This is a de-AI-izer, not a "humanizer". It removes AI tics, it doesn't add human voice.
This tool appears primarily designed to produce text that's clearer and more concise, not to produce text that fools someone into thinking a human wrote it.
You're exactly right. I've found STE100 formatted responses are a godsend to turn AI responses into something concise, understandable and usable; it's great for creating documentation and such. If someone wants AI as a replacement for actual human interaction, you won't fool anyone.
Does it consider word rhythm in a sentence? When I revise what I write, I hear the words in my head and make sure they sound write together, and often across sentences. I wonder where that came from, and if AI can replicate it. Maybe it’s probabilistic?
Not directly. One rule says to vary sentence length so the text does not fall into a fixed pattern, but that is a blunt tool. The sound of the words in your head is still your job.
The biggest failing is that, when I write something, I have a theory of the world and the mind of the person or the people who might read it, and I adjust accordingly. Current text generation systems completely miss this point no matter how nice the writing can be made.
Functionally speaking, i.e. talking about observable output and not making any claims about LLM "mental states", there are plenty of examples where LLMs demonstrate good theories of the world and theory of mind. They can even exhibit, again from a functional perspective, emotional intelligence.
However, they need input to tell them what the world in question is like and what the minds are like. Otherwise, they're just operating from an enormous amount of often-conflicting pretrained data.
You're more likely to see good results in this area when interacting directly with one, so that it has some history of the interlocutor's writing, or when doing something like coding where "the world" is something like a codebase, along with the pretrained information it has about that kind of code.
Code is a good example, because the results they give in that context would not be possible without some equivalent of a theory of the subject. This applies to pretty much any kind of reasoning that models do - in math, logic, physics etc.
When writing for a more general audience, models lack this kind of information, so you would have to supply it in some form.
The other issue is that whatever theories are implied by their writing are not necessary stable, because of issues like context size limitations, compaction, selective attention, etc. Nevertheless, it's a big oversimplification to say that they don't possess the functional equivalent of such theories.
Fair point. It removes AI tics and makes text clearer. It does not add a voice, and it does not try to fool a detector. "Humanizer" is the name of the skill it builds on (blader/humanizer), so I kept it.
Very close. SimpleEnglish uses the same base, STE. The main difference: this one first runs a pass for known AI patterns (from Wikipedia's "Signs of AI writing"), then applies the STE rules. SimpleEnglish has much more measured evals across models. This one does not have that yet.
This is a great idea. I dropped an instruction to use STE-like language in my Claude global instructions. Opus and Fable don't seem to respect it but Sonnet very much does.
wow even just asking it for 80% of the way to ASD-STE100 makes a huge difference in readability. I'm experiencing way less brain fog trying to read opus.
Am I crazy or does the skill itself look like slop, complete with disclaimers back to the user embedded into the skill file itself?
"A warning about the original post. It came with an overview graphic of STE. A later fact check found errors in that graphic. For example, "approximately" is an approved word, and the dictionary has no entry for "in order to". This skill follows the rules of Issue 9, not the graphic."
I'm totally fine with that, but I think there are still quality standards that if a bit higher, would let the skill at least be much smaller and thus less token-consuming. One place where I draw the line is when somebody, with minimal effort using an LLM, could have improved the quality of the LLM generated output. Also, I treat skills like programs and skim all external skills to make sure that they're not doing things they shouldn't be doing.
Let me ask it straight: WHY? Why, for heaven's sake? The only goal is to make AI detection harder, and that directly translates to more uneducated students. So why? What's the point? What's the ethos?
There are a lot of comments here every day about how difficult it is to read AI writing and how grating people find it. I think it would be a big improvement if that discussion could be drawn to a close.
This is a de-AI-izer, not a "humanizer". It removes AI tics, it doesn't add human voice.
This tool appears primarily designed to produce text that's clearer and more concise, not to produce text that fools someone into thinking a human wrote it.
You're exactly right. I've found STE100 formatted responses are a godsend to turn AI responses into something concise, understandable and usable; it's great for creating documentation and such. If someone wants AI as a replacement for actual human interaction, you won't fool anyone.
Does it consider word rhythm in a sentence? When I revise what I write, I hear the words in my head and make sure they sound write together, and often across sentences. I wonder where that came from, and if AI can replicate it. Maybe it’s probabilistic?
Scansion is a technique for measuring this so it is possible to automate without AI
Not directly. One rule says to vary sentence length so the text does not fall into a fixed pattern, but that is a blunt tool. The sound of the words in your head is still your job.
The biggest failing is that, when I write something, I have a theory of the world and the mind of the person or the people who might read it, and I adjust accordingly. Current text generation systems completely miss this point no matter how nice the writing can be made.
It's not that simple.
