| name | de-slop |
| platforms | ["cowork","claude-code"] |
| description | Remove AI writing patterns ('slop') from text to make it sound natural and human-written. Use this skill when the user asks to 'de-slop', 'remove AI tells', 'remove LLM tells', 'make this sound less AI', 'make this sound less like an LLM', 'make this more human', 'clean up AI writing', 'fix the AI voice', 'fix the LLM voice', 'this sounds like ChatGPT', or 'this sounds like an LLM'. Also trigger when the user pastes text that is clearly AI-generated and asks for editing help, or when content-reviewer flags AI voice issues and the user asks to fix them. Do NOT trigger for general writing requests (use brand-voice-writer) or for content evaluation (use content-reviewer). This skill is specifically for taking existing text and stripping out the patterns that make it read like AI output. |
| references | ["brand-voice-guide"] |
De-slop: Strip AI writing patterns
Remove the patterns that make text read like AI output. Inspired by the humanizer skill (MIT-licensed) and informed by Wikipedia's Signs of AI writing catalog, but restructured around our editing workflow and Tiger Data's brand context.
Why this matters: the same patterns that scream "AI wrote this" are also just bad writing habits. Vague where they should be specific, inflated where they should be plain, formulaic where they should be interesting. De-slopping makes text better, not just less detectable.
When to use this skill
- Someone pastes text and asks you to make it sound less like AI or less like an LLM
- Someone runs
/de-slop on a piece of content
- Someone says "this sounds like ChatGPT" or "fix the AI voice" or "remove the LLM tells" or "clean this up"
- Someone refers to AI or LLM writing patterns they want removed (people use both terms interchangeably)
- After content-reviewer flags AI voice issues and the user wants them fixed
- Someone asks to edit AI-generated or LLM-generated drafts before publishing
When NOT to use this skill:
- Writing new content from scratch (use brand-voice-writer)
- Evaluating content quality (use content-reviewer)
- General copyediting for grammar, clarity, or structure
Step 0: Pre-flight check
Invoke the marketing-preflight skill via the Skill tool, passing this skill's name: field value (de-slop) as the args. Do not proceed until it completes successfully. If it stops with an error, follow its instructions and stop.
If marketing-preflight returned a stale-connector footer note, append it verbatim at the end of your final response to the user.
Instructions
1. Read the full text first
Read the entire piece before changing anything. Figure out what it's actually trying to say underneath the slop. Who's the audience? What's the one thing the reader should take away? What tone is appropriate? You need to understand the intent before you can express it better.
2. Load brand context (for Tiger Data content)
If the text is Tiger Data marketing content (blog post, landing page, email, social post, etc.), fetch the brand voice guide from Tiger Den:
get_marketing_reference(slug: "brand-voice-guide")
If Tiger Den is not connected, tell the user to run /setup. See REFERENCES.md in the plugin root for details and error handling.
The brand voice guide already bans some of the same things de-slop catches (em dashes, passive voice, fluffy paragraphs). When both apply, follow the brand voice guide — it's more specific to our context.
If the text is not Tiger Data content, skip this step.
3. Identify the slop
Scan the text for these families of problems. You won't find all of them in every piece — focus on what's actually there.
The inflators
AI text loves to make everything sound important. Watch for language that inflates significance without adding information.
The tells: "pivotal moment," "testament to," "underscores the importance of," "reflects broader trends," "marks a shift," "evolving landscape," "setting the stage for," "deeply rooted." Also watch for stacking: a sentence that claims something "represents," "contributes to," or "symbolizes" something bigger without explaining how.
What to do instead: Say what happened. Use specifics. "The company was founded in 2019" not "The company's founding marked a pivotal moment in the evolution of the industry."
The fakers
AI text fills gaps in knowledge with confident-sounding vagueness.
The tells: "Experts argue," "Industry observers note," "Several publications have highlighted," "Some critics contend" — all without naming a single expert, observer, or publication. Also: "active social media presence," "has garnered attention," and any claim of notability that lists outlets without saying what they actually reported.
What to do instead: Name the source and what they said, or delete the claim. "A 2024 report by Gartner found..." beats "Industry reports suggest..."
The vocabulary cluster
Certain words appear 5-10x more often in post-2023 AI text than in human writing. One is nothing. Three in the same paragraph is a signal.
The words: additionally, bolstered, crucial, delve, emphasizing, enduring, enhance, fostering, garner, intricate/intricacies, interplay, landscape (figurative), meticulous, pivotal, showcase, tapestry (figurative), testament, underscore, vibrant. Also watch for "align with," "key" as an adjective for everything, and "valuable insights."
What to do instead: Replace with the simpler word you'd actually say out loud. "Important" not "crucial." "Show" not "showcase." "Complex" not "intricate." Or often, just delete the word entirely — "valuable insights" is usually just "insights," and "insights" is often just "data."
The structural tics
These are patterns in how AI builds sentences and paragraphs, not which words it picks.
Copula dodging: AI avoids "is" and "are" in favor of fancier constructions. "Serves as," "stands as," "functions as," "represents" — when the sentence just means "is." Fix: use "is."
Forced triples: AI groups things in threes compulsively. "Innovation, collaboration, and excellence." If there are two things, say two. If there are four, say four.
