I Made AI Do the Scrolling for My "Feel for What Trends"—Auto-Finding Benchmark Accounts + Catching Hot Topics

Fair warning up front: this video will probably get throttled, because it covers how to find benchmark accounts automatically and surface, every day, the hot topics you can ride. But I figure my followers will still see it, so let’s get into it.

Start with one word: a feel for what trends

Anyone who has run a creator business knows the phrase “feel for what trends”—you see a piece of content and you sense it immediately: total engagement is still low right now, but over the next two days this is going to be a big topic.

Where does that instinct come from? From scrolling video. Back when I was doing incubation and ongoing coaching, it took roughly five or six hours of video a day to develop that feel. The cost was brutal.

This episode is about using AI to automate the two pieces of it: picking topics and finding your benchmarks.

Step one: let AI work out how to get the data

I started with Claude Code and had it run a multi-agent deep research pass. The question, roughly: what existing best practices or approaches let an agent search out the day’s hot topics on its own and pull as much data as possible on a set of accounts, so I’m not searching one by one?

It gave me three technical routes:

1. Browser automation: use CC to drive a browser and do the browsing for me. The problems: it sits outside the rules, it gets rate-limited, and it only pulls a trickle of data per hour. You also need a browser a command line or an agent can actually drive—Chrome’s dev build, Microsoft’s browser. Cheap, but not that efficient—personally I don’t really recommend it.
2. Third-party data APIs (paid): some third-party services have already collected this data and let you call it over an API, so an agent like CC can pull it however you specify. Least hassle.
3. Official compliant APIs: most platforms simply don’t offer them, because this is their core data—unless you’re a major client, a commercial account, a big advertiser, or a large e-commerce player. The bar is extremely high (say, more than ¥5 million in annual spend for a single brand), so ordinary people can’t get in at all.

For each route the research also lists the actual options (MCP servers, for instance), which platforms each one covers, and how actively maintained it is. That last part matters a lot. Most of these third-party tools are put out free by individual developers, and one platform update can break the old method, so check whether anyone is still maintaining it. If something was still being updated on June 12, it’s probably still working now.

My pick: the third-party data API TikHub

I run five platforms: WeChat Channels, Douyin, Xiaohongshu, Bilibili, and YouTube, and naturally I want data from all five. Once you’ve looked at everything, there’s no real choice to make among third-party APIs—it’s TikHub, full stop. A lot of creators have talked about it too. I signed up and tried it. It has concurrency limits (concurrency stays low at the lower spend tiers), but it’s enough for me.

Technically it’s straightforward: register a GitHub account, hand its integration docs to Claude Code, and it works out “which endpoint gives me which data” without any trouble.

The benchmark account dashboard: leaders / dark horses / adjacent

Next, CC uses what it knows about me and what I’m trying to do (longer videos, long-term value, no news content, audience of industry experts and senior executives) to draw up a couple dozen keywords. Push those keywords through the API and it finds my benchmark accounts on each platform, then generates a benchmark account dashboard: 12 per platform, with follower counts, like-to-follower ratios, and what each one does.

I sort accounts into three buckets:

• Leader accounts: the biggest follower counts serving the same audience with the same kind of content. They set the bar, but I may not be able to learn from them—they’re already at that scale, and copying much of what they do won’t get me the same result.
• Dark horse accounts (my own tier): just started, only a few thousand followers, but a like-to-follower ratio above 5:1 or a save rate above 5%. These grow fast in the short term and are the most worth studying. (My account is a bit over a month old and has nearly 20,000 followers across platforms.)
• Adjacent accounts: they don’t talk about AI, but their audience matches mine (industry experts, senior executives), so they’re worth learning from too.

How to define “blowing up”: virality rate

Accounts alone aren’t enough. You also need the topics that are climbing today. I had CC measure that with a single metric: virality rate = total engagement ÷ hours since publication. A piece with 100 engagements that has only been up an hour has a virality rate of 100. That pulls out whatever is picking up engagement fastest, which gives you something to aim your own topics at.

I also have it pull each platform’s hottest topics of the day through the API—the gaokao, China’s college entrance exam, was one of them a while back. Those go into one master table alongside industry news, and then it gives me combined recommendations. On the gaokao, for instance, it suggested making a “red book” guide to help students with their application choices. That shows off both halves of AI: it can research, and it can turn research into a solution.

Combining hot topics with topic selection: tag and filter

I automated the whole thing. Claude Code runs on a schedule via GitHub and hands me every piece of content from the past 72 hours that meets my criteria, on any platform, with tags attached:

• Platform tag
• Whether it’s relevant to my business / audience
• Whether it came from a dark horse account
• Whether it’s news—I don’t want to make “the sky is falling, this is insane” news-chasing content, so I hide that category outright

Filter, then sort, and what’s left are topics people care about that also have real substance behind them. A lot of what it suggested—using AI to prepare an industry conference keynote, to prepare an investor pitch, to review contracts and investment term sheets, to run due diligence—are topics I’ve already made, and they performed well.

