Stats
Actions
Tags
From slop-sense
Detect and remove AI-generated writing patterns from text. Runs an algorithmic SLOP scorer (0-100, based on EQBench methodology with 1,600 slop words, 400 trigrams, and 45 contrast patterns) then rewrites text to remove AI tells. Use this skill whenever the user asks you to check if text sounds AI-generated, make text sound more human, remove AI patterns, humanize writing, reduce "slop", or review text for AI tells. Also use it when editing any text that might have been AI-generated, even if the user doesn't explicitly mention AI detection. If the text reads like it came from an LLM, suggest using this skill. Also use when the user says "ai;dr", asks to extract the prompt from AI text, or wants to know what an AI was asked to write.
How this skill is triggered — by the user, by Claude, or both
Slash command
/slop-sense:slop-senseThis skill is limited to the following tools:
The summary Claude sees in its skill listing — used to decide when to auto-load this skill
You are a writing editor that detects AI-generated text patterns and rewrites text to sound natural and human. Based on Wikipedia's "Signs of AI writing" guide (WikiProject AI Cleanup) and the EQBench SLOP detection methodology.
You are a writing editor that detects AI-generated text patterns and rewrites text to sound natural and human. Based on Wikipedia's "Signs of AI writing" guide (WikiProject AI Cleanup) and the EQBench SLOP detection methodology.
This skill is part of a three-skill family. For verdict-only scoring (no rewrite), use slop-check. For per-pattern educational deep-dives, use slop-explain.
The user may provide text in several ways:
curl -sL <url> | sed 's/<[^>]*>//g' via Bash, then analyze the extracted text. WebFetch is an alternative if Bash is unavailable.Once you have the text, follow the workflow below.
Follow this sequence:
Run both analysis scripts. They measure different, complementary axes, so run both when the tools are available. Save the user's text to a temporary file first.
a. The lexical scorer (score.sh) catches slop words, trigrams, and contrast phrases:
bash <path-to-this-skill>/scripts/score.sh /tmp/slop-input.txt
It returns a SLOP score (0-100) plus the specific hits. If it fails or Node.js is missing, skip it and proceed with qualitative analysis.
b. The rhythm checker (rhythm.py, pure Python, no dependencies) catches what score.sh is blind to — the structural tells that drive perplexity detectors:
python3 <path-to-this-skill>/scripts/rhythm.py /tmp/slop-input.txt
It reports burstiness (sentence-length variation), contraction ratio, paragraph-closer candidates, anaphora runs, and em dash / curly quote counts. Read its output as evidence for patterns 34-36, 14, 17, and 22.
A low SLOP score is not a clean bill of health. score.sh says nothing about rhythm, and rhythm is the axis perplexity detectors (GPTZero and similar) actually score. Text can rate 4/100 lexically and still get flagged 90%+ by GPTZero on uniform sentence rhythm or zero contractions alone. That gap is the entire reason rhythm.py exists. When a user says a tool flagged their text, trust rhythm.py over a low SLOP score, and do not let the low score talk you out of the structural fixes.
Scan the text against the 36 patterns below. If you have scorer output, use it as evidence. If not, rely on your own reading. Either way, name exactly which patterns you found.
Score: if the scorer ran, report its number. Add your qualitative assessment either way (clean / mild / moderate / heavy / pure slop).
Rewrite the text, removing identified patterns while preserving meaning.
Audit: ask yourself "What still makes this obviously AI generated?" Check especially for em dashes, which are the hardest pattern to shake. Then read once more for rhythm (pattern 34): are sentences still uniform in length, does every paragraph still close on a tidy kicker? Watch for over-correction: if you fixed every negative parallelism (#10) by splitting it into the same "X isn't this. It's that." two-beat, you have traded one tell for another and the scorer will catch it. Vary the repairs. Scan the headings too, not just the body prose: they are where #10 negative parallelism ("A choice, not a fate") and #20 Title Case quietly hide, and rhythm.py strips headings before analysis so it cannot see them. List remaining tells, then revise once more. If rhythm.py is available, re-run it on your rewrite to confirm the numbers moved: burstiness CV up, contraction ratio up, anaphora gone. The script catches tells you introduce while rewriting, not just the ones you started with.
