Making Claude Code My Head of Growth — Handing It a Month of Cross-Platform Data

I’ve been posting daily for nearly a month now — 32 videos so far, roughly ten thousand-plus new followers across all platforms. Not bad. Today I’ll try something different: dump all the backend data from all five of my platforms into Claude Code and see what kind of analysis, and what next-step recommendations, today’s AI can give me at the level of a short-video growth lead.

First, merge into one master table

Open the page and it starts with five core conclusions, telling me roughly where each platform stands. I’d exported the per-video data from every platform as Excel and handed all of it over, so it merges those sheets into one complete master table with the key metrics from each platform folded in — strong performers get a dark background cell, weak ones a light cell. A genuinely clear data overview.

Completion rate can’t be compared across the board

Next it goes platform by platform. For WeChat Channels, for instance, it built a quadrant chart showing me which episodes had both a high completion rate and solid view counts.

On completion rate, it classifies the videos first. Because WeChat Channels currently gives extra distribution support to videos over 10 minutes, it splits videos into two buckets — over 10 minutes and under 5 minutes 30 seconds — and compares within each. Long videos naturally have lower completion rates, so it doesn’t label a long video’s low completion as a bad result.

One more thing that matters a lot: because I write every episode’s full transcript with Claude Code, every one of them is on file, so it can analyze exactly why the underperformers underperformed — was the opening hook badly designed, or did the middle drag long enough that people lost patience? It also generalizes why the strong episodes hit high completion, turning that into validated “good opening formulas.” It does the same generalization for the weak ones.

Xiaohongshu: growth off one viral post isn’t actually healthy

Xiaohongshu is its own animal. The AI’s first read was that Xiaohongshu’s follow conversion looked excellent — two thousand-plus new followers. Then it dug in and found that a single viral post accounted for nearly 75% of that, which makes the number a lot less healthy than it looks. Unlike WeChat Channels, where every post picks up some followers and the platform is fairly forgiving of content, on Xiaohongshu a lot of posts get essentially zero distribution.

It also examined a rumor: the recent claim that content with a high save rate performs better overall on views and follow conversion. But in my own data, cover click-through rate is actually the single biggest driver of overall performance. It also gave me recommendations on how to optimize Xiaohongshu covers specifically for my audience and subject matter.

Douyin: off-platform traffic is the kill shot

Douyin’s problem is worse — after 30-plus episodes I’ve gained only 200-odd followers. The most fatal part: Douyin transcribes videos to text and runs content analysis on them, and the moment it catches you pointing viewers off-platform, it drops that post’s views to zero and tells you the content is unsuitable for public release.

By comparison, Xiaohongshu — and especially WeChat Channels — are far more forgiving; saying “comment and I’ll send you the link” in a video isn’t penalized there. But Douyin’s penalties are extremely strict, and its content red lines are tighter too. Some episodes did great on WeChat Channels — say, how to ban-proof CC, or how to do Netflix subtitles with AI — and got zero distribution on Douyin, blocked outright, with a warning that this category of content will keep causing problems.

Bilibili and YouTube: change the packaging, not the aspect ratio

Bilibili and YouTube share a typical situation: both are primarily landscape long-video platforms, my content is vertical, and I haven’t reworked any of it into landscape yet — so I’m naturally at a disadvantage on both. And both are slow-burn platforms: until the platform is confident you’ll keep producing quality consistently, it won’t push much distribution your way.

Bilibili also has a cover problem: even though I have AI produce covers in three different aspect ratios, those three ratios aren’t tuned per platform — and cover format varies enormously from platform to platform, as does how much it moves views. On Bilibili, content mentioning celebrities or trending topics converts better, while certain topics are very unpopular.

