/de-slop

Make anything an AI wrote sound like a person.

operations · Free · no sign-up

The failure this kills

You ask your AI to write something quick — a post, an email, a note to your team — and it comes back, and you can just tell. It’s the long dashes running through every sentence. It’s the “it’s not just an update, it’s a revolution” line. It’s the same three adjectives in a row, the “delve” and “leverage” and “robust”, and the “Certainly! Here’s your post” it opens with. Everyone who reads it can tell too, and now the thing you’re about to send quietly says a machine wrote it.

So you either send it as-is and it reads like a robot, or you sit there deleting the giveaways by hand before you can hit send. Both cost you something — one costs credibility, the other costs time you don’t have.

What it does

Say /de-slop (or “de-slop this”, “make this sound human”) and the agent takes whatever the AI just gave you and rewrites it the way a person actually talks:

The long dashes turn back into normal commas and full stops. The em-dash pile-up is the single clearest AI tell, and it’s the first thing to go.
The “it’s not X, it’s Y” reversal gets said straight. So does the rule-of-three rhythm and the inflated transitions.
The filler vocabulary goes back to plain words. “Delve” becomes “look”, “leverage” becomes “use”, “robust” becomes whatever you actually meant. The “Certainly! Here’s…” preamble and the “hope this helps!” trailer get cut entirely.
Every fact stays exactly the same. Names, numbers, dates, prices, links — untouched. It changes how the text sounds, never what it says.
It only rewrites. It never sends. You get the clean version back (on your clipboard if you want it), you eyeball it, and you send it yourself.

When to reach for it

Any time your words go out with your name on them. A LinkedIn post. A cold email. A quick note to your team. A reply to a client. Anywhere you’d rather it didn’t obviously read as “the AI wrote this for me” — this hands you the same message in your own plain voice instead.

It’s not just for writers. Founders, ops, sales — anyone who drafts with an AI and then has to sound like a real person gets the tells stripped in one step.

Install

Download SKILL.md and drop it into your agent’s skills directory:

mkdir -p .claude/skills/de-slop
curl -o .claude/skills/de-slop/SKILL.md https://agentropic.ai/skills/de-slop/SKILL.md

Works with Claude Code and any agent that supports the SKILL.md convention. On macOS it copies the cleaned text via pbcopy; on Linux/WSL it falls back to xclip / wl-copy / clip.exe.

The bigger idea

The point isn’t the em-dash. It’s that AI output is only useful when it lands like you — when the person on the other end reads it and hears you, not a model. Teaching your agent the tells to strip, and the facts to never touch, is exactly the kind of small, sharp discipline we install during a 1 Week Sprint: by Friday your AI isn’t producing drafts that scream “AI” — it’s producing things your team can send as-is, in your own voice.

The skill file

This is the entire skill — copy it, or download SKILL.md.

---
name: de-slop
description: Rewrite AI-written text so it stops sounding like AI — remove the tells (em-dash overuse, the "it's not X, it's Y" reversal, rule-of-three parallel triples, filler vocabulary like delve/leverage/robust, hollow openers, emoji decoration, "Certainly! Here's…" preambles) while keeping every fact, name, number, and link exactly as-is. Trigger with "/de-slop", "de-slop this", "make this sound human", "remove the AI tells", or when the user is about to send something an AI drafted and wants it to read like a person wrote it. Optional args name which text to clean (e.g. "/de-slop the LinkedIn post"). Rewrites only — it never sends anything.
---

# de-slop

Take a piece of AI-written text and rewrite it so a reader can't tell an AI made it, without changing what it actually says.

## What counts as "the text to clean"

1. **If the user named a target** in the args or their message ("the post", "that email", "the last answer"), use that.
2. **Otherwise, use the most recent substantive AI-written block** in the conversation — the last draft, post, email, or answer the user would plausibly send. Not your own meta-chatter.
3. **If the user pasted text**, clean exactly what they pasted.
4. When two candidates are equally likely, ask one quick question before rewriting. Otherwise just proceed.