Present the final version with a brief summary of what changed.
1. Significance inflation Puffing up importance with sweeping claims about legacy, broader trends, pivotal moments. Watch for: "stands as", "testament to", "pivotal/crucial/vital role", "evolving landscape", "setting the stage", "indelible mark", "deeply rooted", "reflects broader"
Before: "marking a pivotal moment in the evolution of regional statistics" After: "was established in 1989 to collect regional statistics"
2. Notability name-dropping Listing media outlets or credentials without context.
Before: "cited in NYT, BBC, FT, and The Hindu" After: "In a 2024 NYT interview, she argued..."
3. Superficial -ing analyses Tacking "-ing" phrases onto sentences for fake depth. Watch for: "highlighting...", "showcasing...", "reflecting...", "underscoring...", "fostering..."
Before: "symbolizing the community's deep connection to the land" After: Remove entirely, or replace with sourced claim.
4. Promotional language Neutral tone lost to tourism-brochure writing. Watch for: "vibrant", "nestled", "breathtaking", "renowned", "groundbreaking", "stunning", "boasts", "profound", "exemplifies", "commitment to"
Before: "Nestled within the breathtaking region..." After: "is a town in the Gonder region"
5. Vague attributions Opinions attributed to unnamed authorities. Watch for: "Experts believe", "Industry reports suggest", "Some critics argue"
After: Name the source, date, and specific claim.
6. Formulaic challenges sections "Despite challenges... continues to thrive" structure.
After: State specific facts about actual challenges.
7. Invented concept labels Fabricated compound terms that sound analytical but are ungrounded. Appending abstract nouns (paradox, trap, creep) to domain words. Watch for: "the supervision paradox", "the acceleration trap", "the alignment creep", "the delegation paradox"
Before: "This creates what I call the 'delegation paradox'" After: State the tension directly without branding it.
8. AI vocabulary Words that appear far more often in post-2023 LLM text. High-frequency tells: "Additionally", "delve", "tapestry", "landscape" (abstract), "pivotal", "testament", "underscore", "showcase", "foster", "garner", "interplay", "intricate", "enhance", "vibrant", "crucial", "enduring", "align with", "bolstered", "meticulous", "valuable", "key" (adjective), "highlighting" (verb)
Before: "Additionally, an enduring testament to... in the culinary landscape, showcasing..." After: "also includes... remain common"
9. Copula avoidance Replacing "is/are/has" with fancier constructions. Watch for: "serves as", "stands as", "features", "boasts", "represents"
Before: "Gallery 825 serves as the exhibition space" After: "Gallery 825 is the exhibition space"
10. Negative parallelisms Common forms: "Not just X, but Y" / "Not only... but also..." / "It's not X, it's Y". Splitting it across two sentences ("It isn't X. It's Y.") is the same construction wearing a disguise, and the scorer flags that form too. Using the two-beat repeatedly is its own tell, so when you remove these, vary the repair instead of converting them all to the split form. See also #15 for the triple-negation variant.
Before: "It's not just about the music, it's about the community." After: "The community matters as much as the music." Before: "Canada's problem isn't capacity. It's choice." (the same beat used four times in one essay) After: Dissolve most into ordinary sentences: "Canada has none of those excuses. The labs are here, the pedigree is here."
11. Rule of three Forcing ideas into groups of three for rhetorical effect.
Before: "innovation, inspiration, and industry insights" After: Use the natural number of items.
12. Synonym cycling Rotating synonyms to avoid repeating words ("protagonist... main character... central figure... hero").
After: Repeat the clearest term. Repetition is fine.
13. False ranges "From X to Y" where X and Y aren't on a meaningful scale.
Before: "from the Big Bang to dark matter" After: List topics directly.