YouTube is even more clear-cut, since it’s currently aimed at overseas markets. The analysis shows the US accounts for 10% of traffic, with Southeast Asian countries also a meaningful share. But average watch time from US viewers is pretty long — over two minutes forty seconds — which tells you the people watching my videos on YouTube are actually overseas Chinese, because if you don’t understand Chinese there’s no way you’d watch that long. It also concluded that YouTube’s bottleneck is the cover, plus the language in my video descriptions: I’d naively used mixed Chinese-English titles at the start, but since the audience is Chinese speakers, mixing English in actually hurts overall CTR. So I’ll just optimize for overseas Chinese speakers. And since YouTube only uses 16:9 covers, I can optimize a dedicated version just for it.

Next-step recommendations per platform

After all that analysis, it gives me next-step recommendations per platform — topic selection, delivery pacing. Right now I actually don’t write a transcript at all; I talk as it comes to me, which is the better way for me to keep producing consistently given how busy I am. Even if it wrote one for me, it’d be hard to match my voice and quirks in the short term, and reading from a script also kills efficiency. So it gave me a middle path: skip the full transcript, but really polish and optimize the first 15 to 30 seconds so overall completion and retention improve.

For Xiaohongshu, the P0 is how topics and titles are phrased: listicles, or “here’s the tool I used,” don’t really pull Xiaohongshu users in. What works better is “I solved a bind you’re in, I have a better solution.” It also noted that switching to a business account’s lead-capture component would let me point people off-platform legitimately, instead of getting penalized for it the way I was before.

For Douyin, it recommends I run a self-check as soon as I finish recording, looking for anything that points viewers off-platform plus any sensitive keywords, and it reminds me to fix those at the editing stage. It also suggests publishing both a long and a short version on Douyin — cutting a highlights version. It recommends fixed publishing times on both Douyin and WeChat Channels. And I have found that, since my content is aimed at industry experts, senior executives, and some small business owners, publishing roughly between six and eight in the evening produces the best overall distribution — so publishing in a fixed window and controlling the relevant variables does help. For Bilibili and YouTube it’s a packaging-layer reset: don’t change the landscape-format content, just optimize the cover and the publishing copy.

It remembers on its own and iterates on the next episode

Finally, for later content positioning and topic selection, it built me a checklist and a to-do list, flagging item by item what I can do on each platform next. It also optimizes my daily production workflow the same way — anyone who’s watched my earlier videos knows that every day I finish recording, hand it to Claude Code, say “go publish it,” and it runs the whole pipeline. So it writes the next round of publishing-copy and cover fixes into its own memory and applies them on the next episode’s publishing run.

It also built me a topic bank I can keep pulling from for new content. Or if I just want to cover a certain topic today, it reviews it for me and tells me whether the topic is a fit and which angle I could take for better results — and I do adjust my short-video content to its recommendations.

Let’s come back in another month and see how the next month goes. That’s it for today, and I’ll see you in the next one.

Source: EP0034_audio.mp3 · ASR model gemini-2.5-pro (chunked parallel) · full text of the roughly 10-minute original recording

[00:00] I’ve been posting videos daily for close to a month now. So far I’ve put out 32 videos, and my followers across all platforms are up by roughly 10,000-plus. Still a pretty good result. So today I handed Claude Code all of the backend data from all five of my platforms, and let’s take a look together at what kind of data analysis AI can give us today at the level of short-video operations. And at the same time we can have it give me suggestions for the next round of optimization. So we’ve opened this page, and you can see it starts by listing

[00:26] five core conclusions — roughly what the situation is on each platform, a ballpark. And at the same time, I exported from every platform an Excel sheet with the per-video data and handed it over, and it merges them into one complete table, pulling the key metrics from each platform into that big table. It uses dark color blocks as the background for the ones that did well and light color blocks for the ones that didn’t. It’s a pretty clear master data table.

[00:51] And at the same time we can see it also did a detailed analysis for each platform. Take WeChat Channels — it made a quadrant chart like this, telling us which episodes had a high completion rate and decent view counts too. And then it keeps going deeper into the analysis. Take completion rate — it does a, it classifies the videos, because on WeChat Channels right now videos over 10 minutes get extra distribution.

