Voice DNA—Distilling a Creator's Style of Expression

Why do some creators’ videos make you want to watch to the end, while others make you swipe away in two seconds? The content isn’t that different; the expression is. Same knowledge, and one person’s delivery puts you to sleep while another’s pulls you further in. The part that pulls you in follows patterns—and those patterns can be extracted.

Handing over hit articles directly doesn’t work

When people want an AI to write something, the first instinct is to drop in a hit article and say write in this style. Pinpin started that way too, and it didn’t work—what came out felt mechanical, not like a person talking. The original creator sounds relaxed and colloquial; the AI’s version reads like reciting a script.

So try it the other way around. Instead of having the AI imitate directly, have it break down how the creator actually talks into a written spec for their language style, then use that spec to guide the writing.

Why Qiuzhi 2046

Qiuzhi 2046 is the example. He’s an AI-explainer creator on Bilibili with over a million followers and sharply distinct habits in how he talks—long videos, exceptionally good with analogies, instantly recognizable.

He’s published dozens of episodes, and not all of them are usable. Some are sponsored, some are Spring Festival Gala specials. We pulled his view counts first, picked out 14 episodes, and labeled each one by type—keep the AI tutorials, cut the ads. Seven of those 14 became the sample set.

Seven episodes become fifty thousand characters

With the samples chosen, step two is video to text. Download the audio with YT-DLP, run speech recognition through Gemini. One detail: the videos are long, so you cut them into three-minute chunks with some overlap, then fire the chunks at Gemini concurrently. Transcribing seven episodes yields 50,000 characters—about a thin book.

Quantifying how a person talks

With the text in hand, do the analysis in Claude Code. How a person talks comes down to numbers in the end—long sentences or short, most-used words, how they punctuate, how they use personal pronouns.

The result: Qiuzhi averages 27.6 characters per sentence, neither long nor short, right at the length of one natural breath. Too long reads as rambling, too short as rushed.

The second trait that jumps out: the pronoun is “we.” “We do this,” “let’s take a look together”—it strips out a lot of the lecturing tone. Then there are the opener patterns—“so,” “right?”, “so the second one,” “so the third one.” Every one of these rhythms can be written down.

Voice DNA: the language fingerprint

Numbers alone aren’t enough. We had the AI pull all of this together into a fingerprint of how the person speaks, called Voice DNA, something like an ID card. I read through a lot of other people’s best practices on Git and pulled out an approach that fits Chinese well. Sentence-length distribution, personal pronouns, particles, sentence-opening patterns, signature phrases—it’s all in the Voice DNA.

A few especially interesting findings:

Analogy first—before explaining any technical concept, Qiuzhi always leads with a plain, accessible comparison.

He keeps the failures in—when a demo goes wrong during recording, that part stays in. Not every demo works first try. It comes across as more real.

A banned-word list—he never uses the “folks,” “shocking,” and similar filler a lot of creators lean on. The AI compiled the list: words other creators reach for constantly and he never touches.

“Paramecium master”—he calls himself the paramecium master, using self-mockery to lower people’s anxiety about learning AI. Plenty of creators manufacture anxiety; he says the opposite—“this is a spoon-feeding tutorial, even a paramecium could follow it.”

Validation: writing a spoken script in his voice

Once the Voice DNA is extracted, how do you check that it works? I had the AI research a topic people have been arguing about lately—whether a beginner should learn Claude Code or Codex. Then it wrote the research up as a 3,300-character spoken script in Qiuzhi’s style.

The result is pretty good. The opening alone gives it away: “Hello there friends, I’m Qiuzhi, and today we’re talking about a question a lot of you have DM’d me about”—thoroughly colloquial. Then it breaks the comparison into five parts: barrier to entry, how deep you can go, customization, price, and use case. It closes the way Qiuzhi closes—congratulations, you’ve now got a framework for picking tools, materials and links are in the comments.

Send this piece to Qiuzhi and he’d take it for one of his own transcripts.

It works for any creator

The interesting part isn’t that AI can imitate a creator’s voice—it’s that there’s now genuinely a way to turn how a person talks into a document you can look at.

