What Anti-Slop Actually Is
Every AI coding agent has a house style, and it is not yours. Ask Claude Code, Cursor or Codex for a landing page and you get the page everyone else gets: a sparkle logo, a beta pill above the headline, a fake terminal reporting 0.0001ms latency, and statistics nobody measured. Anti-Slop is a rulebook that exists to stop exactly that.
It is a free, MIT-licensed set of 38 mandatory rules (R-01 to R-38) that an AI coding agent reads at the start of a session and applies to the UI, copy and code it produces. The repo was created in August 2026 and has passed 4,200 stars on GitHub, with a fresh release (v3.2.20) landing this week. Nothing here runs as an app. Anti-Slop ships as agent skills: folders holding a SKILL.md that your agent loads as standing instructions.
A note on the GitHub language label, which says JavaScript. The installer is an npm package, but the substance is markdown: rules, checklists and examples that the agent reads rather than code that you run. Clone this expecting a codebase and you will find a very well-organised rulebook instead.
The one-line summary the project gives itself is the honest one: a filter, not a style guide. It prescribes no colors, no fonts and no layouts. It removes what should not be there, and the direction that replaces it stays yours.
The Problem: AI Output Has a Default Look
Slop is the default AI look and sound, and you have seen all three versions of it.
In UI: the color-fading banner, the identical rounded cards, the "Unlock the power of..." headline, the invented metrics row. The repo's own four-way comparison is blunt about it: one brief, built with no filter, produced a page claiming a "10,000% ROI Synergy Multiplier" and a "5.0 rating from 500,000 founders", neither of which exists. That is not a design flaw. It is an honesty flaw, and it ships because nothing in the agent's instructions forbids it.
In copy: excitement that says nothing. Emoji bullets, hype in every line, fake-candid openers, social proof with no evidence behind it. DevPik's own AI Text Cleaner exists to strip exactly these artifacts out of text after the fact; Anti-Slop tries to stop them being written in the first place.
In code: box-drawing section banners, emoji comments, and a comment on every constant that restates the constant. If you have ever reviewed AI-written code and deleted twenty comments that added nothing, that is the pattern the code skill hunts. Our AI Code Detector flags the same family of tells from the reviewing side.
The commercial point underneath all of this: AI output currently looks alike, so it all reads as cheap. A filter that removes the tells is worth more than another prompt telling the model to be creative.
How It Works: 38 Rules in Three Tiers
The 38 rules are not a flat checklist. They are tiered by how much judgement they allow.
- Hard Gate rules are absolute. No exceptions, no negotiating with them.
- Purpose-Gate rules ask whether a choice has a purpose, and reject technique that has none.
- Quality Locks apply the same test at the polish level: every shadow, animation, stat and section has to earn its place.
On top of the rules sits a Liveliness Toolkit: three dials (ENERGY, RHYTHM and MOTION) and a Design Read, so the result comes out alive rather than merely clean. This is the part that separates Anti-Slop from a lint pass. Removing slop is subtraction; the toolkit exists so that subtraction does not leave you with something sterile.
The rules reference each other by number, and the skills (next section) build on the core instead of duplicating it, so adding a skill never changes the core. Rule R-37 is the one to remember: a sterile result means the direction was missing, not that the filter failed. Anti-Slop never beautifies on its own.
Six Skills, Loaded Only When a Task Needs Them
Anti-Slop ships as six skills. One core always loads; the other five load only when a task needs them, which matters, because an agent dragging copywriting rules into a database migration is wasting context and diluting instructions.
| Skill | Covers | Pick it for |
|---|---|---|
| antislop | The core filter: rules, tiers, the Delivery Gate, liveliness | Everything. Always loaded |
| antislop-ui | Layout, color, components, decoration, motion, structure | UI and visual work |
| antislop-copywriting | Headlines, CTAs, tone, fake stats, anti-AI writing, markdown hygiene | Copy and text |
| antislop-human | Contrast, keyboard, focus, states | Accessibility |
| antislop-layoutmobile | Reflow across screen widths, breakpoints, grids, overflow, tap targets | Responsive work |
| antislop-code | AI-slop comments: which to remove, which to keep | Comment cleanup |
Install one, several, or the core alone. The human skill (accessibility) deserves a specific mention: low-contrast text and keyboard traps are the two most common ways AI-generated pages fail real people, and the skill pushes agents toward the WCAG 2.2 basics they otherwise skip. If you are comparing skill-based tooling in this ecosystem, Cloudflare's security audit skill takes the same load-only-what-you-need model in a security direction.