[01:16] It splits the videos into over 10 minutes and under 5 minutes 30 seconds and compares them by category, because long videos naturally have lower completion rates, so it won’t define a long video’s low completion as a bad result. And likewise it can also get — because the transcript for every one of my videos is written with Claude Code, it has the transcript for every one of them. It can analyze what it is about the content that didn’t do well —

[01:41] why it wasn’t good. Is it because the opening hook wasn’t designed well, or because the middle of the content ran long so people didn’t have the patience to watch it all the way through. And there are also some, different, some already-proven good opening formulas — meaning for the episodes that did well, why they were able to hit completion rates that good. It did a certain amount of generalizing and summarizing.

[02:06] And for the episodes that didn’t do well, what the reason actually was — it summarized and generalized that too. Now Xiaohongshu’s situation is more distinctive. Originally AI’s analysis said Xiaohongshu’s follower conversion rate was excellent, because it gained roughly 2,000-plus followers. But within those 2,000-plus followers, only after its detailed analysis do you see that one breakout post actually accounted for nearly 75% of the growth, so

[02:31] it isn’t a very healthy result. Unlike WeChat Channels, where every single post gains a certain number of followers, and it’s relatively forgiving about content. But over on Xiaohongshu, a lot of the content is honestly in a state of getting no distribution at all. And it also analyzed that rumor about Xiaohongshu — there’s a rumor going around lately that content with a high save rate does relatively better on overall views and follower conversion, that its distribution is good.

[02:57] But actually, from what I’m seeing in practice on my end, it’s still cover click-through rate that has the biggest impact on overall results. And likewise here it also raised, raised how Xiaohongshu covers should be optimized, and given the audience I’m speaking to and the subject matter of my content, it did a — it gave a recommendation like that. Then take this, this Douyin.

[03:22] There’s an even more serious problem in there, because relatively speaking, on this whole platform, after 30-plus episodes went out it only gained 200-something followers. A very serious situation, and it’s because Douyin converts our videos into text and analyzes the content that way. Meaning the moment the content contains anything like off-platform traffic driving, it just takes that post and zeroes out its views. It says it isn’t suitable for public posting.

[03:48] Views just go, go to zero. But relatively speaking, Xiaohongshu and WeChat Channels, especially WeChat Channels, are relatively more forgiving. Mentioning things in the video like, hey, drop a comment and I’ll leave you the link — that content doesn’t get penalized. But Douyin’s penalties are extremely strict, and compared to the other platforms it has stricter content red lines. Like some episodes got great views on WeChat Channels — say, how to ban-proof CC,

[04:14] or how to do Netflix subtitles with AI. But that content actually gets zero distribution on Douyin, it’s completely blocked. So there’s some content where it’ll warn me that going forward this kind of content is going to be a problem. Now on the Bilibili side, and including YouTube, there’s a classic situation with both, which is that they’re primarily a landscape, long-video carrier.

[04:39] And my videos are vertical content, so naturally they aren’t a great fit for these two platforms. Plus I haven’t done the vertical-to-landscape reformatting of the source material yet, so on these two platforms I’m naturally at some disadvantage. And both of these platforms are slow burners — meaning until the platform is convinced you can consistently put out high-quality content, it isn’t going to push a lot of distribution your way straight off. But likewise you can also see

[05:04] that Bilibili is also a cover problem. Because right now, even though AI produces three different aspect ratios of cover for me, those three different aspect ratios of cover aren’t optimized per platform. And actually the standards for covers on different platforms, including how much they affect overall views, differ enormously. So over on Bilibili, when the content mentions a celebrity or some trending topic,

[05:29] its overall conversion is higher. And related to that, some topics are just really unpopular. Over on YouTube it’s even, even more obvious, because YouTube right now is aimed at the overseas market. And after the whole analysis, the US share is 10% of the traffic, and then other Southeast Asian countries also account for a fairly big share. But take the US viewers —