The real point: automating an operator’s feel for what trends

If I did creator business full time, I’d probably run this every two or three hours to catch the newest content and stay on top of trends; the traffic return compounds.

But the bigger point is this: a feel for what trends used to be experience an operator could only build by scrolling five or six hours of video a day, at enormous cost. AI can pin that experience down, scale it, and pull it out and lay it in front of you automatically. For anyone running a creator business, that changes a lot.

I have a lot more to say on this topic, and I’ll be releasing it over the next several episodes.

This is episode one of the “how to run a creator business” series (topic selection and monitoring). Coming up: writing with AI, and doing SEO and GEO with AI. Stay tuned.

Source: EP0045_audio.mp3 · ASR model gemini-2.5-pro (chunked parallel) · full text of the original recording

[00:00] Today’s video will very likely get throttled again, because it’s about how to automatically find benchmark accounts, and how to surface hot topics every day that we can ride. I figure my followers will still see it, so let me just talk about it. When I first started on this, I did it in Claude Code. The prompt I used went roughly like this: I need you to run a deep research pass with multiple agents working in parallel. What existing best practices or solutions can I use to get—

[00:25] social media data, without me searching by hand, so that an agent can automatically go find the day’s hot topics for me, or certain accounts, and pull as much data on them as possible. That’s the approach. And the result it came back with—you can see it actually gave me three technical routes. The first is going through a browser: basically using CC to drive a browser and browse, browse for us. The problem with that one is—

[00:51] well, it isn’t exactly compliant either, and it gets throttled—you can only pull so much data in any given stretch of time. The second category is the paid route, the one that’s least hassle. There are third-party services that have already gathered this data and let us pull it through an API, so an agent like CC can call that data however we ask. Then the third category is the official, compliant API. But most platforms don’t provide an official API like that, because what we’re asking for here is—

[01:16] core data for a lot of these platforms. Unless you’re a big client, a commercial client, or a major advertiser, a big e-commerce player, they won’t give you that kind of data interface. And for those three routes it also walked me through different solutions—MCP-based solutions, for instance. And it laid out which platforms’ data each solution can actually get. Now, I used to—right now I do WeChat Channels, Douyin and Xiaohongshu,

[01:42] plus Bilibili, plus YouTube—five platforms, right. So of course I want to be able to get data from all five. Looking at these solutions, the main things to check are whether the data they can get is complete, and how active they are. Because these third-party tools are usually built by developers like me and then handed out for everyone to use for free. But the platforms keep updating. Once a platform updates, the old method may stop working. So watch the activity. Take this MCP—the last activity on it is June 12,

[02:09] so it was still being updated on the 12th, which means it probably still works right now. Further down it also lists the login requirements for each solution, the risks and some of the problems with them. Then group two is browser automation, right. With browser automation we could actually build our own setup for it. But the one it pushes hardest can get all kinds of data from the platforms I want, and it’s a high-availability version.

[02:34] The author explicitly says commercial use is prohibited. Doing some data research for ourselves is fine. Personally I still don’t recommend the browser-automation route—it saves money, but it’s not that efficient. And then the browser base layer: that means when we go the browser-automation way, we have to set up a browser that a CLI, an agent, can drive. Options there include Chrome’s dev build and—

[02:59] Microsoft’s browser, right. All of those have to be configured in advance. Group three is the official compliant API, which, as I said, isn’t really usable—the bar is extremely high. For example, a single brand has to spend more than 5 million a year to get that data. There’s no way we’re getting that. So the option I chose is actually group four: the third-party API. And having looked at all of them,

[03:24] there’s pretty much no choice—it’s TikHub, that one company. I think a lot of creators have talked about this API too. I went and tried it out myself. All of its endpoints have—there’s a concurrency limit: at a given spend level you only get fairly low concurrency. But it’s enough. The volume of data it can pull at once is enough. So this API, once we’ve got—

[03:49] the endpoints, once we’ve got them, we can start screening accounts the way we want to. For instance, once I have the data I’ll tier the accounts. With benchmark accounts, first there’s the flagship account: same audience, same kind of content—which accounts have the most followers in the industry. That’s a flagship, but I can’t necessarily learn from it, because it’s already at that scale. A lot of what it does, even if—

[04:14] I copy it, won’t necessarily get me the same result. The second one is accounts at my own scale—the ones that just started, right. Look, my account has been running a month, a bit over a month, and it’s at close to—close to 20,000 followers across all platforms, which is a decent number. When you’re just starting out, look at the accounts that only have a few thousand followers but whose like-to-follower ratio is over 5:1, or whose views, plays against saves, whose save rate is over 5%. Those are all what we call dark-horse accounts—

