# Anti AI Slop > Removes AI-slop from any writing before it ships — the tells that make text read as machine-generated: filler openers, hedge-everything qualifiers, rule-of-three padding, "it's not just X, it's Y" antithesis, em-dash abuse, empty transitions, and self-congratulatory adjectives. Two modes — Quick Scan (flag + score, no rewrite) and Deep Rewrite (voice-first reconstruction). Activate on ANY writing or editing task: drafting copy, emails, docs, posts, PR descriptions, marketing content, summaries, or when asked to "make this sound human", "de-slop", "tighten", or "cut the fluff". - Skill: `cleanexpo/anti-ai-slop` (Agent Skill) - Install (CLI): `npx skillmds@latest add cleanexpo/anti-ai-slop` - Raw SKILL.md: https://api.skillmd.com/api/skills/cleanexpo/anti-ai-slop/raw - Safety review: pending - Works with: Claude Code, Claude.ai, OpenAI Codex - Category: Marketing & Growth - Author: cleanexpo (https://skillmd.com/u/cleanexpo) - Updated: 2026-09-17 - Page: https://skillmd.com/skills/cleanexpo/anti-ai-slop --- # anti-ai-slop A quality gate for prose. It catches the mechanical tells of AI-generated text and forces writing back toward a specific human voice: concrete, load-bearing, and unafraid to make a plain claim. Applies to all writing tasks by default. Do not wait to be asked — if you are producing prose a human will read, run the gate. ## Two modes **Quick Scan** (default for short text, review requests, "does this read as AI?") - Read the draft, flag every hit against the banned-phrases table and structural rules below. - Return a slop score (0–100, lower is cleaner) and a bulleted list of specific hits with line references. - Do NOT rewrite. The author decides. **Deep Rewrite** (for "fix this", "make it human", "de-slop", or new drafts) - Run the voice-first workflow below, producing a clean draft. - Show the slop score before and after. - Preserve the author's meaning and facts exactly — cut ornament, never substance. Pick Deep Rewrite when the task is to produce or repair text; Quick Scan when the task is to assess it. ## Banned phrases and patterns | Category | Banned | Use instead | | ---------------------- | -------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------- | -------------------------------------- | | Filler openers | "In today's fast-paced world", "In the ever-evolving landscape of", "When it comes to", "It's important to note that", "It's worth noting that" | Delete. Start at the first real claim. | | Empty transitions | "Moreover", "Furthermore", "Additionally", "That said", "With that in mind", "At the end of the day" | A period. Or nothing. | | Hedge stacking | "may potentially", "could possibly", "generally tends to", "in many cases can" | State it, or state the condition once. | | Self-praise adjectives | "seamless", "robust", "cutting-edge", "world-class", "game-changing", "powerful", "comprehensive", "innovative", "elevate", "unlock", "leverage", "delve", "navigate the complexities" | Say what it does. Show, don't assert. | | Antithesis cliché | "It's not just X, it's Y", "It's not about X — it's about Y" | Make one claim. Drop the fake pivot. | | Rule-of-three padding | reflexive triads: "fast, reliable, and scalable" | Keep the one that carries weight. | | Closing boilerplate | "In conclusion", "Ultimately", "At its core", "The bottom line is" | End on the last real sentence. | | Corporate verbs | "utilize", "facilitate", "streamline", "spearhead" | use, help, simplify, lead | | Vague intensifiers | "very", "really", "truly", "incredibly", "actually" | Cut. If it matters, quantify it. | ## Structural pattern rules 1. **No listicle reflex.** Prose is not always three bullets. Use a paragraph when the ideas connect; use a list only when items are genuinely parallel and unordered. 2. **Em-dash budget.** At most one em-dash per paragraph. AI over-uses them as an all-purpose joint. Prefer a full stop or a comma. 3. **Vary sentence length.** If three consecutive sentences are within ~5 words of each other, break the rhythm. Slop is metronomic. 4. **One idea per sentence, one claim per paragraph.** Padding hides behind subordinate clauses. 5. **Concrete over abstract.** Replace "solutions", "capabilities", "offerings" with the actual thing. 6. **No hedged conclusions.** If the draft argues a point, land it. "It depends" is only allowed when you then say on what. 7. **Cut the meta.** Delete sentences that describe the text itself ("This section will explore…", "Let's dive into…"). 8. **Active voice by default.** Passive only when the actor is genuinely unknown or irrelevant. ## Quality gate thresholds Compute a slop score: start at 0, add per hit — - Banned phrase: +4 - Structural rule violation: +6 - Antithesis cliché or closing boilerplate: +8 Thresholds: - **0–10 — Ship.** Clean. - **11–25 — Revise.** Fix the flagged hits; usually surface-level. - **26+ — Rewrite.** Escalate Quick Scan to Deep Rewrite; the draft is structurally slop. Never pass text scoring 26+ off as final in Deep Rewrite mode. ## Voice-first drafting workflow (Deep Rewrite) 1. **Say it out loud first.** Draft the point as you would say it to one specific person across a table. That sentence is the spine. 2. **Lead with the claim.** Put the load-bearing sentence first. Cut every word before it. 3. **Support with the concrete.** One real example, number, or mechanism beats three adjectives. 4. **Read for rhythm.** Vary sentence length. Read it aloud; if you stumble or drone, cut. 5. **Run the gate.** Score against the table and structural rules. Fix every hit above threshold. 6. **Delete 10%.** After it reads clean, cut another tenth. The strongest version is shorter than you think. ## What this skill will not do - It will not flatten a deliberate, distinctive voice into grey neutrality. The target is human and specific, not sanded-down. - It will not change facts, claims, or the author's argument. It cuts ornament, not substance. - It will not ban a word the context genuinely earns — the tables are defaults, not absolutes. Flag, judge, then decide.