--- name: de-slop description: Rewrites any text (emails, posts, landing pages, proposals, deck copy, scripts) so it reads like a specific, confident human wrote it instead of an AI. Removes AI-tell phrases and patterns, adds concrete detail placeholders, varies sentence rhythm and takes a position. Use when the user says "de-slop", "make this sound human", "this sounds like ChatGPT", "remove the AI tone", "humanize this", or pastes AI-drafted copy before sending it to a client or publishing it. --- # De-slop You turn AI-sounding text into text a sharp human would write. You never invent facts to make it sound specific. Where specifics are missing, you leave a visible `[ADD: ...]` gap so the user fills it with the truth. ## When to run - The user pastes text and asks to make it sound human, less AI, less generic, more like them. - The user says "de-slop this" or "run de-slop". - Before any client-facing text leaves the chat, if the user has asked you to check for AI tone. If the user also has a style guide from the `voice-clone` skill (a block that starts with `VOICE GUIDE:`), apply it on top of these rules. Their voice wins over these defaults when the two conflict (for example, if their guide says they love em dashes, keep a few). ## Step 1: Diagnose (silent, unless asked) Read the text once and note: 1. **Job**: what is this text supposed to make the reader do? (reply, book a call, approve a budget, buy, trust you) 2. **Reader**: who exactly reads it? One person, by role. 3. **Tells**: scan for the 10 categories in `banned-phrases.md` (in this skill folder) (vocabulary, openers, transitions, hedging, significance inflation, intensifiers, chatbot residue, generic CTAs, deck clichés, structural patterns). 4. **Missing specifics**: every place a number, name, date, example or result would make the claim believable. 5. **The real point**: if you had to say it in one sentence to a friend, what is it? If the user asks "what's wrong with it?", output this diagnosis as a short list and stop. ## Step 2: Rewrite with the 12 rules 1. **Lead with the point.** Delete the first sentence if it's warm-up. The reader should know what you want within the first two lines. 2. **Concrete nouns over abstractions.** "Your checkout" not "the user experience". "The 3 clients who churned in May" not "customer retention challenges". 3. **Numbers, names, dates.** Use every real one the source gives you. Where one is needed and missing, write `[ADD: number of X]`. Never make one up. 4. **Plain words.** use, not utilize. help, not empower. show, not showcase. start, not embark. If a 10-year-old wouldn't know the word and there's a simpler one, use the simpler one. 5. **Vary the rhythm.** Mix very short sentences (2–6 words) with a few long ones (20–30 words) that carry detail. Never three sentences in a row of similar length. 6. **One claim per sentence.** Split sentences that do two jobs. 7. **Take a position.** Replace "it depends" and both-sides endings with a recommendation and who it's for. If the source is genuinely unsure, say what it depends on and pick a default. 8. **No banned phrases.** Nothing from `banned-phrases.md` (in this skill folder). Don't swap them for fancy synonyms; replace them with facts or delete them. 9. **Kill the AI shapes.** No "it's not X, it's Y". No "The result? ...". No triplet lists unless there really are three things. No em dashes (use a period or a comma). No "From X to Y" ranges. No moral at the end of every paragraph. 10. **Write to one person.** "You", singular. Contractions are fine (you're, it's, don't). 11. **Earned emphasis only.** No "truly", "incredibly", "very". If something is impressive, the number shows it. 12. **End on the action.** The last line says exactly what happens next, with a time or a step. No "Let me know if you have any questions." Keep: the meaning, the facts, the format (email stays an email, bullets stay bullets if they help), and roughly the same length or shorter. Shorter is almost always better. ## Step 3: Self-check before you answer Go through this list. If any answer is "no", fix it before replying. - [ ] Zero phrases from the banned list? (Search your output for: delve, landscape, leverage, seamless, robust, moreover, furthermore, additionally, crucial, pivotal, foster, unlock, elevate, empower, journey, testament, "not just", "it's important to note", em dash.) - [ ] At least one concrete specific (or `[ADD: ...]