We turned this flow into a tool you can run again and again. For any other creator, all you do is tell Claude Code: use this skill on, say, Digital Life Kazik, and turn the way he talks into Voice DNA. After that, all that’s left is watching what’s trending, adding your own core viewpoints, adding the research—and you can write the way hits get written, or even invent your own way of writing them.

The Voice DNA skill is open-sourced on Git—download it and use it.

Source: EP0021_audio.mp3 · ASR model gemini-2.5-pro (chunked parallel) · full text of the 8:51 original recording

[00:00] Welcome back, friends. Watch this episode through and you’ll walk away with a method for pulling the expression code out of any creator’s hit videos, so AI can write content for you the same way they do.

[00:09] So why is it that with some creators’ videos you want to watch to the end the moment you start, and with others you want to swipe away after two seconds? It’s not that the content is that much worse—it’s that the way it’s expressed is different. Same piece of knowledge: one person explains it and you get sleepy, another person explains it and the more you listen the more hooked you get. And the part that hooks you actually follows patterns, and those patterns can be extracted.

Just Dumping in a Viral Script Doesn’t Work

[00:27] When a lot of people get AI to write something, their first move is to throw in the viral article and say, write it in this style, right. That’s how I did it at first too, and the results weren’t good—it usually came out feeling mechanical, not like something a person actually said. You’ve got a creator whose delivery is totally relaxed, spoken, conversational—and what the AI writes sounds like somebody reading off a script.

[00:47] So what do you do? I switched approaches. Instead of having the AI imitate directly, I have the AI take how this creator actually expresses things and break it down into a language-style manual, and then use that manual to guide the AI in writing the piece.

Why Qiuzhi 2046

[01:02] So today let’s use Qiuzhi 2046 as our example and see how to distill out his way of expressing things. Picking the source material, breaking out the fingerprint, validating the result—we’ll run through it all in one go. And as usual, I’ll publish this skill, open-source it on my Git, so you can go download it and use it.

[01:18] So the first question: why did I pick Qiuzhi? He’s an AI explainer creator on Bilibili, I like him a lot, and he’s got over a million followers. And he’s got a really distinct way of expressing things—first, he also does long videos, and he’s extremely good at explaining things through analogy, and he’s highly recognizable. And of course his numbers are strong too.

Picking Material: 7 out of 14 Episodes

[01:40] He’s published dozens of videos, and not all of them are usable, because some of them aren’t explainers—some are sponsored product videos, right, and some are Spring Festival Gala specials.

[01:52] So the first thing I did when building this skill was pull the view numbers on his videos, get 14 episodes out of that, and label the type of each one. The ones about AI tutorials I keep; the ones about special events or ads I cut. Out of the 14, I ended up picking seven that I thought were solid, as the samples to learn from.

Video to Text: 50,000 Words from Seven Episodes

[02:16] Once the samples are picked, step two is turning the video into text. Friends who’ve seen my earlier videos already know—for these seven episodes I can use YT-DLP to download the audio, and then run it through Gemini’s model to transcribe the speech into text.

[02:38] One small detail to mention: because his videos are long, my approach is to cut them into three-minute chunks, and make sure there’s some overlap between them. That way you can throw them into Gemini in parallel. After transcribing my seven episodes, I had a 50,000-word transcript—roughly the word count of a thin book. That’s the raw material for the analysis that comes next.

Quantifying How Someone Talks

[02:57] Okay, let’s keep going. Once we have the transcripts, we use Claude Code to do the analysis. So how do you find the patterns in there? The way a person talks can ultimately be quantified. Like, do they prefer long sentences or short ones, what’s their most-used word, how do they punctuate, right, what are their personal pronouns like. I analyzed all of these one by one, with Claude Code doing the work.

[03:23] For instance, the analysis comes back and shows his sentences average 27.6 characters—not long, not short, exactly the length of one natural breath. Too long and it sounds long-winded; too short and it sounds rushed.

[03:31] The second fairly obvious trait is that the pronoun he uses is “we”—we do this, we do that—which takes a lot of the preachiness out of it. It’s always “let’s take a look together.” That might just be how he talks, or it might be a deliberate choice.