Functionally speaking, i.e. talking about observable output and not making any claims about LLM "mental states", there are plenty of examples where LLMs demonstrate good theories of the world and theory of mind. They can even exhibit, again from a functional perspective, emotional intelligence.
However, they need input to tell them what the world in question is like and what the minds are like. Otherwise, they're just operating from an enormous amount of often-conflicting pretrained data.
You're more likely to see good results in this area when interacting directly with one, so that it has some history of the interlocutor's writing, or when doing something like coding where "the world" is something like a codebase, along with the pretrained information it has about that kind of code.
Code is a good example, because the results they give in that context would not be possible without some equivalent of a theory of the subject. This applies to pretty much any kind of reasoning that models do - in math, logic, physics etc.
When writing for a more general audience, models lack this kind of information, so you would have to supply it in some form.
The other issue is that whatever theories are implied by their writing are not necessary stable, because of issues like context size limitations, compaction, selective attention, etc. Nevertheless, it's a big oversimplification to say that they don't possess the functional equivalent of such theories.
Fair point. It removes AI tics and makes text clearer. It does not add a voice, and it does not try to fool a detector. "Humanizer" is the name of the skill it builds on (blader/humanizer), so I kept it.
Curious how this is different from simple english [0][1]. I've tested that one across multiple harnesses, and it works pretty well.
[0] https://github.com/AminBlg/SimpleEnglish
[1] https://news.ycombinator.com/item?id=49114639
Very close. SimpleEnglish uses the same base, STE. The main difference: this one first runs a pass for known AI patterns (from Wikipedia's "Signs of AI writing"), then applies the STE rules. SimpleEnglish has much more measured evals across models. This one does not have that yet.
It's quite similar.
This is a great idea. I dropped an instruction to use STE-like language in my Claude global instructions. Opus and Fable don't seem to respect it but Sonnet very much does.
I find 79% to be the sweet spot.
wow even just asking it for 80% of the way to ASD-STE100 makes a huge difference in readability. I'm experiencing way less brain fog trying to read opus.
ISO 24495-1 is the way for technical writing.
Very cool, thanks for this! Looks super useful for all kinds of copy..
Wonder if it also defeats text watermarking.
Should look into tropes.fyi
Thanks, I will look at it for patterns the list does not have yet.
Am I crazy or I can't even find the SKILL.md
https://github.com/Dodaios/ste80-humanizer/blob/main/SKILL.m...
Am I crazy or does the skill itself look like slop, complete with disclaimers back to the user embedded into the skill file itself?
"A warning about the original post. It came with an overview graphic of STE. A later fact check found errors in that graphic. For example, "approximately" is an approved word, and the dictionary has no entry for "in order to". This skill follows the rules of Issue 9, not the graphic."
Yes it absolutely is slop, full of AI tells. Why do we need all this ridiculous ceremony for a text file of instructions anyway?
If I want to control how it writes I just put a style_guide.md in the context.
Honest question: why does it matter if the skill is slop? It's designed for use by an LLM, I'm not planning on reading it.
I'm totally fine with that, but I think there are still quality standards that if a bit higher, would let the skill at least be much smaller and thus less token-consuming. One place where I draw the line is when somebody, with minimal effort using an LLM, could have improved the quality of the LLM generated output. Also, I treat skills like programs and skim all external skills to make sure that they're not doing things they shouldn't be doing.
Well it matters if you’re trying to adjust writing quality higher yet giving it terrible writing in the context.
The more slop you have in your context the sloppier the text you get out will be.
Why does a SKILL.md include “sources and credits”? The entire thing is pure slop.
> Karpathy talks about readability, not about sounding human. This skill adds that part, because the two methods cover each other's gaps.
I guess the author forgot to use his own skill.
Fair hit. That sentence had the exact pattern the skill removes. Fixed.
> That sentence had the exact pattern the skill removes.
Even that sentence sounds AI to me. Humans don't write like that, with "the skill removes" at the end of a sentence.
Maybe instead of using this, one should just have some pride and learn to write properly, instead of relying on a machine for everything.
I agree for your own writing. I use it mostly on text an LLM already wrote, like docs drafts and long answers, to make them faster to read.
thanks, good job
Let me ask it straight: WHY? Why, for heaven's sake? The only goal is to make AI detection harder, and that directly translates to more uneducated students. So why? What's the point? What's the ethos?
There are a lot of comments here every day about how difficult it is to read AI writing and how grating people find it. I think it would be a big improvement if that discussion could be drawn to a close.
Write better, read easier.
You can't SKILL.md your way to a soul.
Japanese Shintoism believes that all things have a soul.
Yes, there is no means of getting to imaginary things like souls.
Anthropic is still trying: https://soul.md/