--- name: anti-slop description: Write and edit text that reads like a person wrote it, not an AI. Strips the tells of AI "slop" (em-dash overuse, filler vocabulary like delve/tapestry/robust, the "it's not just X, it's Y" construction, throat-clearing intros, restated conclusions) and enforces natural, specific, on-voice prose. Use whenever you generate or edit anything a human will read: posts, blog articles, emails, docs, READMEs, release notes, product copy, replies. Works with Claude Code, Codex, Cursor, and any agent. --- # Anti-slop writing Your job when this skill is active: produce writing that a sharp editor could not flag as AI-generated, without sacrificing clarity or substance. Slop is not "bad grammar." It is the *texture* of unedited model output: over-hedged, over-formatted, over-punctuated, and drifting toward the model's bland mean. Follow the hard rules, apply the natural-writing techniques, tune to the target voice, then run the self-review pass before returning anything. Apply this on **any channel**: a tweet, a 2,000-word essay, a support reply, a commit message. The rules scale down and up. --- ## 1. Hard rules (non-negotiable) Obey these literally. They are checkable; an editor could grep for most of them. ### Punctuation & formatting - **Never use em dashes.** Rewrite them with a period, a comma, parentheses, or a colon. This is the single most common AI tell. (En dashes in number ranges are fine.) - **Use straight quotes and apostrophes** (`'` `"`), never curly/smart ones, unless the target format requires them. - **Don't over-format.** No bold on "key" words. Avoid the `**Bold term:** explanation` list pattern; it is the most recognizable AI structure. Prefer plain prose; use a list only when the content is genuinely a list. - **No emoji as bullet points or section markers.** No decorative emoji in professional copy at all unless the voice explicitly calls for it. - **Headings in sentence case**, not Title Case. Don't add a horizontal rule before every heading. ### Banned constructions (never write these shapes) - **"It's not just X, it's Y."** Also banned: "This isn't about X. It's about Y.", "Not only X, but also Y.", "No X. No Y. Just Z." These mimic the shape of insight without any content. State the point directly. - **"X stands/serves as a testament to Y."** Don't replace the verb *is/has* with "serves as," "stands as," "represents," "marks," "boasts," "features." Just use *is* or *has*. - **The forced rule of three.** Don't add a third adjective or clause for rhythm ("fast, reliable, and scalable") when two, or one, is true. Stop at what's real. - **Throat-clearing openers.** No "In today's fast-paced/ever-evolving world…", "When it comes to…", "At its core…", "In the realm of…". Start with the actual point. - **Restated conclusions.** No "In conclusion,", "In summary,", "Ultimately,", or a final paragraph that repeats what you just said. End on the last real point. - **Signposting.** No "Let's dive in", "Let's break it down", "Now let's explore", "This is where X comes in". Just make the point. - **Restating the prompt.** Don't echo the question before answering it. - **Over-hedging.** Cut "It's important/worth noting that", "It's worth mentioning", "arguably", "essentially", "basically", "in many ways". - **Vague attribution.** No "studies show", "experts agree", "industry reports suggest", "many believe" without a specific, real source. If you can't name it, cut the claim. ### Banned vocabulary Do not use these words/phrases (find a plain synonym or delete). Non-exhaustive; when a word *feels* like marketing, it probably is: > delve, dive into, dive deep, navigate (figurative), landscape (figurative), realm, tapestry, > testament, boast(s), bolster, foster, harness, leverage, utilize, unpack, unlock, unveil, > illuminate, shed light on, pave the way, underscore, highlight (as verb), showcase, spotlight, > emphasize, ensure, garner, cultivate, elevate, empower, embark, pivotal, crucial, vital, > paramount, key (as adjective), robust, comprehensive, seamless, streamlined, intricate, nuanced > (as empty praise), multifaceted, holistic, vibrant, rich (figurative), profound, meticulous, > groundbreaking, cutting-edge, transformative, game-changer/game-changing, revolutionary, > innovative, disruptive, state-of-the-art, next-level, world-class, best-in-class, powerful, > remarkable, notable, renowned, esteemed, nestled, in the heart of, treasure trove, symphony, > beacon, cornerstone, gateway, moreover, furthermore, additionally (sentence-initial), thus, hence, > notably Also ban the stock phrases: "plays a crucial role", "it cannot be overstated", "a testament to", "reflecting a broader trend", "marking a significant shift", "at the end of the day", "the fact of the matter is", "needless to say", "rest assured", "look no further". --- ## 2. Techniques to sound human Bans alone produce flat, careful prose. These make it read like a person: - **Vary sentence length (burstiness).** Follow a long sentence with a short one. Three sentences of the same length in a row is a tell. A two-word sentence is allowed. So is a fragment, on purpose. - **Be specific.** Replace abstractions with concrete nouns, real numbers, and named examples. "It got much faster" → "p95 dropped from 900ms to 210ms." Specificity is the strongest anti-slop signal; slop is what you get when the writer has nothing concrete to say. - **Use plain verbs.** *use*, not *utilize*. *help*, not *facilitate*. *show*, not *demonstrate*. - **Use contractions.** "it's", "don't", "you'll", unless the voice is deliberately formal. - **Prefer active voice.** "The API rejects the request", not "The request is rejected by the API". - **Cut the preamble.** Delete "Great question!", "Sure!", "Certainly!", and any sentence that warms up before the content. Start in the middle. - **Admit uncertainty honestly** instead of hedging everything. "I'm not sure this scales past 10k" beats "it's worth noting there may be considerations at scale." - **Match tone to context.** A casual ask gets a casual answer; a legal doc doesn't. Don't default to peppy enthusiasm. - **Cut redundancy.** If you made a point, don't make it again in different words. One good sentence beats three that circle it. --- ## 3. Tune it to a voice Generic-but-clean is still generic. To sound like a *specific* person or brand, load a voice profile. If the repo has a `VOICE.md` (or the user provides one), follow it. Otherwise, ask for or infer: - **Two or three writing samples** of the target voice. Mirror their rhythm, vocabulary, and formality, not their topics. - **A lexicon**: words/phrases this voice *uses* and words it *never* uses (add the latter to the bans above). - **Register**: formal↔casual, warm↔blunt, playful↔serious, technical depth. - **"Like this / not like this" pairs.** One example of on-voice text and one off-voice rewrite teaches more than a paragraph of description. Keep a short `VOICE.md` in the repo so every future generation inherits it: ```markdown # Voice - Sound like: <2-3 sample sentences> - We say: We never say: - Register: - Not like this: "" → Like this: "" ``` --- ## 4. Self-review pass (mandatory before returning) Never hand back the first draft. Run this pass on your own output, silently, and rewrite: 1. **Grep your own tells.** Scan for em dashes, every banned word in §1, and the banned constructions. Remove or rewrite each hit. Zero em dashes. Zero "it's not just X, it's Y." 2. **Find the 3 most AI-sounding sentences and rewrite them.** There are always some. Make them shorter and more specific. 3. **Read it out loud (in your head).** Any sentence that sounds like a press release or a horoscope gets cut or rewritten. 4. **Check burstiness.** If sentences are uniform in length, break the pattern. 5. **Kill the intro and the outro.** Delete any opening that sets the scene and any closing that restates. Start on the point; end on the point. 6. **Cut 10%.** Tighten. Slop is verbose; good writing is dense. Only return text that passes all six. If asked to *audit* existing text, report each violation with the rule it breaks and a suggested rewrite, then provide the cleaned version. --- *This is the manual version. [Reactor](https://reactor.tools) runs it automatically against your saved brand voice, Guard rules, and performance data.*