`) per paragraph? - [ ] Sentence lengths visibly varied? - [ ] The first line carries the point? - [ ] The last line is a specific next step? - [ ] Nothing invented? Every number and name came from the source or is an `[ADD]` gap? - [ ] Would a smart friend reading it out loud sound like a person? ## Step 4: Output format Return exactly this: ``` REWRITE GAPS TO FILL - [ADD: ...] → what to put there and why it matters (one line each) (or "None") WHAT CHANGED (3 lines max) - ... ``` If the user asked for "just the text", return only the rewrite. ## Examples The specifics in each "after" came from the user (listed under **Facts given**). That's the point: the skill pulls real details out of you, or leaves `[ADD]` gaps. It never makes them up. ### 1. Cold email **Before (AI):** > Subject: Elevating Your Digital Presence > > Hi Sarah, I hope this email finds you well. In today's competitive landscape, having a robust online presence is crucial. Our agency specializes in crafting tailored solutions that empower businesses like yours to unlock their full potential. Would you be open to a quick call to explore how we can help? **Facts given:** I tried booking on her site; the form has 11 fields before a time slot; I fix booking flows for dental clinics in about 2 days. **After (human):** > Subject: Your booking page loses people at step 3 > > Hi Sarah, I tried to book a cleaning on your site last night. The form asks for 11 fields before it shows a time slot. I gave up at the insurance question. Most of your visitors probably do too. > > I fix this kind of page for dental clinics. Usually it's a 2-day job: move the time slot first, ask for details after. [ADD: one result from a past client, e.g. "Dr. Rossi's bookings went from X to Y a month"] > > Want me to send a 2-minute screen recording of what I'd change? No call needed. Why it works: a real observation replaces the flattery, the offer is one concrete fix, the CTA asks for something tiny. ### 2. LinkedIn post **Before (AI):** > In today's fast-paced world, productivity isn't just about working harder — it's about working smarter. 🚀 Here are 3 game-changing tips that transformed my workflow: 1️⃣ Leverage AI tools 2️⃣ Prioritize ruthlessly 3️⃣ Embrace deep work. The result? I've never been more productive. What are your thoughts? 👇 **Facts given:** No Slack before 11am since March; 4 complaints in week one; shipped the billing rewrite that month. **After (human):** > I stopped answering Slack before 11am in March. > > First week: 4 people complained. Second week: nobody noticed. By the end of the month I'd shipped the billing rewrite that had been "almost done" since January. > > The rule is dumb and that's why it works. No system, no app. Phone in another room until 11. > > What's the one notification you'd kill first? Why it works: one specific change, real timeline, a small admission ("the rule is dumb"), and a question someone can actually answer. ### 3. Landing page hero **Before (AI):** > Revolutionize Your Workflow with Our Cutting-Edge AI Platform. Seamlessly streamline your operations, enhance collaboration, and unlock unprecedented productivity — all in one powerful solution. **Facts given:** The product records a sales call and drafts a proposal; editing takes about 10 minutes. **After (human):** > Turn every client call into a signed proposal in 10 minutes. > > Record the call. We draft the proposal from what the client actually said: scope, budget, deadline. You edit, send, done. [ADD: proof line, e.g. "Used by X freelancers / Y proposals sent last month"] Why it works: one job, one number, the mechanism in three steps, a gap for real proof instead of a fake one. ### 4. Strategy memo paragraph **Before (AI):** > Moving forward, it's crucial that we adopt a holistic, data-driven approach to customer retention. By leveraging actionable insights across key touchpoints, we can drive meaningful engagement and foster long-term loyalty. **Facts given:** Roughly 25% churn in month two; churned accounts mostly skipped the integration. **After (human):** > We lose 1 in 4 customers in their second month. [ADD: confirm the exact figure from the cohort report] Almost all of them never set up the integration. So for Q1 we do one thing: every new account gets a 15-minute setup call in week one. If second-month churn doesn't drop below [ADD: target]% by March, we stop the calls and try in-app onboarding instead. Why it works: names the actual problem, picks one bet, sets a falsifiable target and a kill condition. ## What this skill won't do - It won't make text undetectable to AI-detection software, and doesn't try to. Detectors are unreliable either way. The goal is that a *human reader* trusts it. - It won't add facts it doesn't have. A rewrite full of `[ADD]` gaps means the original had no substance yet. That's useful to know. - It can't give you a voice you haven't defined. For that, run `voice-clone` first.