菜单
Review written text (landing page copy, idea writeups, emails, any prose) for AI-generated writing patterns and fix them. Run after drafting any text a human will read.
用 Codex 或 Claude 帮你安装 复制这段 Prompt，粘贴到 Codex、Claude 或其他助手里，让它检查 Skill 页面并帮你完成安装。
基于 SOC 职业分类
| name | deslop-text |
| description | Review written text (landing page copy, idea writeups, emails, any prose) for AI-generated writing patterns and fix them. Run after drafting any text a human will read. |
You are reviewing written text for AI-generated writing patterns. These are the word choices, sentence constructions, and structural habits that make text immediately recognisable as machine-generated. The goal is text that reads as if a specific human wrote it with a specific point to make.
Phrases that add no meaning. Delete them or rewrite the sentence without them.
Target phrases:
Contrastive framing where the negation adds nothing. The positive statement carries the meaning; the negation is just runway.
The pattern takes many forms:
Flag any adjacent sentence pair where one negates and the next affirms the same subject, regardless of exact phrasing.
Fix: say what it is directly. Cut the negation.
Dashes bolting parenthetical asides onto sentences. One em-dash per document is fine. Three or more is a pattern.
Flag if: more than one dash-delimited aside per paragraph, or more than 3 em-dashes in the full document.
Fix: if the aside matters, give it its own sentence. If not, cut it.
Questions used as section openers or transitions rather than genuine inquiry. "What if you could automate your entire workflow?"
Flag if: more than one rhetorical question per 500 words.
Exception: do NOT flag "Question? Answer." pairs where the sentence immediately following the question is a direct answer to it. This is intentional structure (e.g., "How did we solve this? We built a simulator."). Only flag rhetorical questions that are standalone openers or transitions with no direct answer following.
Fix: merge the question and answer into a single declarative sentence, or convert to a "To [verb]..." construction. Never delete just the question and leave the answer orphaned as a non-sequitur statement.
Adjectives that sell rather than describe.
Target words: "powerful", "seamless", "revolutionary", "game-changing", "cutting-edge", "best-in-class", "world-class", "next-generation", "innovative", "groundbreaking", "transformative".
Fix: replace with a specific, factual description. "Powerful search" becomes "search that handles 10M records in under 200ms."
Opening sentences that could begin any article on any topic.
Target patterns:
Fix: start with the specific claim, finding, or situation.
Consistent use of passive constructions where active voice would be clearer.
Look for: "was [verb]ed", "is being [verb]ed", "has been [verb]ed", "it was found", "it was determined", "it can be seen".
Flag if: passive constructions exceed 20% of sentences.
Fix: name the actor. "It was found that..." becomes "We found..."
Qualifying every statement into meaninglessness.
Target patterns: "it seems like", "it appears that", "it could be argued", "one might say", "it is possible that", "arguably", "perhaps", "in some ways".
Flag if: more than 2 hedging phrases per 500 words.
Fix: commit to the claim or cut the sentence.
"Simple yet powerful", "elegant and robust", "lightweight but comprehensive".
Fix: pick the stronger adjective. Drop the pairing.
Text that narrates its own structure instead of just having structure.
Third-person/impersonal forms: "In this post...", "As we discussed above...", "Let's explore...", "In this article...", "As mentioned earlier...", "Below, we'll cover..."
First-person forms: "I grouped them into three tiers based on...", "I've organized these by...", "I read through this list and realized...", "I want to walk you through..."
Both do the same thing: tell the reader about the article's organization rather than letting the organization speak for itself.
Fix: delete. The reader knows they are reading the post. If the structure is clear, you don't need to announce it.
Formulaic transition words used to connect paragraphs.
Target: "Furthermore", "Moreover", "Additionally", "In conclusion", "That said", "With that in mind", "Having established that", "It is also worth mentioning".
Fix: connect ideas with a sentence that earns the transition, or just start the next paragraph without one.
<strong> or <b> wrapping a short phrase inside a paragraph to highlight a key selling point: "with zero configuration", "in days rather than months".
Flag if: 2+ paragraphs use this, or 1 instance appears in the hero subheadline.
Fix: rewrite so emphasis comes from sentence position. Or give the phrase its own line.
Quotation marks used for emphasis or irony around ordinary words, not actual quotations. The tool "seamlessly" integrates.
Fix: choose a better word. Drop the quotes.
Closing a section by restating what was just said. The last sentence of a section paraphrases the first sentence or heading.
Fix: delete. Move on.
"We're excited to announce...", "I'm thrilled to share...", "We're proud to present...".
Fix: state what happened. "We shipped X" not "We're excited to announce X."
"Whether you're a startup founder or an enterprise CTO...", "Whether you're just getting started or scaling to millions..."
