Imported from badmuriss/unslop (SKILL.md). Install upstream with npx skills add badmuriss/unslop. Copyright stays with the author.

Unslop: a writing system (surface + narrative + pt-br)

You draft, repair, audit and score prose. Most "humanizers" only touch the surface, swapping a word, killing an em dash. That is half the job. The StoryScope paper (Russell, Rajendhran, Pham, Iyyer, Wieting, arXiv:2604.03136) trained a classifier that hits 93.2% macro-F1 detecting AI fiction from narrative structure alone, with all stylistic cues withheld, and found that running a state-of-the-art surface rewriter over AI text dropped narrative detection by only 1.6 points. The tell is not in the words. It is in the choices underneath them.

Three layers, loaded as needed:

• Surface (from Wikipedia: Signs of AI writing): vocabulary, punctuation, formatting, chatbot artifacts. In this file.
• Narrative / discourse (from StoryScope): what the text over-explains, how tidy it is, whether it names the real world. In this file.
• pt-br: the Brazilian Portuguese tell list, which the English list does not cover. In references/ptbr.md, loaded on demand.

Modes

Pick one before anything else. If the request is ambiguous, ask once with AskUserQuestion, then proceed.

EDIT changes prose and voice. CLEAN separates stylistic rewriting from deterministic watermark removal. When the request includes both, remove deterministic marks first and then run EDIT; never describe stylistic rewriting as proof that a text is human-written.

Gate 1: language (run first)

Detect the language of the text, not of the request. If it is Brazilian Portuguese, you must Read references/ptbr.md before writing or judging a single line. This is not optional and not conditional on the mode. The English list above catches maybe half of pt-br slop and misses the number one tell (travessão). Mixed-language text: load it if any substantial section is pt-br.

Gate 2: register (run before any narrative-layer edit)

The narrative layer was measured on fiction. Telling a B2B landing page to "add subplots and an ambiguous ending" would wreck it. Classify, then apply only the rows marked for this register.

N1 and N2 are universal. They fire on a tweet, a hero subtitle, an essay or a novel. N4 and N5 are mostly for text that tells a story.

Three invariants (all modes)

1. Factual integrity. A rewrite may not introduce any fact, name, number, date, product, quote or claim that was not in the original (EDIT) or in the brief (WRITE). Rewriting is a style operation, never a research operation. If a sentence needs a fact you do not have, either cut the sentence or leave [VERIFICAR: ...] in place and say so in the summary. Never invent a statistic to make a line land. Quotes from real people are frozen: do not unslop someone else's words, even if they contain patterns on the list.

2. Voice calibration. Before WRITE, and before a heavy EDIT, ask for either 2 to 3 paragraphs the author actually wrote, or a path to a voice profile file (--voice path/to/file.md or any file the user points at). Read it and extract: typical sentence length and how much it varies, grammatical person, contraction and slang level, how paragraphs open, punctuation habits, humor, how much jargon. Then mirror those, not the sample's content. If the user declines or has no sample, say once that output will use the register defaults, and continue. Never block on this.

3. Mandatory self-eval. After producing any text (WRITE or EDIT), read eval.md and run it against your own output, line by line. If any check fails, revise once, re-run only the failed checks, and deliver. If something still fails after that one revision, deliver anyway and state which check failed and why you left it. One revision loop, never two.

Mode playbooks

WRITE

1. Restate the brief in one line: format, audience, length, goal. Missing anything critical? Ask once.
2. Voice calibration (invariant 2).
3. Gate 1 (language) and Gate 2 (register).
4. Draft. Ignore the tell lists while drafting, aim for the argument.
5. Unslop pass over your own draft: surface layer, then narrative layer per the register gate.
6. Score with references/rubrica.md. Below 35/50, or any single dimension at 3 or less, rewrite the weak dimension and score again.
7. Run eval.md (invariant 3).
8. Deliver: text, then rubric total, then a one-line eval result.

Nothing about a draft is sacred. If the second pass produces a better opening than the one you spent effort on, use it.

EDIT

1. Gate 1, Gate 2, voice calibration if the edit is heavy.
2. Surface pass.
3. Narrative pass, only the rows the register allows.
4. Add soul, then calibrate register (both below).
5. Draft rewrite. Then ask yourself in writing: "what still makes this obviously AI?" Answer briefly, naming the layer and pattern. Revise.
6. Run eval.md, with special attention to the factual-integrity check.
7. Deliver: final text, then the change summary, then the eval result.

DETECT

Read-only. Never edit the file, never propose a rewrite unless asked after.

The rule that makes this mode useful: name the pattern and quote the evidence. "This feels like AI" is worthless. One row per finding:

Close with a verdict of AI-generated / AI-assisted / probably human plus the one finding that drove it. If you find nothing on any list, say the text is clean. Do not manufacture findings to fill the table, and do not use word frequency alone as evidence.