[04:40] accounts that put on followers really well in a short window. Those are the ones we want to see. Then there are the ones I’d call cross-over accounts: they’re not talking about AI like I am, but their audience is the same as mine—industry experts and senior executives at companies. Accounts like that are worth learning from too. After that come some of the related technical pieces. Really, all we need is a GitHub account. Take its integration docs—

[05:06] and hand them to Claude Code to read, and it easily works out which endpoint gets which kind of data. Then step two: it goes off what it knows about me and what I want to do, right. I want longer-form video content with long-term value—the kind where you come back to it a few months later and it’s still inspiring, still gives you something. I’m not in the news business. And who my target audience is. Off all of that, off what I want to do, it lists out for me—

[05:31] about twenty-something keywords. I had CC list them for me. Then when you spread those keywords across the API and search, I can find what I’d call my benchmark accounts on every platform. So I had it generate a benchmark-account dashboard. All four domestic platforms are done—their follower counts, their like-to-follower ratios, and then what these accounts are about, all listed here. And this is only the first pass.

[05:56] For each platform it listed 12 of these accounts I can benchmark against and learn from. So what does it do next? Next it works two angles. I’ll have CC go find, through some keywords, the topics that are running hot in my industry right now. And how do you define hot? You add up all the engagement numbers and divide that whole figure by the number of hours from when it was posted until now. Say its engagement is 100

[06:21] and it’s only been up for one hour—then its heat rate is 100. That’s how I find the content in my industry that’s picking up engagement fast, and use it as reference for my own topics. And of course I’ll also have CC use this API to pull the hottest topics of the day on those platforms. For example, a while back the hottest topic was definitely the gaokao, China’s college entrance exam, right? So how do I tie into the gaokao? At that point it takes the day’s hot topics, the day’s industry hot topics,

[06:46] and lists them all in one master table for me. Then it takes what’s in that table and makes suggestions. Take the gaokao, right. Its suggestion was that I could build a “red book”—a handbook to help students with their application choices. That shows off AI’s research ability, and AI’s ability to turn research results into a solution. And it gives the people who are learning AI an idea to work with.

[07:11] And it shows off my technical chops to my own client base. So that’s how I normally think about picking topics. All I’ve done is automate that thinking—have Claude Code run on a schedule, go up to GitHub and do this for me. And once it’s built, it hands me everything from the past 72 hours that meets my criteria, whatever platform it’s on. And it tags them. First, it tags the platform.

[07:36] Second, whether it’s related to my business, to my audience—I tick a box and it filters. Then whether it came from a dark-horse account. Then whether it’s news. Like I mentioned, I don’t want to make that “the sky is falling,” “this is insane,” “a news story just dropped, you all need to know” stuff. I really don’t want to do that, so I hide that type from the list. Once it’s hidden, I can sort. Look at this one—live test, the full 45-minute dry-and-sanitize cycle on the actual unit, the Robam AI dishwasher, whatever it is.

[08:02] Or this one: gave AI 100,000 yuan to trade stocks, the big reveal 16 days later. From the time that content was posted to the time I scraped the data, it was averaging over 400 engagements every hour. That tells you it’s really hot—people are watching, the general public is interested, and it’s still a professional-content topic. So after that filtering, it gives me topic suggestions, and it factors in my own—

[08:27] approach to topics. It’ll say: you could do “AI helps me prepare a keynote for an industry conference”—I’ve actually done that one, and it went really hot. “AI helps me prepare an investor pitch.” Or this one, “AI reviews a contract, an investment term sheet,” right—that’s also a topic I’ve done, and it did really well. “AI does due diligence.” Those are all topics it gives me that I can use, short-term or long-term.

[08:52] Sometimes it also gives me hot-topic ideas, plus an angle on how I should tie into that hot topic. So by rights, if I were doing creator content professionally—if I spent all day figuring out how to run my channels—I’d probably choose to run this every two hours or every three hours, grab the newest content and chase the trend. The traffic payoff from that would definitely be an exponential payoff.

[09:17] Anyone who’s run a creator business probably knows this one thing: that thing people call a feel for what trends—it comes from scrolling videos. Back when I was doing incubation and ongoing coaching, I was scrolling maybe five, six hours of video a day—that’s what it took before I could see one video and instantly sense that even though its total engagement isn’t high right now, in the next day or two it’s definitely going to become a really hot topic. That’s the feel for what trends. But the price you pay for it is enormous. With AI in the mix,

[09:43] this whole topic-picking process can take that operator’s instinct and make it something you can actually see—pull it out and lay it out at scale, automatically. That’s a huge help for running a creator business. There’s a lot more I can say on this topic, so I’ll put it out over the next few episodes. Do keep an eye out for what’s coming. See you next episode, bye-bye.