Why Your "Remove AI Tone" Skill Barely Works

This episode flopped on the first take. I recorded it with my digital avatar, loaded up a full suite of human-tone skills, and the result was — polished nonsense. The language sounded right, but there was no logical thread, no “here’s the problem, here’s how it works” structure that I’m giving you now in person. So I scrapped it and re-shot. The flop itself is the best proof of what I want to talk about today: removing AI tone is not a one-skill problem.

People in my community have been downloading all sorts of de-AI-tone skills — including ones from well-known creators — and the feedback keeps coming back the same: barely works. My own “language-expression DNA” skill got the same complaint. So where does the problem actually lie?

Let’s start with what AI tone even is. A few things. AI invents words — compounds that don’t exist in natural Chinese but look plausible enough that the model just coins them. Then there’s the translation layer: most major models think in English and convert to Chinese on output, so you get this unmistakable translationese — awkward “it is not X, it is Y” constructions everywhere. Even a word like “loop” gets rendered as something closer to “ring” instead of the natural Chinese equivalent. That’s where the smell comes from.

Most de-AI-tone skills stack three layers of constraint. Layer one is formatting: continuous paragraphs, no numbered headings, no formulaic summary sentences, sometimes even banning formal punctuation. Layer two is speech habits: my skill, for instance, makes the model learn my verbal tics, throw in filler words like “好吧,” even sprinkle in the occasional typo on purpose. Layer three is discourse structure: some people naturally argue top-down, conclusion first, then evidence; others narrate linearly, guiding the reader step by step like a tour guide; some lean on analogies, some quote classics. All of that goes into the language DNA constraints. Three layers deep — and it’s still not enough.

Because here’s what people forget: every skill runs on top of a model, and models are wildly different. Even within the same company — 4.8, Opus 5, Opus 4.6 — the gap is huge. If you write with AI regularly, I’d suggest doing what I do: take the models with strong Chinese output — Kimi’s K3, DeepSeek’s latest V4, Opus 4.6 — run the same prompt and skill through each, compare side by side, then pick. In my experience, Opus 4.6 wins most rounds.

Still not enough. Take Claude Code: it has three modes — Chat, Cowork, and Code. Same prompt, same skill, different model API — and the three modes produce completely different output. Anyone who’s used it notices that Cowork output is notably better, genuinely human-feeling; even casual conversation in Cowork mode feels like talking to a real person. Code mode? Pure engineering-report style. The reason is the system prompt baked into Claude Code’s harness. Cowork is specifically optimized for text output — it essentially has that human-tone skill built into the foundation. Code mode’s system prompt opens with something like “You are an engineer who helps people write code.” With that sitting at the bottom of the stack, what skill you bolt on top barely matters.

This isn’t unsolvable, though. Claude Code has a setting called output style. The latest version ships with four built-in styles — Default, Proactive, Explanatory, Learning — but honestly none of them are great for this purpose. What you need to do is have Claude Code write a custom writing-focused output style, drop it into the output style options, pull it up via the slash-config command, close the window, reopen, and only then does it load a “speak like a human” system prompt at the base layer.

Even after all that, the output genuinely sounds human — but it still might not sound like a person with a point to make. That’s exactly why my first take flopped. The digital-avatar version had flavor in the language but couldn’t hold a logical thread.

So AI writing is really a two-stage job. Stage one is the first draft: what’s the core argument, what evidence supports it, what’s the narrative arc. At this stage you don’t worry about tone — you worry about getting all the information down in a coherent order and structure. Stage two is the polish — that’s where system prompts and skills earn their keep, shaping the output into something that reads like a person actually wrote it.

Try to collapse both stages into one skill and you get what my digital avatar produced: sentences that sound human, wrapped around an explanation that makes no sense. Companion materials are linked for this episode. See you tomorrow.

