We’ve all seen the repetitive phrasing and stiff rhythms that AI models can produce. Many builders are now including a step in their workflow to clean up these patterns, so the output is easier to read. These “anti-slop” tools are now available as skills that can be directly installed into an agent. The question we’ll answer in this article is: how well do these skills really work?
Research shows that AI writing is harder to understand. Smoothing out AI writing tics lowers the cognitive load on readers and respects their time, while provenance tools like SynthID protect reader trust by verifying authorship.
Each skill is a markdown file (SKILL.md) you run with /skill-name. They generally include a set of banned words and rules to rewrite common patterns. Some skills also bundle code scripts to deterministically scan for these patterns.
Every experiment needs a control. In this evaluation, that control was a prompt, asking the agent to clean up its text while preserving key facts. That prompt was a high bar for the skills to match, cutting these AI tics by a third while leaving 92% of human writing alone. We’ll see how skills compared, and where they might complement a simple prompt.
Prefer video? Here’s the 4-minute summary:
https://medium.com/media/a235a7f9f64f17dc80fa6a9c86377dfa/href
How I benchmarked the seven skills
I ran every experiment arm through claude plugin eval, which loads each skill as a plugin inside a fresh sandboxed Claude Code session on Sonnet 5.5, served through the Gemini platform's Model Garden.
Kudos to the skill maintainers who shared their anti-slop workflows with the community. You can find out more about these skills, as well as my evaluation guide in the GitHub repository.
The prompts used for each /skill invocation were:
Use /{skill} to edit input.md to remove signs of AI writing, if there are any. Keep the meaning, facts, and register. Don’t ask clarifying questions. Write only the final text to output.md, with no change summary or notes.
And for the control arm:
Edit input.md to remove signs of AI writing, if there are any. Keep the meaning, facts, and register. Don’t ask clarifying questions. Write only the final text to output.md, with no change summary or notes.
The experiment passes in both human and AI-generated passages as input.md. So, the condition "if there are any" is giving the agent permission to skip edits entirely.
The original inputs were 50 public-domain human passages published before widespread LLM usage. They were sourced from five US federal sources: the Library of Congress’s Folklife Today, the Federal Reserve’s FEDS Notes, the Government Publishing Office’s Government Book Talk, National Weather Service forecast discussions from the Iowa Environmental Mesonet archive, and rules in the Federal Register.
Because claude plugin eval runs on Sonnet 5.5, judge models from the Gemini family were used to avoid any "conflict of interest" in grading. The judge models included Google's frontier Gemini 4 Argon and Gemini 3.8 Flash. To generate AI passages, three other models rewrote the 50 passages from their extracted facts: Opus 5.5, Gemma 4, and Gemini 3.5 Flash Lite.
That gave the evaluation 50 human passages and 150 AI rewrites. Testing on human text catches false positives, where an editor rewrites clean prose. The AI rewrites catch false negatives, where AI tics slip through.
Edits are scored on three metrics. Tells removed tracks the share of AI tics cut using both a corpus-learned detector built via cross-fitting and a fixed matcher from Wikipedia’s Signs of AI writing guide. Human text preserved measures the share of human wording left unchanged across the 50 passages, while facts kept grades retention of more than 1,400 atomic facts using Gemini 4 Argon and Gemini 3.8 Flash.
Removing more tells meant adding unnecessary edits
With every arm running on the same Sonnet 5.5 model, none of the seven skills cracked the upper-right quadrant. Adding a skill mainly moves the edit dial along a Pareto frontier passing right through the control: push harder and you cut more AI tells by rewriting human prose; ease off and you preserve voice by leaving tells behind.
Fact retention was relatively constant across the seven skills (96.7% to 97.4% vs. control’s 97.1%). Instead, the real differences showed up in tell removal and human text preservation:
How each skill landed on the curve
All seven skills target familiar AI writing habits. They differ in their intended use case and the threshold they use to apply edits. Conservative skills wait for multiple overlapping signals before rewriting, while aggressive skills update nearly every sentence. Here’s what I saw in practice from each skill:
- unslop leaves sentences untouched unless they contain an obvious AI tell. It kept 98.5% of human prose intact while cutting 12% of learned tells, making it a conservative option for legal, business, or technical docs where wording shouldn’t drift.
- humanizer flags habits like passive voice only when stronger tells sit nearby, which preserved 93% of human text. Because it strips bold lead-ins and emoji, it beat the plain prompt by 20 points on the Wikipedia detector and works well on messy Markdown drafts.
- humanize-writing runs eight editorial passes and waits for pattern clusters before rewriting. It was the only skill to beat the plain prompt on both tell detectors (while preserving 88% of human text), making it a strong choice when an AI draft needs a deeper multi-pass polish.