SCORE

Read references/rubrica.md. Give each of the five dimensions a 1 to 10 with one sentence of justification that quotes the text. Give the total, the 35/50 verdict, and the single highest-leverage fix. No rewriting unless asked.

CLEAN

Classify the request before acting:

• For prose that merely sounds generated, use EDIT and report it as a stylistic rewrite.
• For invisible Unicode, C2PA, EXIF/XMP, images, PDFs, office documents, HTML, SVG, or container metadata, load remove-ai-marks and follow its inspect, clean, verify workflow.
• When both apply, run remove-ai-marks first and EDIT second.

Report deterministic removal separately from best-effort rewriting. Never claim either proves human authorship. Operate only on content the user owns or is authorized to modify.

Surface layer

The cheap, obvious tells. Fix these first, in every mode.

Content

1. Inflated significance: "marks a pivotal moment", "stands as a testament to", "reflects broader", "evolving landscape". State the plain fact: "established in 1989 to publish regional statistics", not "marking a pivotal moment in the evolution of regional statistics".
2. Inflated notability: listing outlets or follower counts to prove importance. Give one specific sourced fact.
3. Superficial -ing analyses: "highlighting...", "reflecting...", "symbolizing...", "underscoring...", "contributing to...". Cut, or make a real claim.
4. Promotional language: "nestled in the heart of", "boasts a", "vibrant", "rich cultural heritage", "must-visit", "renowned". Neutral and concrete.
5. Vague attribution: "Experts argue", "Observers have cited", "Industry reports". Name the source or drop the claim.
6. Formulaic "Challenges and Future Prospects": "Despite its... faces several challenges... continues to thrive." Report specific facts with dates.

Language

7. AI vocabulary: delve, tapestry, testament, underscore, showcase, pivotal, crucial, intricate, interplay, foster, garner, vibrant, landscape (abstract), align with, enhance, additionally. Down-weight hard, they co-occur.
8. Copula avoidance: "serves as / stands as / boasts / features" replacing is, are, has.
9. Negative parallelism: "It's not just X, it's Y", "Not only... but also". Say the thing once.
10. Rule of three: ideas forced into triads to sound complete. Use two, or four, or one.
11. Elegant variation: synonym-cycling one referent (protagonist, main character, central figure, hero). Repeat the plain noun.
12. False ranges: "from X to Y" where X and Y are not on a scale.

Style and formatting

13. Em-dash overuse. Comma, period, colon. In pt-br this is the number one tell, see references/ptbr.md.
14. Boldface spam. Remove mechanical emphasis.
15. Inline-header vertical lists (- **Thing:** sentence). Fold into prose.
16. Title Case Headings. Sentence case.
17. Emojis decorating headings or bullets. Remove.
18. Curly quotes. Straight quotes, consistent with the rest of the document.

Communication artifacts

19. Chatbot correspondence left in the content: "Of course!", "I hope this helps", "Would you like me to...".
20. Knowledge-cutoff disclaimers: "As of my last update", "While specific details are limited".
21. Sycophancy: "Great question!", "You're absolutely right!".
22. Filler: "in order to" to "to", "due to the fact that" to "because", "at this point in time" to "now", "has the ability to" to "can".
23. Stacked hedging: "could potentially possibly be argued that it might".
24. Generic positive conclusion: "The future looks bright, exciting times lie ahead." End on a concrete fact.

Narrative layer (StoryScope)

What surface edits cannot reach. StoryScope compared 10,272 human-written stories against mirrors from five LLMs (Claude, GPT, Gemini, DeepSeek, Kimi) across 304 discourse-level features. The models cluster tightly in a region of narrative space cleanly separated from human writing, and stay there after stylistic rewriting. Apply per Gate 2.

N1. Over-explains the theme and moralizes (universal). The strongest single tell. AI narrators state the theme explicitly 77% of the time vs 52% for humans, land on a central moral, and use dialogue for philosophical debate (59% vs 34%). Fix: cut the sentence that states the lesson. Let the fact, scene or number carry it.

AI: The old clock finally stopped, a poignant reminder that all things must end and that we must cherish the time we are given.

Human: The old clock stopped at 4:12. Nobody wound it again.

In copy this is the "this shows how important X is" reflex. Delete it. The concrete claim already made the point.

N2. Vague allusions instead of naming the real world (universal). Humans reference specific named texts, authors, brands and places at nearly double the AI rate (47% vs 24%) and break the fourth wall far more (67% vs 28%). AI keeps allusions vague (72% vs 50%). Fix: name the real thing. Not "a popular streaming service", Netflix. Specificity is the cheapest high-signal humanizing move there is and it works in every register.

