Human after all.
It's become a bit annoying and it is everywhere. Your LinkedIn post, every Reel's script, blogposts, news articles, even music lyrics!
Everything reads and feels the same nowadays. It's the silent, corrosive fingerprint that AI is leaving behind. Before the LLM boom, we could say that every piece of content on the Internet had its own voice. Even if it was crappy content, it was genuine.
Writing, drawing, and even thinking require practice and creativity, and it's been this variability and plethora of human science and art that has allowed us to train frontier models. By itself, that alone is a good reason to keep feeding our collective human knowledge to make AI and, more importantly, ourselves better.
I think we'll eventually find our way to AGI or super-intelligence and, one day, even agree on what those two things are in the first place. I refuse, though, to accept a world where we surrender to the monopoly of AI thought and human stupidity. I feel a bit guilty here for how much I enjoy those afternoons running exciting vibecoding sessions with my AI friends. Being able to swiftly create working stuff is fascinating. But I do think it makes us a bit dumber not taking some time to exercise our brains.
When was the last time you wrote a letter from scratch? An email? A thank-you or love letter?
I've always seen coding, maths, and writing as fulfilling and nourishing disciplines for the mind. That's why I still write most of my blog posts from scratch.
Psychologist Paul Bloom believes humans are essentialists. Meaning, we not only find value in things for what they are, but also for where they come from or from how they came to be. Bloom gives examples such as Britney Spears' chewing gum that has reached prices of up to USD 14K at auction. Here, it's not really the intrinsic value of the chewing gum that people are willing to pay for, but where it comes from (It's Britney, B****!).
I see people (now and in the future) giving more value to human-made creativity over AI-made paintings, books and songs. As for coding, I think we need to give ourselves some respect here and allow ourselves to still understand what machines are doing. Programming languages should be written and documented by humans for human understanding (let's see how well that statement ages...). Sure, we can fine-tune AI models to follow and excel at them, but a world where all the backbone of data, data processing and data manipulation is solely understood by machines is, at the very least, frightening.
In summary: write yourself that love letter, make that essay your own. Put a bit of yourself into whatever you do. I'm not telling you to stop using AI (that would be plain dumb), but to keep your brain in good shape.
However, I know that nowadays we do need to use AI to write stuff. In this blog post I'll show you the most common AI slop signals in text and how to avoid them with a simple skill. With Claude's help, I've developed a humanizer skill that is 12 times shorter than the most popular one, and I put both to the test.
AI slop in text
Detecting AI slop is an ability that I've unproudly developed. AI tells are a set of recurring themes, exaggerations, formatting, and artifacts that LLMs leave in almost anything you ask them to do. These artifacts reduce, in general, the quality of your text.
When life gives you lemons, make lemonade 🍋, as the saying goes.
Let's see how AI would fail miserably to say the same by introducing 8 of the most common AI slop signals in text:
Here's the thing: when life gives you lemons, it's not a setback — it's an invitation. Don't get me wrong, sour days are real, and nobody is asking you to enjoy them. But at its core, making lemonade is about perspective, patience, and purpose. It's a chance to delve into the rich tapestry of who you are when things don't go to plan — a crucial reminder that growth rarely arrives on schedule. So the next time life hands you a “problem,” don't see an obstacle. See an ingredient. Because the squeeze is the point.
It doesn't matter which model you choose. Due to the nature of the text corpus used during LLM pre-training as well as decisions made during fine-tuning, these tells are fused into AI guts and that's why they aren't that easy to avoid. The sad story here is that the more AI text drowns the internet, the more common these AI tells will be.
How to humanize AI text?
This has already been answered before. Maybe the most straightforward way to humanize AI-written text is using Blader's Humanizer skill (BHS from now on). You install it in Claude Code by running:
/plugin marketplace add blader/humanizer
/plugin install humanizer@humanizerand then just invoke it by asking Claude (or your AI of choice) to humanize the following crappy text. This skill works very well and you'll end up with fewer AI-sloppy texts. Now, if you have some time to write the text yourself, please do yourself and humanity a favor and do so.
Can this be simpler? Introducing the Short Humanizer Skill (SHS)
I've lately been playing with my Hermes agent; her name is Asuka. She's become my personal and family assistant. During our initial interactions, the infamous fingerprint was there: AI tells in all her texts, advice, chats, everything. To me it was a bit of overkill to inject more than 8K tokens in every conversation by invoking BHS.
I wondered if I could achieve similar results with a shorter skill or just a smaller prompt. The good news is that this is possible by using a shorter prompt that focuses on which tells to avoid and not elaborating that much on the examples.
So I created (with Claude's help, clearly) the Short Humanizer skill (SHS from now on). Use the short prompt below to fix the most common AI tells / AI slop in text. It uses 12 times fewer tokens than BHS. This makes it perfect for integrating with your OpenClaw or Hermes agent, since you can directly inject this prompt into your agent's SOUL.md and make your conversations a bit more... human... if we could put it that way:
# Short Humanizer
Rewrite the text so it reads like a careful human writer. Keep every claim, opinion and detail. Do not add facts, names, numbers or quotes.