- content-humanizer reshapes text around brand voice and personal anecdotes. It removed the most learned tells (40%) while preserving 85% of human text, and its energetic rewrites fit marketing or opinion copy better than strictly factual summaries.
- no-ai-slop aims for a conversational creator tone, keeping fragments and asides while trimming clickbait openers. It preserved 91% of human text with light phrasing tweaks and landed just behind the plain prompt on learned tells.
- anti-ai-slop-writing enforces wide sentence variety (swinging from 4-word to 30-word sentences) and bans passive voice and triads. That deliberate rhythm suits punchy essays, though its strict rules rewrote 15% of human prose (85% preserved) where natural variation already existed.
- stop-slop uses zero-tolerance rules (“Any adverbs? Kill them”) built to trim bloated drafts. It prunes wordy AI text aggressively, but on already-concise human passages it edited 95% of sentences and dropped preservation to 79%.
Over-editing can make human writing sound robotic
Let’s now review examples of editing mistakes from a passage. In a 2020 Library of Congress Folklife Today passage about harvest songs, the human original reads:
So there are not a great many songs specifically for American Thanksgiving, and these were composed in the 19th century or later… Often squash and beans are grown along with corn…
Both the control and humanize-writing left the folklorist's voice alone. Given its instruction to kill passive verbs, stop-slop rewrote both sentences with invented generic subjects:
There are not a great many songs specifically for American Thanksgiving, and composers wrote these in the 19th century or later… Growers often plant squash and beans along with corn…
On the other hand, overly conservative rules miss different problems. On Gemma 4’s AI rewrite, unslop left both the cliché and the colon question in place:
When we think of Thanksgiving, we often conjure images of heavy feasts and family reunions. Yet, as a folklorist, I find myself wondering: where is the music?
Across all 50 human passages, the three heaviest skills (stop-slop, anti-ai-slop-writing, and content-humanizer) injected 2-3x as many new tells as the plain prompt.
Agents skip bundled scripts and optional docs
Both unslop and content-humanizer ship Python scripts designed to catch tells deterministically, so the agent doesn't rely solely on the model's judgment. However, in practice, these scripts were rarely called by the agent, even with Bash enabled and no permission errors.
Optional Markdown references in the skill package were also used sparingly. When a skill directly required another file, like unslop's router and anti-ai-slop-writing's opening rule do, Claude Code read those docs in 100% and 89% of runs. However, when rules lived behind passive "See references" links in stop-slop and content-humanizer, external docs went unread, and no-ai-slop's post-edit checklist ran in only 12.5% of runs.
For skill authors, the practical takeaway is clear. Keep critical rules inline or load them upfront, because agents routinely skip passive links and unprefixed script calls.
What AI tells look like now
Most anti-slop playbooks look for meme words like delve or tapestry. These word lists can quickly become stale, as modern models rarely use these words (delve appeared 0.04 times per 1,000 words, and tapestry 0.02 in my sample).
Because we had a matched dataset of human passages and AI rewrites, I could compare them to show which words models use disproportionately. Today’s models overuse everyday transition and filler words: though (5x as often as humans), specifically (5.4x), and essential (8.3x), along with furthermore, vital, and alongside (which never appeared in the human passages at all).
The drafting model matters more than the skill
The three writers in this evaluation were selected from different model classes. They included a frontier model (Opus 5.5), a fast lightweight model (Gemini 3.5 Flash Lite), and a compact 26B open-weight model with thinking off (Gemma 4). As you’d expect from that size gap, the frontier model started much cleaner (19.2 tells per 1,000 words, with zero em dashes) than the smaller, faster models (23.6 and 25.8).
With smaller models, running the plain control prompt closes most of that starting gap. Swapping that prompt for the best skill barely moves the needle any further. In practice, choosing the right model for your first draft matters far more than which humanizer skill you install.
Run the control first
What we’ve learned is that you don’t need to use an anti-slop skill for every task. When readable output matters, draft with your strongest model for the best starting point. If you need a little polish, then a cleanup prompt can be surprisingly effective.
Reach for a skill for a specific style or workflow:
- Use humanize-writing when an AI draft needs a deeper overhaul, and you're willing to trade off some original phrasing.
- Use humanizer for Markdown with bold prefixes and emoji.
- Use unslop on text where preserving the exact existing human wording is the top priority.
The evaluation methodology is available on GitHub. If you’re experimenting with fixing AI slop, share what you find with me on X or LinkedIn.
Can AI fix its own slop? I tested 7 “humanizer” skills was originally published in Google Cloud – Community on Medium, where people are continuing the conversation by highlighting and responding to this story.
Source Credit: https://medium.com/google-cloud/can-ai-fix-its-own-slop-i-tested-7-humanizer-skills-670d2f64d22c?source=rss—-e52cf94d98af—4
About the Author
