| name | ai-slop-killer |
| description | The final anti-AI-slop editing pass for any drafted or revised prose — blog posts, newsletters, LinkedIn/social, emails, marketing and web copy, documentation, essays, fiction. Detects documented AI writing tells (negative parallelism "not X, it's Y," rhetorical Q&A, rule-of-three and anaphora pile-ups, em-dash and unicode overuse, corporate-AI vocabulary, wind-up intros and signposted conclusions, bold-first-bullet listicles, low-burstiness rhythm, hedging, vague attributions, participial "-ing" filler, grandiose stakes, invented concept labels), scores the passage, rewrites in place, and reports what it caught. Runs LAST, after any other writing or editing skill. Trigger on "clean this," "de-slop," "kill the AI slop," "copyedit," "does this sound like AI," "make this sound human," "humanize this," or automatically after any skill produces prose meant to be published. |
AI Slop Killer
The last skill to touch prose before it ships. Whatever drafted or revised it, this runs the passage through one job: make it read like a person wrote it on purpose, then prove what it removed. Works on any genre — business writing, marketing, docs, essays, fiction.
The pattern catalog is built from documented sources (Wikipedia's Signs of AI writing, the tropes.fyi directory, working editors' guides — see reference/slop-catalog.md), not one person's taste. It ships with sensible defaults and an optional per-user config for house rules.
The loop: Detect → Score → Fix → Prove
- Read the passage cold, start to finish, before touching anything. Slop clusters — you're looking for stacked tells, not one stray word.
- Detect. Run every check in
reference/slop-catalog.md. Match on the pattern, not exact wording. A single instance of most tells is human; the tell is the repetition and clustering.
- Score. Compute the slop score using
reference/scoring.md. If it's ≤5 per 500 words, say "clean" in one line and stop — don't manufacture edits.
- Fix in place. Rewrite the tells. Preserve the author's meaning and voice. If a voice/style skill or a config is available (see below), honor it. Never flatten a correct term of art into vagueness.
- Prove. Return the cleaned passage, then the before→after score and an itemized list of what was cut and why — plus anything left in on purpose.
The non-negotiables
- Kill negative parallelism — "It's not X, it's Y," "not because X but because Y," "The question isn't X. The question is Y." The single most-cited AI tell. At most one earned instance per piece.
- Cut structural scaffolding. Wind-up intros ("In today's fast-paced world…"), signposted conclusions ("In conclusion," "At the end of the day"), false-suspense transitions ("Here's the kicker," "Here's the thing"), and rhetorical-question pivots ("The result? Devastating."). Delete and start at the real sentence — the true article usually starts in paragraph two.
- Break the rhythm. No stacked rule-of-three triplets, no anaphora (three sentences opening the same way), no page of same-length "rectangle" paragraphs. Vary sentence length hard (long, explanatory sentence → short punch). This is burstiness, the thing readers and detectors both notice.
- Two weights of emphasis, used sparingly. No paragraph that's mostly em-dashes (a human uses 2–3 per piece, not 20). No bold-first-bullet on every list item. Normalize curly/straight quotes and stray unicode arrows.
- Plain words beat AI vocabulary. Replace the documented offenders (delve, leverage, robust, seamless, tapestry, landscape, realm, testament, elevate, harness, streamline, foster, pivotal, myriad…) with a concrete noun or an active verb. The pub test: would you say this to a colleague over a beer?
- Substance, not fake authority. No vague attributions ("experts say," "studies show") without a named source. No grandiose stakes inflation. No invented concept labels ("the supervision paradox") used as if they're established terms. No fake first-person experience ("Picture this…").
- Watch the humanizer over-correction. Thesaurus salad ("paramountly significant") from de-detection tools is its own tell. If a word makes the sentence harder to say aloud, revert it.
Don't over-correct
One em-dash, one triad, one rhetorical question, one metaphor per piece is human. Deliberate fragments, real lists of real steps, correct technical terms, and a genuine strong stance all stay. The failure mode cuts both ways: leaving obvious machine cadence, OR scrubbing prose until it's beige and voiceless. Aim for lived-in, not sanitized.
Config (optional, for house rules)
If a file named slop.config.md exists in the working directory or alongside the content, load it. It can define:
banned_words: extra words to hard-ban on top of the defaults (e.g., a brand's forbidden list).
allowed_words: documented-tell words this author legitimately uses (removes false positives).
voice: the name of a voice/style skill to preserve (e.g., a personal or brand voice skill), or a short voice note.
citations: required | preferred | off — whether every factual claim needs an inline source.
channel_defaults: per-channel tweaks (e.g., allow emoji in social, forbid in docs).
With no config, defaults apply: standard documented banlist, citations preferred, preserve whatever voice the text already has.
Fiction mode
Say "fiction mode" or run this on manuscript prose. Adds the section-7 checks in the catalog: one negative-parallelism per book max, don't explain a metaphor after the scene already made it, don't beat one metaphor to death (the "dead metaphor" trope), prefer real checkable texture over invented brand names, and catch paste artifacts (duplicated/garbled sentences). Series-specific metaphor limits belong in slop.config.md.
Output
Slop score: N → M (per 500 words)
[cleaned passage]
Cut / fixed:
- [pattern] xN — [what you did]
...
Left on purpose:
- [thing] — [why]
If clean on first read: one line saying so, with the score. Stop.