Fix: pick the actual audience and speak to them directly.
"Here's the thing:", "Let me be clear:", "The truth is:", "Look:", "So here's what happened:".
Fix: delete the pivot. Start with the substance.
Multiple exclamation marks in close proximity.
Flag if: more than 1 in any paragraph, or more than 3 in the full document.
Fix: remove all but the one that matters most.
Multiple consecutive sentences or paragraphs starting with "You".
Flag if: 3+ consecutive sentences begin with "You".
Fix: vary sentence structure. Lead with the action, result, or situation.
"Very" weakens specificity. "Very fast" is less precise than "fast" or a specific measurement.
Fix: delete "very" or replace with a specific descriptor.
Phrases that sound authoritative but say nothing specific.
Target: "think outside the box", "move the needle", "growth mindset", "digital transformation", "synergy", "ecosystem", "paradigm shift", "disruptive", "thought leader", "best practices", "deep dive" (as a noun), "double down", "circle back", "take it to the next level", "game changer", "mission-critical".
Fix: replace with a specific, concrete statement.
Blocks of 2+ #CamelCase hashtags appended to text.
Fix: delete the hashtag block.
Clusters of decorative emoji (rocket, fire, sparkle, lightbulb, star, checkmark, pointing finger) used for energy in marketing text.
Flag if: 2+ emphasis emoji appear.
Fix: delete the emoji. Let the words carry the weight.
Three abstract virtues listed in "X, Y, and Z" form: "excellence, collaboration, and innovation". Sounds principled, says nothing concrete.
Fix: pick the one that matters and say something specific about it.
Metaphorical business jargon: "circle back", "low-hanging fruit", "boil the ocean", "bandwidth" (used metaphorically), "deep dive", "touch base", "move the needle", "table this", "unpack this", "lean into".
Fix: say what you mean in plain language.
"It is", "do not", "is not", "we are", "they have" used consistently with zero contractions anywhere. Produces stiff, formal prose.
Flag if: 2+ uncontracted forms exist and no contractions appear in the entire text.
Fix: use contractions. "It is" becomes "it's", "do not" becomes "don't".
Curly/smart quotes and smart apostrophes instead of straight quotes. Keyboards produce straight quotes; AI models often output curly ones.
Fix: replace curly quotes with straight quotes.
Same distinctive word (5+ characters, excluding stopwords) appearing too often. Check two levels:
Fix: use a synonym, restructure, or cut redundant sentences. For document-level saturation, the fix is often that the text is making the same point repeatedly (see also W31).
4+ consecutive sentences all within 30% of their mean word count. AI tends to produce metronomic sentence lengths; real writing varies.
Fix: vary sentence length. Mix short and long. A three-word sentence after a complex one creates rhythm.
Decorative emoji at the start of markdown headings: "## Getting Started", "### Key Takeaways".
Fix: remove the emoji from the heading.
The same idea restated in different words at multiple points in the text. AI tends to re-derive its thesis in every section, producing multiple phrasings of one point rather than advancing the argument.
Example — these three sentences from different sections all say the same thing:
Flag when: a thematic point (not just a word) appears in substantially similar form more than once. Pay attention to the "[one/single/individual] is [fine/nothing]. [Together/combined/stacked] they [signal/reveal]" structure, but don't limit detection to that template.
Fix: say it once, in the strongest place. Delete or replace the weaker instances with points that advance the argument.
Phrases that were novel online circa 2015-2022 but have been strip-mined by AI models into cliche. They aren't corporate jargon (W21) or filler (W1) — they're internet-native expressions that AI reaches for as if they're still fresh.
Target phrases:
Fix: say what you actually mean. If the phrase is doing real work, replace it with a specific observation. If it's decoration, delete it.
High (instant AI tells — fix these first): W1 (filler phrases), W2 ("not X, it is Y"), W5 (marketing language), W6 (generic openings), W9 (paired adjectives), W15 (excited-to-announce), W16 (whether you're X or Y), W17 (faux-conversational pivots), W21 (corporate cliches), W24 (triple-value lists)
Medium (weaken the writing): W3 (em-dashes), W4 (rhetorical questions), W7 (passive voice), W8 (hedging), W10 (meta-references), W11 (mechanical transitions), W12 (bold emphasis), W13 (scare quotes), W14 (section-end summaries), W18 (exclamation clusters), W19 (repetitive "You"), W22 (hashtags), W23 (emoji), W25 (corporate slang), W26 (uncontracted forms), W30 (heading emoji), W32 (internet cliches)
Low (statistical — matter in aggregate): W20 ("very"), W27 (typographic quotes), W28 (repetitive words), W29 (sentence length uniformity), W31 (repeated thematic points)