Negative seesaws: "Not just X, but Y." "Not merely a tool — it's a movement." "It's not about the technology, it's about the people." This construction appears once in a while in human writing. AI uses it every other paragraph. Fix: state the point directly.
Synonym rotation: AI has anti-repetition training, so it cycles through synonyms for the same thing. The protagonist... the main character... the central figure... the hero. Fix: pick one word and reuse it. Repetition is fine. Repetition is human.
Fake ranges: "From X to Y" where X and Y aren't really ends of a spectrum. "From ancient traditions to modern innovations." Fix: just list the things.
The formatting tells
Em dashes: AI overuses them for dramatic pauses and parenthetical asides. Our brand voice bans them entirely. Do not swap the em dash for another dash character — --, an en dash (–), or a hyphen are all still dashes and are not fixes. Read what the dash is doing and repunctuate or rewrite the sentence accordingly: a comma for a light aside, a period to split two independent clauses, parentheses for a true parenthetical, or a colon when the second half explains the first. If none of those read cleanly, rewrite the sentence so it doesn't need the break at all. The goal is a sentence with no dash, not a sentence with a different dash.
Bold-header lists: Bullet points where each item starts with a bolded label and colon ("Performance: The system is fast"). Convert to prose or use plain list items.
Mechanical bold: Bolding terms like a textbook highlighting vocabulary words. Remove unless there's a real reason for emphasis.
Title Case Headings: AI capitalizes every word. Use sentence case.
Emoji garnish: Rocket ships and checkmarks decorating headers. Delete.
The chatbot residue
Text that was clearly pasted from a conversation without cleanup.
Pleasantries: "Great question!", "I hope this helps!", "Certainly!", "Let me know if you'd like me to expand on any section!"
Hedges from training: "As of my last update," "While specific details are limited," "based on available information." Either find the information or don't mention it.
Filler connectives: "In order to" (just "to"), "due to the fact that" (just "because"), "it is important to note that" (delete, then state the thing), "at this point in time" ("now").
Sycophancy: "You're absolutely right!", "That's an excellent point!" Delete on sight.
Hallmark-card endings: "The future looks bright." "Exciting times lie ahead." "This is just the beginning." End the piece when it's done. The last paragraph should contain information, not warm feelings.
4. Rewrite — and put a person behind it
Stripping slop is the first job. But a scrubbed text that's grammatically clean and factually accurate can still read as obviously AI-generated if nobody's home. The absence of bad patterns isn't the same as the presence of a human voice.
Things that make writing feel like a person wrote it:
Uneven rhythm. Real writing doesn't march at a steady pace. Some sentences are three words. Others unspool over a full paragraph because the idea needs room. If every sentence in a stretch is 15-20 words, something is off.
Actual reactions. Human writers respond to their own material. "That number surprised me." "I keep coming back to this." "Honestly, I'm not sure what to make of it." AI reports facts. People react to them.
Willingness to not know. "I'm not sure this is the right framing" or "there's probably a better way to think about this" — these signal a real mind at work. AI always sounds certain, even when hedging.
Mess and tangent. Not every paragraph needs to land one clean point in perfect order. Sometimes a parenthetical aside, a half-formed thought, or a detour into a related story is what makes a piece feel alive. Don't overdo it, but don't sand everything smooth either.
Specificity over abstraction. "The dashboard loads in 200ms" instead of "the dashboard is fast." "We tested this on a customer's 500GB Postgres instance" instead of "we tested this at scale." Concrete details are harder for AI to generate and more interesting to read.
When rewriting, also preserve the core meaning, the intended audience and tone, any sourced data or quotes, and the author's point of view if one exists beneath the slop.
5. Check your own work
After rewriting, read it back and ask: "Would I believe a person wrote this?" Be honest about what survives. Common second-pass catches:
- Every paragraph is still roughly the same length
- The piece is suspiciously well-organized (thesis, evidence, evidence, conclusion — too clean)
- You introduced new AI patterns while removing old ones (it happens)
- You replaced em dashes with
--, en dashes, or hyphens instead of actually repunctuating or rewriting (search the text for any remaining dash characters and fix them)
- The closing still trails off into vague optimism
- It's clean but forgettable — technically fine, zero personality
Fix what you find. This second pass matters.
6. Present the output
Provide the rewritten text and a brief changes summary grouped by what you fixed (not a line-by-line diff). If the piece was long, note which sections needed the most work so the user knows where to focus their review.
Voice matching
If the user mentions who should have written the piece (e.g., "de-slop this to sound like Matty," "clean up Jacky's draft"):
- Call
get_voice_profile with the author's name
- Use their writing samples and voice notes to guide the rewrite — not just removing AI patterns, but actively matching that person's rhythm, tone, and verbal habits
- Note voice-matching adjustments in the changes summary
If no specific person is mentioned, skip the voice profile.
Relationship to other skills
- brand-voice-writer writes new content. De-slop edits existing text.
- content-reviewer evaluates quality. De-slop fixes a specific class of problems.
- If content-reviewer flags AI voice issues, de-slop is the fix.
- After de-slopping Tiger Data content, suggest running content-reviewer as a final check.
Credits and references
Pattern catalog informed by Wikipedia:Signs of AI writing, maintained by WikiProject AI Cleanup. Approach inspired by the humanizer skill by blader (MIT license).