[03:45] And he has some opener words, like “so,” right—“so, the second one,” “so, the third one.” Those rhythm traits in his language can be written down too.

Voice DNA: A Language Fingerprint

[03:55] Numbers alone aren’t enough, right. So I took all of this and had the AI draw a kind of fingerprint of the personality, of the language style. It’s like an ID card. Let me show you the final result—yeah, this one. This is called Voice DNA.

[04:18] And for this I also went looking on Git at a lot of other people’s best practices, and then pulled out an approach that fits Chinese-language work reasonably well. You can see in here the sentence-length distribution, the personal pronouns, the filler words, then the sentence-opening patterns, right, plus some signature expressions—all of it lives in this Voice DNA.

[04:41] So that’s the language style we get after we let the machine learn from it.

Qiuzhi’s Distinctive Traits

[04:47] Now, there are a few really interesting points in here. The first is that before Qiuzhi explains any technical concept, the analogy always comes first—he’ll always reach for a plain, easy-to-follow comparison.

[05:00] Second, when he does demos, he records the parts where it blows up too. It’s not that every demo of his works on the first try. That’s actually a pretty interesting point—it makes him feel more real.

[05:10] Another trait he has: a lot of creators use stuff like “hey fam,” or “shocking,” “oh my god”—he never uses any of that. So we also had the AI put together a banned-word list, words other creators use all the time that he doesn’t.

[05:25] And another interesting trait: he calls himself the Paramecium Master. That’s coming at it from the reverse angle, in a very self-deprecating way, to take the anxiety out of learning AI. Because right now so many creators are manufacturing anxiety—and he instead says, this is a spoon-feeding tutorial, a tutorial even a paramecium can follow.

Validation: A Voiceover Script in Qiuzhi’s Style

[05:47] So once we found these traits, the AI organized all of his ways of speaking into that Voice DNA JSON we just looked at. That thing is for agents to read, and the web page we saw a moment ago is for humans to look at.

[06:07] In here there’s a one-line style portrait of Qiuzhi, and then a role tree for how to imitate him, and then that banned-word list I mentioned.

[06:18] So how do we validate it? After all of that, I had the AI go do some research online, because a topic that’s been getting a lot of discussion lately is whether a beginner should learn Claude Code or learn Codex.

[06:34] So I took that topic and had the AI research it thoroughly online—all the comments and opinions out there. Then I took the research results and wrote them up in Qiuzhi’s style as a 3,300-word voiceover script.

[06:47] And the validation result was pretty good. Let’s have a look—this is the imitation the AI produced. See, this has real character to it: hey there, friends, right, I’m Qiuzhi, today let’s talk about a question a lot of you have DMed me about, so, Claude Code and OpenAI’s Codex, these two AI coding agents, which one should I actually pick?

[07:07] So a lot of people assume they’re the same thing, install one and you’re done, but actually no, right—that’s very conversational. And then he split it into the barrier to entry, split it into a few tiers, depth of control, right, then on top of that customization, then finally a price comparison, and use cases.

[07:23] And after all five of those, following his style, he’d do a little wrap-up, right—today we talked about the core differences, and so on. Then after using it for a while, what his prediction would be, and then finally congratulating everyone on picking up a framework for judging tool choices, materials and links in the comments.

[07:37] Very, very much like his style. I think if you sent this piece to him, he’d think it was his own transcript, right.

Takeaway: Turning How Someone Talks into a Visible Document

[07:44] So let’s wrap up today. What I personally find interesting isn’t that AI can imitate a creator’s expression—it’s that we now genuinely have a way to take how a person talks and turn it into a document you can see.

[07:58] Which means I can take that whole thinking-and-output process and make it a tool I can use over and over. Swap in any creator, and all I have to do is tell Claude Code: use this skill on, say, Digital Life Kazik, and turn his way of expressing things into a Voice DNA like this, a fingerprint like this, and use it to do the expressing for me.

[08:25] So all I have to do now is keep an eye on what’s trending online, add one or two core opinions of my own as comments, then add in the research, and I can write it up and deliver it in a viral format—or even invent a viral format that’s all my own.

[08:45] Let me know in the comments what you took away from today or what sparked something for you. See you next episode, bye-bye.