[05:56] their watch time, average watch time, is actually pretty long, a watch time of two minutes forty-something seconds. And that actually tells you the viewers watching my videos on YouTube are overseas Chinese, because if they didn’t understand Chinese there’d be no way to watch my video for that long a stretch. And likewise it worked out that YouTube’s bottleneck is actually the cover, including the wording in my video descriptions. At the start I just assumed

[06:21] that, that combining English and Chinese content was the way to do the title and the copy. But since what we saw this time is that my content is actually aimed at Chinese-speaking users, mixing Chinese and English hurts the overall CTR conversion. So I might as well just optimize the content for Chinese speakers, for overseas Chinese. Including, since YouTube only uses the 16:9 cover, I can give its cover

[06:46] its own separate optimization. And likewise, after it finishes this whole run of analysis, it gives me a next-step recommendation for each platform. For instance on topic selection, on the rhythm of the voiceover — because right now I don’t write a transcript at all, I just say whatever comes to mind. For me, as busy as I am, that’s a better way for me to keep putting out content consistently. So I’d give it the feedback that I actually have no way to do the transcript thing, and even if it wrote one for me, I think in the short term

[07:12] it still couldn’t match my, call it, language style and my quirks. Plus reading off a script actually lowers efficiency. So I compromise — the suggestion it gave me is that I can skip the transcript, but I can take the first 15 to 30 seconds of the opening and polish and optimize it better, so that my overall completion and retention get a good lift. And then Xiaohongshu’s P0 is actually topic selection and the way the headline is phrased.

[07:39] Meaning checklists, or what tool I’m using, don’t really pull in Xiaohongshu users. What works better is: I solved some predicament the user is stuck in, I have a solution — that’s a better, better approach. And it also mentions that switching, switching to a business account’s lead-capture component could catch the incoming traffic better, so we wouldn’t hit the situation from before where collecting leads got us penalized. And for Douyin it’ll also

[08:06] tell me I can, can run a self-check after this piece of content is recorded — is there any traffic-driving content in it, including any of these sensitive keywords — and it can prompt me to do some remediation later at the editing stage. And it also suggests I post two versions of different lengths on Douyin, that I could cut a highlights version. Including for both Douyin and WeChat Channels, it recommends I post at a fixed time. And first off, we did also, also find

[08:31] that because my content targets industry experts, corporate executives, and some small-business owners, basically posting in that 6 p.m. to 8 p.m. window gives the best overall distribution. So it does also recommend I post in a fixed time window, to control for some of these related variables. And as mentioned earlier, for Bilibili and YouTube I can actually do a degree of rework at the packaging level, rather than

[08:56] rebuilding the content as landscape. Meaning optimize, optimize the cover a bit, optimize the posting copy. And then including the positioning of the content going forward and the topic selection, it actually drew up a checklist for me, a to-do list — the things I can do next on each platform, marked out item by item. And likewise it’ll use this same approach

[09:21] to optimize my daily output workflow. Because those of you who’ve seen my earlier videos know that every day, once the video is recorded, I hand it to Claude Code and say go publish it, and it runs the whole flow end to end. So how the specific posting copy should be, should be optimized going forward, including how the cover should be optimized — it’ll commit all of it to memory itself and iterate on it during the next episode’s publishing flow. And going forward, so that my video content can better attract my audience, it also drew up

[09:46] a whole topic bank for me. So I can pick topics out of that bank and keep putting out my content. Or if today I just want to talk about a particular topic, it can run a review for me and say whether this topic is, is, is a fit, or that I could come at it from a certain angle to get a better result. And I’ll adjust the short-video content according to its recommendations. So let’s come back in another month and see how next month’s results look.

[10:11] So that’s it for today’s content. See you next episode, bye.