/unslop

name: unslop
description: >
  Removes signs of AI generation from English text and prevents them from
  appearing in new writing. Works on three levels: typography and mechanics,
  vocabulary and rhetoric, structure and epistemics. Use whenever the user asks
  to "humanize", "unslop", "de-AI", "make this sound human", "remove the AI
  style", "make it not sound generated". Also apply from the first draft when
  writing ANY English text for people: posts, articles, emails, reports, docs,
  landing pages, product copy. Do not wait to be asked. If the text is in
  English and a person will read it, this skill applies.

Unslop

Goal: text that reads like a thinking person wrote it, not like a model assembled it. The target is this author writing well, not the median human: average human writing is also slop, just a different flavor. Without a style profile the output is, honestly, the model's own house voice; say so once when it matters (see step 1).

The key fact this skill is built on. A 2026 study by the University of Maryland and Google DeepMind (61,608 texts) showed that when AI text is edited to remove clichés and surface artifacts, detection based on structural features barely drops: from 95.5% to 93.9%. What gives text away is not the word list but the way it is built: what the author chews over, what they name, where they allow unevenness. Swapping "delve" for "explore" fixes nothing. Vocabulary tells also decay on their own: "delve" collapsed in 2025, GPT-5.1 suppresses em dashes. Structure persists. So the skill works on three levels, cheapest first:

1. Typography and mechanics. 2. Vocabulary and rhetoric. 3. Structure and epistemics. This level is the one that matters.

How to work

1. **Check the author profile.** If `references/style-profile.md` exists, read it in full. Profile settings override the defaults in this file (except the Epistemics section, which no profile can switch off). No profile: use defaults. If no profile exists and the text will be published under the user's name, mention once that the default output carries the model's own house voice and offer calibration. Don't nag; once per conversation. 2. **Pick the mode.**

• *Rewriting*: the user gave you text. Keep the meaning, facts, and genre; remove the AI patterns.
• *Writing*: the text is new. Apply the rules from the first draft, not as a cleanup pass afterwards.

3. **Identify the genre and keep it.** A press release stays a press release, a post stays a post, an email stays an email. Humanizing means removing artificial rhetoric, not repainting everything as an analyst's memo. 4. **Pick the volume mode** (choose yourself if the user didn't, and say which if it matters):

• *Free edit* (default for "humanize this"): length may shrink noticeably; empty phrases get deleted, not reworded.
• *Careful edit* (the text must fill a slot: client document, fixed format): structure survives, length stays within 80–110%.
• *Minimal edit* (on explicit request): only unambiguous AI constructions are touched.

5. **Read `references/blacklist.md` and `references/prose-benchmarks.md`** before touching the text. The first holds the stop lists for level 2; the second holds the positive benchmarks of plain good prose (sentence-length variance, slack sentences, uneven confidence) used in the final check. 6. Run the three levels. Then the final check.

Level 1. Typography and mechanics

Defaults below marked as conventions can be overridden by the author profile.

• **Em dashes: avoid.** The single most famous AI tell, especially the spaced, formulaic kind used to punch up parallelisms. Restructure: comma, colon, parentheses, or two sentences. Hyphens stay in compound words. (Convention; a profile may allow sparing, unspaced use.)
• **Quotation marks: pick one style and hold it.** Mixed curly and straight marks in one text is a chatbot artifact. Default for web writing: straight. (Convention.)
• **Sentence case in headings.** Title Case On Every Main Word is an AI habit; capitalize the first word and proper nouns.
• **Bold**: not on every "key term". Bold only where the reader would get lost without it.
• **Lists**: do not convert connected prose into vertical lists with a bolded inline header and a colon in every item ("**Durability**: ..."). Fewer than four items, or items that logically flow: that is a paragraph.
• **Emoji** in headings and bullets: no, unless the genre demands it.
• **No "Conclusion" / "In summary" / "Overall" sections** that restate what was already said. If a long text needs an ending, the ending contains something new: a decision, a next step, a consequence.
• **Markdown artifacts and chatbot debris**: `---` breaks before headings, ```code fences around prose, `utm_source=chatgpt.com` in links, `turn0search0` and `oaicite` fragments, placeholder text like `[Your Name]`. Delete silently when rewriting.

Level 2. Vocabulary and rhetoric

18 categories. One example each here; full marker lists with fixes are in `references/blacklist.md`. A note on shelf life: the AI wordlist changes with each model generation, so treat lexical items as hints and the underlying habits as the actual target.

**1. Inflated significance.** "Stands as a testament", "plays a vital role", "underscores its importance", "left an indelible mark". Before: "The festival plays a key role in the city's cultural life." After: "The festival draws about 20,000 people every August."

**2. Promotional tone.** "Boasts a rich heritage", "vibrant", "nestled in the heart of", "renowned". Before: "A company with a rich history and unique expertise." After: "The company has been around since 2011; its main business is groupage freight."

**3. Negative parallelisms.** "Not just X, but Y", "not only X but also Y", "It's not X, it's Y", and the Grok-flavored "X rather than Y". Before: "This isn't just a redesign, it's a new product philosophy." After: "The redesign changed the navigation and removed two menu levels."

*