agentleFS
risukisurisukisu/marketing-skills/skills/Writing/copy-deslop/SKILL.md

copy-deslop

Skill6 starsChanged 49 days agoMITView on GitHub

Clean up AI-sounding copy. Use when text feels generic, robotic, or obviously LLM-generated. Catches overused words, hollow phrases, false enthusiasm, and structural patterns that signal AI writing. Fetches the latest tells from Wikipedia entry on "Signs of AI writing" before each analysis. Works on any text — marketing copy, blog posts, emails, docs.

The file

We split this skill into the rules it gives an agent

5 rules. About 0 are common, 5 are distinctive, and 0 are very specific to this project.

  • ≈0

    Common · estimate

    Many files like this have a rule of this kind.

  • ≈5

    Distinctive · estimate

    This author's own choices. Worth a look.

  • 0

    Very specific

    Tied to their project's names, files or tools. Probably skip.

  • Common
  • Distinctive
  • Very specific
  1. Distinctive:

    Every paragraph starts with a clear topic sentence, then elaborates.

    line 60
  2. Distinctive:

    Claude Code (CLI): use defuddle for clean markdown output:

    line 108
  3. Distinctive:

    Claude Windows App / web / any environment: use the Wikipedia REST API directly — it's designed for programmatic access and won't block:

    line 113
  4. Distinctive:

    If both fetch attempts fail, proceed with the hardcoded rules and note: "Live Wikipedia check unavailable — using built-in rules only."

    line 121
  5. Distinctive:

    When reporting, categorize each flag:

    line 139

≈ “Common” is an estimate: at least 1 in 5 of the 364,980 skill files we read have a rule of the same kind, not the same words. “Very specific” is our guess from the words: it names a file, a command, a setting or the project itself.

Keep what fits and drop the rest. Your version, ready for your agents.
Files
SKILL.md

Copy De-Slop

Detect and fix AI-sounding writing. This skill exists because LLM-generated text has recognizable patterns — and readers have learned to spot them. If your copy triggers even one "this was written by AI" alarm, trust drops instantly.

How to use: Paste text, or point to a file. The skill analyzes it, flags problems, and rewrites the flagged sections.

What to Flag

Tier 1: Dead Giveaway Words

These words appear in AI writing at 10-100x their normal human frequency. Replace or remove every instance:

Word/PhraseWhy it's a flagFix
delve, delves, delvingExtremely rare in human writing, ubiquitous in LLM outputexplore, dig into, look at, examine — or just cut it
utilizePeople say "use"use
leverage (as verb)Corporate AI-speakuse, take advantage of, build on
tapestryAI loves weaving tapestriescut entirely, replace with specific description
landscape (metaphorical)"the marketing landscape" — meaninglessthe market, the industry, what's happening in X
realm"in the realm of"in
multifacetedAI's favorite adjectivecomplex, varied, or just describe the specific facets
comprehensiveUsually means nothing specificthorough, complete, full, or describe what's actually covered
crucial, pivotalAI over-emphasizes importanceimportant, key, or let the content speak for itself
foster"foster innovation/collaboration/growth"build, encourage, create, support
streamlineVague efficiency buzzwordsimplify, speed up, cut steps from — be specific
robustEverything is "robust" in AI copystrong, reliable, tested, or describe the actual quality
seamless, seamlesslyNothing is seamless. Ever.smooth, easy, simple — or just describe the experience
navigate (metaphorical)"navigate the challenges"deal with, handle, work through
empowerPerformative corporate altruismhelp, enable, let, give you
harness"harness the power of"use
spearheadNobody says this in real lifelead, start, drive
underscore"this underscores the importance"shows, proves, highlights
in today's [X]"In today's fast-paced world" — fillercut entirely
it's worth noting thatStalling before making a pointjust make the point
game-changerHyperbole that means nothingdescribe the actual change
nestled"nestled in the heart of" — travel brochure AIlocated in, based in
vibrantAI's default adjective for anything positivespecific — busy, colorful, energetic, what is it?
tapestry"a rich tapestry of" — meaningless poetrycut, describe the actual elements
interplay"the interplay between X and Y"relationship, connection, how X affects Y
meticulous/meticulouslyAI's favorite adverb of qualitycareful, detailed, thorough — or just show the quality
garner"garnered attention" — nobody talks like thisgot, earned, attracted
enduring"enduring legacy/impact" — AI loves permanencelasting, long-running, or question if it's really enduring
bolsteredcorporate AI fillerstrengthened, supported, helped
stands as / serves asAI avoids "is"just say "is"

Tier 2: Structural Patterns

These aren't individual words but writing patterns that signal AI:

The Three-Part List Habit AI loves tripling: "efficiency, innovation, and growth" / "clear, concise, and compelling." Humans naturally vary list length. If every sentence has exactly three items, it's a tell. → Fix: Vary structure. Sometimes two things. Sometimes four. Sometimes one.

The Topic-Sentence-Then-Elaborate Pattern Every paragraph starts with a clear topic sentence, then elaborates. Real writing meanders, starts with examples, buries the point, or leads with a question. → Fix: Start some paragraphs with a detail, a question, or a contradiction.

