How to Stop Shipping AI Slop: The Companion Guide

🧰 The full list with every install command is on the free Anti-AI-Slop Toolkit page. This post covers how to use it.

Merriam-Webster named "slop" its 2025 Word of the Year. Its definition is digital content of low quality, produced usually in quantity by means of AI.

If you build with AI, some of what you ship probably fits. I know some of mine did. This guide covers how to catch it, which tools help, and which parts you still have to do yourself.

First, a disclosure

The presenter in the video version is my digital avatar, built with HeyGen, not me on camera. The research and picks are mine. Nothing here is sponsored.

The test

Simon Willison's framing from 2024 holds up: not all AI-generated content is slop. It's slop when it's "mindlessly generated and thrust upon someone who didn't ask for it."

So the test for anything you're about to publish is simple: did a person with taste stand behind this before it shipped?

Slop has three habits worth remembering:

It looks competent at a glance.
It's cheap to make and expensive to judge.
It's easy to make a thousand of.

The second one is the real cost. Your audience pays for your shortcut with their attention.

Spot it in your own work

Run your last post, script or landing page past these lists before you reach for any tool.

In writing:

words like "delve" and "tapestry"
"It's not just X, it's Y" sentences
openers that clear their throat for a paragraph before saying anything
em dashes in every other sentence
lines that sound deep and say nothing
"experts agree," with no experts named

In design:

Inter on every piece of text
purple-to-blue gradients
rounded cards nested inside rounded cards, all with the same soft grey shadow
frosted glass panels
icons sitting in little circles
a testimonial from "Sarah Chen, CTO," who the model invented

Anthropic's own frontend-design skill adds a sharper list of what AI design clusters around right now:

tracked-out all-caps labels above every heading
meta text joined with middle dots
an arrow tacked onto every button
near-black backgrounds with one acid-green accent
a warm cream background with a terracotta accent, which the skill notes sits close to Anthropic's own Claude color

Our confession: the graphics in my previous video used several of those: all-caps labels, dots between words, arrows everywhere, near-black with neon green. The edit for this video was rebuilt without them.

Taste adds, tools subtract

Every tool below removes defaults. None of them gives your work a point of view. You need both halves, so keep that in mind while you install things.

The design workflow

Install these in this order. You don't need everything; the rule is one generator, one finish gate, one DESIGN.md. Five skills with conflicting opinions is its own kind of slop.

1. Pick a system before any code: UI/UX Pro Max

UI/UX Pro Max (~127K★) looks up your product type and hands your agent a design system before it writes CSS: style, palette, type pairing, and rules for your industry. It ships 192 reasoning rules and 192 palettes, so a spa site doesn't come back looking like a purple SaaS dashboard.

Try this after installing:

Before writing any code, use UI/UX Pro Max to generate a design system for [your product, e.g. "a booking page for a mobile bike repair business"]. Show me the style, palette, type pairing and the anti-patterns to avoid, then wait for my approval.

The "wait for my approval" is the important part. That pause is where your taste goes in.

2. Set the floor: Anthropic's frontend-design

Anthropic's frontend-design skill (883.6K installs on skills.sh) makes the model commit to a direction specific to the brief, then check its own plan against the generic defaults above before it builds. Leave it on for every UI task.

3. Give it rules: a DESIGN.md

A DESIGN.md is a plain-text design system your agent reads before building. awesome-design-md (~116K★) collects 73 of them extracted from real websites. Start from one you admire, then edit it until it describes your brand. Your agent then works from rules you chose.

Read DESIGN.md before building. Follow its colors, type and spacing exactly. If something the page needs isn't covered, ask me instead of inventing it.

4. Gate the finish: Impeccable

Impeccable (~68K★) is the audit before you publish. Run /impeccable init once so it knows your product, then /impeccable audit and /impeccable polish on the page. Its 61 deterministic detector rules run without an API key.

Generation and review are separate jobs. Pro Max builds the system; Impeccable checks the result.

Optional: Taste Skill, for marketing pages only

Taste Skill (~87K★) gives your agent three 1–10 dials for design variance, motion and density. Turn them up for a landing page or portfolio. It gets loud quickly, so pair it with a DESIGN.md so loud still looks like you.

The writing workflow

Humanizer or No AI Slop

Humanizer (~48K★) is the best-known option, built on Wikipedia's Signs of AI writing. Read what it changes, because any rewrite tool can touch prose that was already fine.
No AI Slop (~9.3K★) is the pick when your voice matters most. Use /no-ai-slop is this slop? to get flags without rewrites.

Stars aren't upkeep

A field guide measured 68 of these writing skills in July 2026 and ranked them by maintenance. Its top pick for English, Avoid AI Writing, has a fraction of Humanizer's stars and the best upkeep in the survey. Stop Slop, with over 17,000 stars, hasn't been updated since March 2026 and got an F.

Before you install anything, look at the last commit date, not the star count.

About detectors

Detectors flag slop; they don't fix it. Most are small, early projects, so treat them as experiments: an in-browser 270M-parameter model, a YouTube extension that flags AI scripts, a regex CLI with no AI at all, and a GitHub Action that closes slop pull requests. The full list is on the toolkit page.

Paid "undetectable" humanizers are built to beat detectors. That's a different goal from sounding like a person with taste.

The seven rules no tool does for you

Start from something real. Your notes, a voice memo, something you actually did, never "write a post about X."
Add one detail only you could write. A number from your own business, a mistake you made, a name.
Take a position. Models default to safe and hedged, so pick a side and say why.
Cut the fingerprints. Run the lists above, or a skill that runs them for you.
Read it out loud, and check the UI at full zoom. Slop hides in the parts you skim.
Fact-check before you polish. A clean sentence carrying a wrong claim is worse than a rough true one.
Keep the last pass human. The final read before publishing is yours.

The short version

Pro Max for the system, frontend-design for the floor, Impeccable for the audit, Humanizer or No AI Slop for the words, plus a DESIGN.md and the seven rules. Every install command is on the Anti-AI-Slop Toolkit page.

Want someone to tell you when your work reads like slop, before your audience does? That's what the AI Creators Roundtable is for: creators and solo business owners trading real workflows and honest feedback.