication — it's in transforming topic selection from something that "depends on inspiration" to something that "depends on a system."

Each of the three dimensions has a distinct role: practice ensures originality and depth; trending topics ensure timeliness; benchmarking provides market validation. Together they complement each other, avoiding both the insularity of working in a vacuum and the aimlessness of chasing every trend.

Background: The Philosophy of Human-AI Division of Labor

The "AI does the grunt work, humans make the judgment" logic aligns closely with mainstream findings in Human-AI Collaboration research. Cognitive science shows that decision quality degrades significantly in information-overloaded environments — psychologist Barry Schwartz's "paradox of choice" theory demonstrates that once the number of options exceeds a certain threshold, both decision-maker satisfaction and efficiency drop. Manually browsing dozens of articles daily is itself a form of cognitive drain ("decision fatigue") that erodes a creator's judgment in the moments that truly matter. From an information theory perspective, AI's core value lies in separating signal from noise — compressing massive amounts of raw information into a small set of high-quality candidates, reducing the information entropy humans must process to reach a decision, while preserving final judgment for the human who understands context. AI's strength is retrieving, classifying, and ranking large volumes of information in milliseconds — but it cannot truly understand a creator's personal narrative style, their audience's emotional needs, or their current creative state. Multiple experimental studies from MIT and Stanford have shown that in creative tasks, a hybrid model of "AI generates candidates + human selects" significantly outperforms either pure AI generation or pure human brainstorming. This complementarity makes "AI narrows the candidate set, human makes the final call" the most efficiency-optimal collaboration mode in content production.

For creators looking to use AI to improve content production efficiency, the biggest takeaway from this system is: don't let AI create for you — let AI handle the information grunt work that happens before topic selection, and focus your energy on the steps that genuinely require judgment and creativity. According to JK, the next installment will cover how to use AI to help write viral voiceover scripts — definitely worth following.

Key Takeaways