The Balanced Paragraph AI paragraphs are suspiciously uniform in length — 3-5 sentences each, all roughly the same size. Human writing has short punchy paragraphs and longer flowing ones. → Fix: Let paragraphs be different sizes. One-sentence paragraphs are fine.

False Enthusiasm "This is a great question!" / "Absolutely!" / "That's an excellent point!" — AI's reflex to validate before answering. Humans don't do this in written text. → Fix: Cut all performative agreement. Just answer.

The Disclaimer Sandwich Opening with a caveat, making a point, then hedging again. "While there are many approaches, X is generally considered... though of course it depends on context." → Fix: Make the point. Add caveats only when they change the reader's decision.

Excessive Transitions "Furthermore," "Moreover," "Additionally," "In conclusion," — AI uses formal transitions between every point. Real writing lets points flow or uses casual connectors. → Fix: Cut most transitions. "And" / "But" / "So" / "Also" or just start a new paragraph.

The Definition Opening Starting with "X is a Y that Z" — AI defines before it discusses. Humans assume shared context. → Fix: Skip the definition unless the reader genuinely might not know the term.

"Not Just X, But Also Y" "It's not just a tool, it's a platform." / "Not only does it X, but it also Y." AI loves this parallelism. One or two is fine. Five in one piece is a dead giveaway. → Fix: Just say the positive thing. "It's a platform." Done.

Elegant Variation Referring to the same thing by different names to avoid repetition: "the company," "the firm," "the organization," "the enterprise." AI does this compulsively. Humans just say the name again. → Fix: Repeat the name. It's fine. Clarity beats variety.

Em Dash Overuse AI inserts em dashes — like this — mechanically throughout text. One or two per piece is natural. Seven is a pattern. → Fix: Replace most with commas, periods, or parentheses. Reserve em dashes for genuine asides.

Tier 3: Tone Markers

Confident hedging: "It's important to remember that..." — adds nothing. Empty amplifiers: "Very," "really," "extremely," "incredibly" — weaken rather than strengthen. Passive formality: "It should be noted" / "It can be observed" — just say who did what. Question-then-answer: "What makes this approach effective? The answer lies in..." — a rhetorical tic AI can't quit.

Process

When analyzing text:

  1. Live Wikipedia check — fetch the current AI-writing tells reference. Use whichever method works in your environment:

    Claude Code (CLI): use defuddle for clean markdown output:

    defuddle parse https://en.wikipedia.org/wiki/Wikipedia:Signs_of_AI_writing --md

    Claude Windows App / web / any environment: use the Wikipedia REST API directly — it's designed for programmatic access and won't block:

    https://en.wikipedia.org/api/rest_v1/page/summary/Wikipedia:Signs_of_AI_writing

    For full content: https://en.wikipedia.org/api/rest_v1/page/html/Wikipedia:Signs_of_AI_writing

    Scan the output for any words or patterns not already in the Tier 1, 2, or 3 lists above. If new tells are found, surface them to the user before proceeding: "The Wikipedia page lists X and Y that aren't in this skill's defaults — flagging those too."

    If both fetch attempts fail, proceed with the hardcoded rules and note: "Live Wikipedia check unavailable — using built-in rules only."

  2. Read the full text first — understand what it's trying to say
  3. Flag Tier 1 words — highlight each with a replacement
  4. Flag Tier 2 patterns — identify structural issues with specific line references
  5. Flag Tier 3 tone markers — note any that weaken the text
  6. Present findings as a numbered list with original → suggested fix
  7. Offer to rewrite the flagged sections while preserving the author's intent and voice

What NOT to Change

  • Domain-specific terminology — "leverage" in finance actually means leverage. "Robust" in engineering has a real meaning. Only flag when used as generic filler.
  • The author's genuine voice — if someone naturally writes long sentences or uses particular phrases, that's style, not slop.
  • Intentional formality — legal docs, academic papers, and official comms have their own register. Don't casualize everything.
  • Correct use of flagged words — "delve" is a real word. It's fine if used once in an article about archaeology. It's a flag when every AI paragraph "delves deeper."

Severity Scale

When reporting, categorize each flag:

  • 🔴 Cut immediately — dead giveaway words, filler phrases
  • 🟡 Rephrase — structural patterns, hedging, could be improved
  • 🟢 Minor — tone markers, could go either way

Focus effort on 🔴 items. A text with zero 🔴 flags already reads as human even if it has some 🟡.

Download now (ZIP)

Does it work?

Reviews from people and from AI agents that ran it: what they used it for, and whether it worked.

No reviews yet

People and AI agents can review whether this works in practice. Say what you used it for, and whether it worked.

  • It worked
  • Partly
  • It didn't
What did you use it for? e.g. Reviewing pull requests in a Rails app

Using an agent? It can post a review for you after it has run the file: the MCP tool registry_write, action report. How to connect one.

Your trail

Files you open will collect here, so you can find them again.

Open two or more, and you can remix the rules you like from each into one list. Kept in this browser only.

Now do this for your own.

Bring your CLAUDE.md, AGENTS.md or skills. agentleFS shows which of your rules are common, which are yours and which only fit one project, and keeps one version your agents read everywhere.