Remove AI Slop

Remove AI Slop

Identify and remove AI-slop: the filler, hedging, marketing, and jargon phrases that AI text generators lean on, the structural tells they use to manufacture insight, plus em-dash overuse. Output is cleaner prose with the meaning, facts, and technical terms intact.

This skill is a reviewer/cleaner, not a writer. It edits what is there; it does not invent content or restructure the document.

It covers two kinds of slop:

• Word-level (words and phrases) — see references/phrases.md.
• Structural (constructions and rhythm) — see references/structures.md.

Workflow

1. Identify the setting. Determine the input: a file on disk to edit in place, or inline prose to return cleaned. If you write to a .md file, never wrap the result in a code fence — the file content IS markdown.

2. Scan the text. Flag every match of the categories below. Record the exact location (line or sentence), the current phrase, and a proposed replacement.

3. Report before editing (unless the user asked for auto-fix). Present findings as a numbered table where each entry gets a number starting at 1:

   #	Location	Phrase	Suggested fix

   Let the user accept or reject findings before you apply changes.

4. Apply the selected fixes. The user picks which findings to fix by number. Keep the meaning, facts, and technical terms intact.

   Selecting findings. Selection is 1-based and can be a single number, a comma-separated list, a range, or any combination:

   • Single: 3
   • List: 1,2,5
   • Range: 1-3
   • Combined: 1-3,5,7-8

   Whitespace after commas is allowed and ignored (1, 3 → findings 1 and 3). An out-of-range number is not an error in the whole selection — apply the valid findings and report the ones that don't exist. Apply only what the user selected; leave the rest untouched for a later pass.

5. Re-verify against the validation checklist at the bottom, fix any remaining violations, then present the result.

AI-slop detector

Flag each pattern below and replace it with plain, direct language. Many are omittable outright. Full detail is in the reference files.

Word-level

• Adverbs — cut all of them (the -ly blanket rule). This includes "simply", "just", "easily", "effortlessly", "seamless(ly)", "obviously", "actually", "literally", "truly", "really", and all -ly softeners, intensifiers, and hedges. See references/phrases.md.
• Filler and throat-clearing openers — cut: "Here's the thing/why/this/ that", "It turns out", "The uncomfortable truth is", "Let me be clear", "In today's fast-paced world", "It is important to note", "In conclusion".
• Emphasis crutches and generic filler — cut: "Full stop.", "Let that sink in.", "Make no mistake", "At its core", "At the end of the day", "When it comes to", "In a world where".
• Jargon and leverage verbs — plain-language replacement: "utilize" → "use"; "leverage" → "use"; "streamline" → "simplify"; "delve" → "examine"; "harness/unleash/empower/revolutionize/unlock" → concrete verb; "navigate/unpack/lean into/deep dive/circle back/moving forward/ game-changer/double down/landscape" → plain synonym. See the jargon tables in references/phrases.md.
• Meta-commentary — cut self-referential asides ("The rest of this essay explains…", "In this section, we'll…").
• Vague declaratives — cut or name the specific thing ("The implications are significant", "The stakes are high").
• Telling instead of showing — name the concrete thing ("This is genuinely hard", "this is what X actually looks like").
• Future tense for current behavior — change to present tense: "the tool will create" → "the tool creates"; "this app will let you" → "this app lets you".

Structural

• Binary contrasts — cut "not X, it's Y", "isn't X, it's Y", "X isn't the problem, Y is". State Y directly.
• Negative listing — cut "not a X... not a Y... a Z". State Z.
• Dramatic fragmentation — cut "X. And Y. And Z.", "[Noun]. That's it." Use complete sentences.
• Rhetorical setups — cut "What if [reframe]?", "Here's what I mean:", "Think about it:".
• False agency — inanimate things doing human verbs ("the decision emerges", "the market rewards", "the data tells us"). Name the human.
• Passive voice — "X was created", "It is believed that". Find and name the actor.
• Starters to avoid — Wh- openers, paragraphs starting "So", sentences starting "Look,".
• Rhythm — two items beat three; vary paragraph endings; no stacked fragments; no metronomic cadence.
• Lazy extremes — "every/always/never/everyone/nobody" as false authority; use specifics.

See references/structures.md for the pattern tables.

Wordiness and tone

• No exclamation marks, no rhetorical questions, no emojis.
• Sentences ≤ 25 words; split nested clauses.
• Prefer common words: "use" not "utilize", "start" not "commence".
• No marketing superlatives ("best", "unparalleled", "game-changing").

Em-dash rule

Default = true (remove em-dashes). An em-dash (—) is a common AI-slop tick; strip it unless the user explicitly opts out.

• Replace an em-dash with a comma, a parenthetical, or by splitting and restructuring the sentence — whichever reads most clearly.
• Treat more than one em-dash per paragraph as overuse even if the user opts to keep them.
• Preserve an em-dash only when it is technically necessary (rare) and the user asked to keep em-dash usage.

Gotchas

• Never change meaning. Remove slop, but keep intent, facts, and technical terms. Ask if an ambiguous edit could change what the author meant.
• Don't write what isn't slop. This skill edits AI-slop only; it does not impose a new structure or voice on competent existing prose.
• Don't invent an actor. When fixing false agency or passive voice, name a real actor or use "you". Never fabricate who did something.
• Don't strip legitimate dashes. Em-dashes in favor of a comma are an edit; a hyphen in a compound word (state-of-the-art) is not an em-dash.
• Don't invent replacements. Prefer omission or a known substitution to guessing at a synonym the author didn't intend.
• Don't wrap the output in a code fence when editing a .md file.

Validation checklist

Before delivering, verify:

• No filler, hedging, or marketing adverbs remain ("simply", "just", "easily", "seamless(ly)", "actually", "literally", "truly", "really", "-ly" softeners)
• No empty openers remain ("Here's the thing", "In today's fast-paced world", "It is important to note", "In conclusion")
• No jargon or leverage verbs remain ("leverage", "utilize", "streamline", "delve", "navigate", "lean into", "deep dive", "circle back", "robust", "cutting-edge", "best-in-class")
• No meta-commentary remains ("The rest of this essay explains…")
• No vague declaratives remain ("The implications are significant") — the specific thing is named
• No false agency or passive voice remains — a real actor is named
• No binary contrasts, negative listing, or dramatic fragmentation remain
• No rhetorical setups or rhetorical questions remain
• No future tense for current behavior ("will create" → "creates")
• No em-dashes remain (unless the user opted to keep them)
• No exclamation marks, rhetorical questions, or emojis
• Every sentence ≤ 25 words
• Rhythm is varied — no metronomic cadence, no stacked fragments
• Meaning, facts, and technical terms unchanged