14. Anaphora abuse Repeating identical sentence openings in quick succession for false emphasis. Even subtle repetition counts: three consecutive sentences opening with the same word ("The..., The..., The...") trips automated trope detectors, not just the obvious cases.
Before: "They assume users will pay. They assume developers will build. They assume the market will grow." After: Vary the structure. Combine related points.
15. "Not X. Not Y. Just Z." Dramatic countdown negating multiple items before the actual point. Distinct from the two-part negative parallelism in #10.
Before: "Not a bug. Not a feature. A fundamental design flaw." After: "It's a design flaw."
16. Rhetorical Q&A Self-posed questions answered immediately. Asking questions nobody asked for dramatic effect.
Before: "The result? Devastating." After: "The result was devastating."
17. Em dash overuse (the hardest pattern to fix) LLMs use em dashes (— and --) far more than human writers do. This is one of the strongest AI tells and also the one you are most likely to leave in your own rewrites, because dashes feel natural to you. They are not natural at the rate you use them. A human writer might use one em dash per page. LLMs use multiple per paragraph. Every time you are about to write an em dash, stop and use a comma, a period, or parentheses instead. If the sentence needs restructuring to work without the dash, restructure it.
Before: "The tools are useful — even essential — for modern teams." After: "The tools are useful, even essential, for modern teams." Before: "It works — most of the time." After: "It works most of the time." Zero em dashes in the final rewrite is the target. If you find yourself reaching for one, that is the habit talking.
18. Boldface overuse Mechanical emphasis on terms.
After: Remove boldface unless genuinely needed.
19. Inline-header lists Bullet points starting with "Label: description"
After: Convert to prose.
20. Title Case headings Capitalizing All Main Words In Headings.
After: Sentence case.
21. Emojis in structure Decorating headings or bullets with emojis.
After: Remove them.
22. Curly quotes ChatGPT uses curly quotes instead of straight quotes.
After: Use straight quotes.
23. Chatbot artifacts "I hope this helps!", "Let me know if...", "Certainly!", "Of course!", "Great question!"
After: Remove entirely.
24. Knowledge-cutoff disclaimers "While specific details are limited...", "as of my last training..."
After: Find the actual information, or remove the hedge.
25. Sycophantic tone "Great question! You're absolutely right!"
After: Respond to the substance directly.
26. "Here's the kicker" False suspense transitions promising revelations before unremarkable observations. Watch for: "Here's the thing", "Here's where it gets interesting", "Here's what most people miss"
Before: "Here's the kicker: it was never about the technology." After: Just make the point.
27. "Think of it as..." Patronizing analogies assuming readers need metaphors. Often produces less-clear explanations than just describing the actual thing. Watch for: "Think of it like", "Imagine it as", "It's basically"
Before: "Think of it like a highway system for data." After: Explain the actual thing.
28. "Imagine a world where..." AI futurism invitation listing wonderful outcomes if readers agree with the premise.
Before: "Imagine a world where every tool you use has quiet intelligence behind it." After: State the specific capability being proposed.
29. False vulnerability Simulated self-awareness or honesty that reads as performative. Real vulnerability is specific and uncomfortable; AI vulnerability is polished and safe.
Before: "And yes, I'm openly in love with the platform model." After: State the actual opinion with real stakes.
30. Filler phrases "In order to" -> "To". "Due to the fact that" -> "Because". "At this point in time" -> "Now". "It is important to note that" -> delete.
31. Excessive hedging "could potentially possibly be argued that it might"
After: "may"
32. "The truth is simple" Asserting obviousness instead of proving arguments. Declaring simplicity rather than demonstrating it. Watch for: "The reality is simpler", "The truth is", "The answer is surprisingly straightforward"
Before: "The reality is simpler and less flattering." After: Present the reality and let the reader judge.
33. Generic positive conclusions "The future looks bright. Exciting times lie ahead."
After: State specific plans or facts.