## The rule that overrides everything: keep the meaning

Rewriting must not change a single fact, name, number, date, price, quote, or link. You are changing how it *sounds*, never what it *claims*. If cleaning a sentence would risk altering its meaning, keep the meaning and accept a slightly less clean sentence.

## The AI tells, and what to do with each

| Tell | What it looks like | Fix |
|---|---|---|
| **Em-dash overuse** | "This works — and it's fast — every time." | Use commas, periods, or "and"/"so". Aim for zero em-dashes. |
| **The reversal** | "It's not just fast, it's effortless." | Say it straight: "It's fast, and it's easy to use." Drop the "not X, it's Y" scaffolding. |
| **Rule-of-three rhythm** | "Clear, concise, and compelling." | Break the parallel. Keep one or two real points, drop the third if it's only there for rhythm. |
| **Filler vocabulary** | delve, leverage, robust, seamless, tapestry, testament, realm, navigate (figurative), underscore, boast, elevate, foster, garner, pivotal, crucial, notably, moreover, furthermore | Swap for the plain word: use, strong, smooth, area, show, key. Cut it if it adds nothing. |
| **Hollow openers/closers** | "In today's fast-paced world…", "In the realm of…", "In conclusion…", "Whether you're X or Y…" | Delete. Start on the actual first real sentence; end on the last real one. |
| **Chatbot scaffolding** | "Certainly! Here's the post you asked for:", "I hope this helps! Let me know if you'd like any changes." | Delete the preamble and the trailer entirely. |
| **Over-hedging** | "It's important to note that…", "Generally speaking,", "It's worth mentioning that…" | Cut the throat-clearing; keep the point that follows. |
| **Emoji decoration** | ✅ bullets, 🚀 in headings, a sprinkle per paragraph | Remove decorative emoji. Keep an emoji only if it carries real meaning the user clearly wants. |
| **Formatting spam** | bolded phrases in every line, a heading over two sentences, bullet lists for prose | Reduce to what a person would actually format. Turn bullet-spam back into sentences where it reads better as prose. |
| **Inflated transitions** | "That said,", "At the end of the day,", "When it comes to X," | Replace with a plain connective (so, and, but) or nothing. |

## Match the person, not a generic "human"

- **Keep the user's own voice if there's a sample.** If earlier messages show how the user writes (their length, their bluntness, their contractions), lean the rewrite toward that. When there's no sample, default to plain, direct, everyday English with contractions.
- **Keep it the same length ballpark** unless the user asks to shorten. De-slopping usually removes 10–25% of the words on its own; don't pad to refill.
- **Don't over-correct into slang.** The goal is "a real person wrote this on a normal day", not "a teenager texting". No forced casualness.

## Steps

1. Identify the text to clean per the rules above.
2. Rewrite it, applying the tells table, preserving every fact/name/number/link.
3. Produce the cleaned text as plain output — nothing but what the user would send.
4. **If the user wants it on the clipboard** (they said "copy" or it's clearly headed somewhere): write the cleaned text verbatim to `<scratchpad>/de-slop.txt` with the Write tool, then copy it — macOS `pbcopy < <path>`; fallbacks Linux `xclip -selection clipboard < <path>` / `wl-copy < <path>`, WSL `clip.exe < <path>`. Otherwise just show it.
5. **Show a one-line change summary** so the user trusts it, e.g. "Removed: 6 em-dashes, 1 'not X, it's Y', 3 filler words (delve, leverage, robust), the 'Certainly!' opener. Facts unchanged." Keep it short.

## Guardrails

- **Rewrite only — never send.** This skill produces text; it does not message, post, or email anyone.
- **Facts are sacred.** Never add, drop, or alter a claim, name, number, or link to make a sentence flow better. When in doubt, keep the meaning.
- **Don't invent a voice the user doesn't have.** If you have no writing sample, stay plain and neutral rather than guessing at a personality.
- **Show your output** before the user sends it anywhere, so they can eyeball it.
- If the text is already clean (few or no tells), say so and change little — don't rewrite for the sake of it.

Watch the walkthrough.

Every skill ships with a video on our YouTube channel — one discipline per episode, field-tested.