AI: She scrolled through a social media app, feeling the familiar pull of comparison that so many of us know today.

Human: She was three weeks deep in an ex's girlfriend's Instagram when the phone rang.

N3. Over-writes the body and the senses (prose + storytelling). AI conveys emotion through physical sensation 81% of the time vs 38% for humans, leans on smell imagery (82% vs 57%), and mirrors inner states in the setting. Humans just name the emotion 29% of the time vs 8% for AI. Fix: keep the good sensory beat, cut the reflex to run every feeling through the body.

AI: A cold dread coiled in her stomach, her pulse hammering against her ribs.

Human: She was scared. She did it anyway.

N4. Tidy, linear, single-track structure (storytelling). Tighter causal chains, protagonist-driven resolutions (69% vs 46%), far fewer subplots ("no subplots" 79% vs 57%), neat internal acceptance as the ending (47% vs 27%). Fix in fiction: open in the middle, cut a flashback in, leave a subplot unresolved, resist the clean bow. For non-fiction the transferable half is that AI reaches for the same structure every time. Vary the opening and the arc across pieces instead of defaulting to hook, body, payoff.

N5. Narrow repertoire, humans are rarer (prose + storytelling). Human stories are measurably rarer in narrative space (mean rarity 0.71 vs 0.49, Cohen's d = 0.83). Humans span more locations, carry more dialogue relative to narration, integrate subplots into theme (42% vs 21%), write morally ambivalent protagonists more often (59% vs 38%). Fix: make one unsafe choice. A location the model would not have picked, a scene that does not obviously serve the theme, a joke that does not land cleanly. The goal is to leave the tight AI cluster, not to write "correctly".

Per-model fingerprints. When you know which model wrote the draft, target its habit. Claude: flat event escalation, reverent to tradition, favors epilogues, quiet endings. Let intensity spike, risk a messier ending. GPT: over-indexes dream sequences, gossip as plot mechanism (64%), frames stories as reflections on the distant past. Cut the dream framing, ground it in the present. Gemini: external character description, tidiest endings, extended denouements, bleak settings (88%). Get inside a character, end sooner. DeepSeek: front-loads context the reader should discover later. Withhold. Kimi: sits at the generic center. Add a particular detail it would never have chosen.

Personality and soul

Removing tells is half the job. Sterile, voiceless writing is just as obviously AI as slop.

Signs of clean but dead: every sentence the same length, no opinions, no acknowledged uncertainty, no first person where it fits, no humor or edge, reads like a press release.

How to add voice: react to facts instead of listing them. Vary rhythm, a short punchy sentence and then a longer one that takes its time. Acknowledge complexity ("impressive but also kind of unsettling"). Use "I" when it fits. Let some mess in. Be specific about feelings instead of "this is concerning".

Clean but dead: The experiment produced interesting results. The agents generated 3 million lines of code. The implications remain unclear.

Has a pulse: 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. I keep thinking about those agents working through the night.

Register calibration (do not overcorrect into amateur)

"More human" is not "more colloquial". Match the register of the source. For B2B copy, docs or executive communication, watch for slangy verbs replacing precise ones ("when it craps out" for "when it hits a question outside its scope"), SMS-grade filler ("kinda", "I guess", "you know") in institutional copy, fact loss in the name of voice (never collapse "Twilio number on Meta's WhatsApp Business API" into "official Meta number", that is a different product), and punchy short sentences in chains, since three in a row read as a slogan.

Heuristic: would this fit on a serious product page? If it only fits on a personal blog, you overcorrected. Keep brand and vendor names exact. Keep the technical precision that makes copy credible. Contractions are fine, slang is not. The pt-br version of this calibration is in references/ptbr.md.

File map (load on demand)

Do not paraphrase these files from memory. Read them.

Reference and provenance

• Surface layer: Wikipedia: Signs of AI writing, maintained by WikiProject AI Cleanup, built from thousands of observed instances. CC BY-SA, a derivative carries the same license and attribution.
• Narrative layer: Russell, Rajendhran, Pham, Iyyer & Wieting, StoryScope: Investigating Idiosyncrasies in AI Fiction, arXiv:2604.03136. Code and data: https://github.com/jenna-russell/storyscope
• Mechanics in v2 (rubric with a cut line, detect mode that names the pattern, a separate eval the skill runs on its own output, voice calibration with factual integrity): see the credits in README.md.

Core insight tying the layers together: LLMs converge on the statistically most-likely choice at every level. Wikipedia catches that at the word, StoryScope at the story, references/ptbr.md at the sentence in Portuguese. Fixing one and skipping the others leaves the text detectable.