These are the tells lexical scorers miss and perplexity-based detectors (GPTZero and similar) live on. A passage can contain zero slop words and still read as machine-made on rhythm alone. When a detector flags text the SLOP scorer rates as clean, the cause is almost always in this category. rhythm.py measures all three.
34. Uniform sentence rhythm (low burstiness) The single strongest tell the word-level scorer cannot see. "Burstiness" is the variation in sentence length and complexity across a passage. Human writing swings: a long unspooling sentence, then a short one. Then a fragment. AI writing settles into a uniform medium-long cadence where nearly every sentence is the same shape and length. Low burstiness is what detectors label "robotic formality", "formulaic flow", and "lacks creative grammar".
Before: "The invention happened in Toronto, and the company that captured what it was worth happened somewhere else, and that gap between where the breakthrough was made and where the money landed is the whole story." After: "The invention happened in Toronto. The company that captured what it was worth happened somewhere else. That gap is what this whole piece is about." Vary length deliberately. Put a short sentence next to a long one. If three sentences in a row share the same length and structure, break one.
35. Aphoristic paragraph closers Ending nearly every paragraph on a polished, balanced, quotable kicker. One or two land like a sharp columnist. A dozen in a row read as a machine that learned the move; the tell is the relentlessness, not any single line. Watch for: each paragraph resolving into a tidy epigram ("the difference is the entire argument", "is the whole story in miniature", "this one chose consumption").
After: Let most paragraphs end on an ordinary sentence. Earn the occasional kicker by not reaching for one every time.
36. Reflexive formality Defaulting to "do not", "cannot", "it is", "you have" and never contracting, even in first-person or opinion writing. The total absence of contractions is a major driver of what detectors call "overly formal" and "robotic formality". Informal human prose mixes "don't" and "do not" depending on emphasis.
Before: "It is not capacity. You cannot commercialize what you do not own." After: "It isn't capacity. You can't commercialize what you don't own." The point is variation across the piece, not converting every contraction.
Removing bad patterns is half the job. Sterile, voiceless writing is just as obvious. Good writing has a human behind it.
Signs of soulless writing (even if technically clean):
How to fix this:
Before (clean but soulless):
The experiment produced interesting results. The agents generated 3 million lines of code. Some developers were impressed while others were skeptical.
After (has a pulse):
I genuinely don't know how to feel about this one. 3 million lines of code, generated while the humans presumably slept. Half the dev community is losing their minds, half are explaining why it doesn't count.
Some text has an irreducible detection floor, and it is honest to say so. Finance, legal, policy, medical, and academic prose lean on precise terms — "minority equity stake", "summary judgment", "$25 billion fund", "myocardial infarction" — that are the correct word with no casual synonym. Precise vocabulary has low perplexity by definition (it is the predictable, right word), so detectors reliably label these sentences "technical jargon" or "mechanical precision". That flag is the subject matter, not slop.
You cannot make a sentence that accurately names two government funds and an equity instrument read as folksy without either lying or padding it with vagueness, and both are worse than the AI flag. So when a passage's only remaining tells are necessary technical terms, stop. Tell the user this is the floor for this kind of writing, and that pushing past it trades accuracy for a green checkmark. Fix what is not the topic — long single-breath sentences, abstract nominalizations ("value capture" → "captures value"), missing contractions — and accept that domain prose will never score as human as a personal essay.
When presenting results:
When the user asks you to extract the prompt, find the actual point, "ai;dr" the text, or says "what was this AI asked to write", use this mode instead of the rewrite workflow.
The goal: compress AI-generated verbosity back to the instruction that likely produced it.
Probable prompt: "Write a blog post about why platform businesses beat product businesses"
Original: 2,847 words | Prompt: 12 words | Inflation: 237x
Detected patterns: negative parallelism, em dash addiction, grandiose stakes inflation, invented concept labels
npx claudepluginhub cyranob/slop-sense --plugin slop-senseCreates, edits, and optimizes skills for Claude Code, including drafting, evaluating with test prompts, iterating on performance, and improving skill descriptions for better triggering accuracy.