Source: EP0080 self-recorded on camera (re-recorded 2026-08-06, 4:47) · text transcribed with gemini-2.5-pro and corrected line by line · timecodes taken from the burned-in subtitles

[00:00] How to strip the AI flavour out of your writing This is a really hot topic Recently, a well-known content creator released his own skill for stripping out AI flavour But a lot of people in my group chat tried it out and found it’s pretty much useless I also get a lot of feedback saying that people used the one from my previous video the ‘Linguistic DNA’ skill and the articles it writes still have that AI flavour So what’s really going on here? Today, I’m going to get to the bottom of this First of all, what is this ‘AI flavour’? It really comes down to a few things First, AI makes up words

[00:28] Words that don’t actually exist in Chinese but they look plausible It’ll invent some obscure-sounding word Another thing is that many of the big overseas AI models think in English So when they output Chinese they’re actually performing a translation So the logic is English logic but the expression is Chinese And you can see this really strong translation-ese flavour Especially the way it says ‘it’s not this, it’s that’ You see that kind of phrasing all the time Or even, like I’ve mentioned before the word ‘loop’ In Chinese, we’d say ‘xúnhuán’ for ‘loop’ but the AI might translate it as ‘huán’, which just means ‘ring’

[00:55] So the question is, why do these skills we download from other people these de-AI-ifying skills just not work? Most of these skills contain several layers of constraints The first layer is a constraint on writing style It requires the AI to use natural language and express itself in continuous paragraphs It can’t use those “1, 2, 3, 4” headings and none of those formatted summary sentences The more extreme ones even forbid the AI from using formal punctuation That’s the most basic layer

[01:21] Then, going further, like in my own skill I make it learn my verbal habits I’ll add in some verbal tics like “alright” I even get a little clever and add some typos That’s the second layer So what’s the third layer? A person’s habits of expression Even in purely spoken language, some people have a pyramid structure Conclusion first, then the arguments Others have a linear narrative thread, like a tour guide leading you step-by-step to understand the content Everyone actually has a different stylistic structure to their language

[01:49] including whether they’re good with analogies or if they like to quote the classics These are all things that in my “language DNA” I put constraints on But don’t forget no matter what skill you use it’s always running on different… on top of the model even models from the same company 4.8, Opus 5 Opus 4.6 have huge differences So if you often use AI to write your scripts I suggest you do what I do Take Kimi’s K3 DeepSeek’s latest V4 and Opus 4.6 models with relatively strong Chinese language abilities

[02:15] use the same prompt and skill run a round of output and then compare and finally make your choice From what I’ve seen Opus 4.6 in most cases performs better But even that isn’t enough Take Claude Code for example We know that Claude Code has Chat, Cowork and Code modes, right? And in these three modes you can use the same prompt, the same skill and different models’ APIs to write the script and the results will be different too

[02:42] Especially for those of you using Claude Code you’ll find that Cowork’s output is exceptionally good It’s very human-like Even when you’re just chatting with it the conversation feels like it’s with a real person But Code mode is very engineering-report style, and that’s because in Claude Code inside its harness at its very core, there’s a system prompt Now, Cowork is specifically optimized for text output It’s as if it took what we just talked about that skill and built it right in, you know? But in Code mode the first line of its system prompt is

[03:09] something like “You are an engineer who helps people write code” With a foundational prompt like that at its core what kind of skill can you put on top that would actually work? Of course, this isn’t a completely unsolvable problem Now, for Codex I’m not too sure but in Claude Code there’s a setting called “output style” In the latest version it has four built-in output styles But none of the four are actually very good We need to ourselves get Claude Code to write a dedicated writing output style and put it into the

[03:35] output style setting in the options there then use the slash-config command find ‘output’ ‘style’ and load the one we wrote call up that config file then close the window and reopen it and it will be able to load that system prompt that makes it write like a person One last thing to keep in mind even when we’ve gone to this extreme it really does write like a person but it still might not sound like a person I actually had a failure with this very video because at first, for this video I tried to use my digital human to make it I loaded a whole set of

[04:02] ‘write-like-a-person’ skills for it but it was just spouting eloquent nonsense The language itself had a lot of flavour but it didn’t explain the whole thing clearly not like how I’m explaining it to you now, as a real person so clearly, step-by-step like, “oh, this is how it works” It didn’t have that kind of logical thread So you see, with writing with AI writing it’s actually a two-stage process The first stage, I call it the first draft stage In this stage, it’s about the core ideas the supporting evidence, and the entire narrative logic of how I should tell it It might not need to worry about the language It needs to get all the information right

[04:29] and output it in a logical order with a logical structure Then the second step is the polish the refinement stage So it’s by controlling the system prompt and by controlling the skill that we can finally generate a normal human-sounding piece of writing So that’s all for today I’ve left the resources for you See you tomorrow bye bye