Remove these habits of AI prose:
1. Contrast used as emphasis: denying a claim nobody made so the real point sounds bigger. It hides in several shapes: a denial and an affirmation inside one sentence; the reversed form that names the preferred option and then the rejected one; a denial sentence followed by an affirming one; a clipped negative tail after a comma; a closing line that sets the point against a weaker version of itself. Check every paragraph for these. State the point directly. Keep a contrast only when it corrects something the reader actually believes, or when both halves carry information.
2. Closing lines: the last sentence of a paragraph or a reply that sums up, draws a moral or repeats what came before; rows of dramatic fragments. End on the last concrete fact.
3. Ordinary points dressed as deep truths or aphorisms.
4. Run-ups that announce the point instead of making it.
5. Rebuttals of objections no one raised.
6. Forced triads: three words, phrases, examples or clauses in parallel when the idea needs one or two. Check every list of three and keep it only when each item adds something the others do not.
7. Several sentences in a row that open the same way.
8. Em and en dashes. Use commas, periods, colons or parentheses.
9. Stacked hedges.
10. Hyphenated compounds after the noun.
11. Passive voice that hides who acts.
12. Stock AI vocabulary.
13. Ordinary facts framed as turning points or legacies; upbeat send-offs.
14. Vague links where the relationship could be named.
15. Participle tails tacked on to add significance.
16. Promotional, brochure adjectives.
17. Unnamed experts or lists of outlets used as authority.
18. A grander verb or verb phrase where is, are or has would do. Use the simple verb.
19. Bold in prose, including bold labels at the start of list items. Use bold only when the format calls for it.
20. Title Case or emoji headings.
21. Curly quotes. Use straight ones.
22. Chatbot greetings, offers and sign-offs.
23. Knowledge-cutoff disclaimers.
24. A section whose first sentence restates its heading. Open with the first new fact.
25. Text describing itself instead of its subject.
26. In a conversation, restating what the user said or what was already covered before answering. Start with the answer.
Leave quotations, proper names, genuine asides and odd specific details alone. Vary sentence length. The result should still sound like a person with opinions.
Does it really work?
I decided to put this to the test in the most scientific way I could. To test whether the SHS yielded at least similar results to BHS, I compared both on the following tasks:
A writing challenge consisting of rewriting three base texts generated with ChatGPT:
A blog post on why the world is less colorful nowadays.
An opinion article explaining why Mufasa should not have died in The Lion King.
A guide on how to start a vegetable garden.
A chatting challenge where we prompted an agent to avoid the aforementioned AI tells during chat.
For the writing challenge, we first (when I say we here, I refer to my friend Claude and me) generated these base texts using GPT-6.1 Sol at medium reasoning, with no skill. Then, we used the same model to rewrite each base text 12 times with each skill (BHS and SHS). That's a total of 72 rewrites of the same three texts. This was done to measure with some level of statistical significance how much AI tells diminished when you applied each skill, and not concluding that one is better than the other by pure luck. AI errors were measured by counting the number of AI tells every 100 words.
As for the chatting challenge, we need to recognize that this is a different kind of task because here we're not necessarily asking the model to rewrite something but to avoid AI tells. In this case we simulated 20 chat conversations with each skill and also counted the number of tells per 100 words.
AI tells were measured both deterministically and with Claude agents. Things like formatting, em dashes, etc. can easily be spotted with a Python script. When it came to more sophisticated blunders, such as "It is not X, but Y", we used 8 Claude sub-agents (the judges) to detect these errors in each generated or rewritten text plus another 8 sub-agents that audited the job of the judges. This audit was worth doing since the original judges tended to over-count AI tells, generating a lot of false positives.
The results
Both skills are good and useful. I wouldn't say one is significantly better than the other since they apply to different use cases. The SHS was slightly better at the chatting challenge since it yielded 0.37 fewer tells per 100 words than BHS (p < 0.001). This is mainly because the short skill was better at avoiding formatting tells (particularly bold text). If it wasn't for that, there wouldn't be a significant difference between the two.
Regarding the writing challenge, both skills performed very similarly. BHS was better at rewriting the blogpost, but on the other two writing assignments it was a tie.
Conclusion
Using any skill is way more important than the actual skill or prompt you use. I'd bet you would achieve good results by just asking your AI to remove AI slop (most likely it'll look them up and apply them).
BHS is still a very good skill because it's detailed. This is particularly important for low-end LLMs or models reasoning at low levels.
The Short Humanizer skill, being 12 times smaller, would be particularly useful for agent usage, where API token usage and context windows matter. It's also useful if you want to tweak it and make it your own, as a quick copy-paste prompt you can use on any AI tool.
Again, if you have the time, please write whatever you want to write. It'll end up having that Human after all touch, people will notice, and it'll make content out there better (btw, I came up with that forced triad on my own).
Github Repo: https://github.com/thecodebeatz/short-humanizer
Peace 🍋 🤖