The Delivery Gate: A Pass/Fail Report Before Anything Ships
Rules that nobody checks become suggestions. The Delivery Gate is Anti-Slop's enforcement mechanism: before anything ships, the agent must run a mandatory PASS/FAIL report in four blocks that you can read and challenge. A build that fails the gate does not get quietly shipped with a note attached; it gets fixed, or it stays unshipped.
In practice this changes the texture of working with the tool. You are not trusting the agent to have absorbed your taste. You are reading a checklist verdict at the end of every session, the same way you would read a CI result before merging.
One honest caveat, taken from the project's own guide: the gate only bites where Anti-Slop arrives as a skill. Inside a coding agent, a skill is an instruction the agent is built to follow, and the gate runs. Paste the single-file version into a chat window and the rules become reference material instead: still genuinely useful for tone and structure, but no chat app runs the Delivery Gate for you.
Two Modes: Guide New Builds, or Audit What You Already Have
Anti-Slop runs one of two ways, and by default the agent asks which at the start of every session. Both end with a report you can check; they differ in when the rules apply.
- During guides the work while it is built, and the session closes with the Delivery Gate report. Use it when you are building something new.
- After audits work that already exists: a numbered findings list, you approve which items to fix, then a follow-up report. Use it to clean up output you already have.
Being asked every session gets old, so a saved preference is opt-in (it arrived in v3.2.20, the current release). Set it once with npx antislop-ai --mode during. Swap during for after to default to audits, use ask to restore the question, or run --mode alone to see the current setting. The shared settings file is plain JSON (~/.config/antislop/settings.json on Linux and macOS, %APPDATA%\antislop\settings.json on Windows), so if you hand-edit it, running it through a JSON formatter first is worth the ten seconds. When a saved preference is active the skill announces it once ("antislop active: during (global preference)"), and an explicit instruction in the current chat always outranks the saved setting without overwriting it.
Works With the Agent You Already Use
The README lists thirteen supported agents: Claude Code, Codex, Antigravity, OpenCode, Cursor, Cline, Amp, Gemini CLI, Hermes, GitHub Copilot, Kimi Code, Pi and Oh My Pi. The installer covers twelve of them and detects which ones you have. Seven (Claude Code, Antigravity, Codex, Cursor, Kimi Code, Cline and Oh My Pi) also have plugin doors that load Anti-Slop straight from this repository with nothing copied into your project, and Pi installs through its own package manager.
This makes Anti-Slop part of the fast-growing agent-skills ecosystem rather than a single-vendor add-on. The same load-a-markdown-skill model powers Anthropic's Claude for Financial Services agents and the audit skill linked above: standing instructions that ride along with the agent instead of an app you context-switch into. If you are exploring how far that model goes, BuilderIO's agent-native approach is the adjacent read.
Getting Started in One Command
Most people should use the installer. It asks which skills you want, where to install them (this project or everywhere) and which agents you use, then copies the folders and writes the pointer that reloads Anti-Slop every session:
npx antislop-ai
You need Node.js 18 or newer. Two alternatives deliver the same content. The skills.sh directory route, npx skills add miqdadbadjuber/anti-slop, copies the same folders and writes no pointer, so Anti-Slop reloads by description alone. Or take a plugin door: in Claude Code, /plugin marketplace add https://github.com/miqdadbadjuber/anti-slop followed by /plugin install antislop@anti-slop. Every route (install, update, removal) is walked through per agent in the GUIDE.md.
No terminal, or want it in a chat window? The single file antislop.md is a complete filter on its own: fetch it with curl -o antislop.md https://raw.githubusercontent.com/miqdadbadjuber/anti-slop/main/antislop.md and paste it in.
Two operational notes. Skills load when a session starts, so after installing or updating, open a new session or the old rules are still the ones in force. And Anti-Slop does not update itself: npx antislop-ai --update replaces what the installer put down, while plugin installs update through each agent's own plugin command. If you installed through the installer, rerun it in each project to refresh the project pointers.
The Honest Limitation: A Filter Is Not a Style Guide
The most important sentence in the README is the caveat, quoted on its own closing banner: "antislop is a filter, not magic."
The division of labour is explicit. Anti-Slop removes what should not be there; a DESIGN.md file you write supplies the direction that fills the space. The repo's own four-way comparison shows why both halves matter. Anti-Slop alone produced honest copy on a plain, restrained layout ("Plain, because beauty is not its job"). DESIGN.md alone produced the directed look with the invented statistics still in it, because direction does not filter. Both together were the only one of the four that was clean and directed.
So the limitation is real, and disclosed: install Anti-Slop expecting it to art-direct your product and you will get clean, honest output that still looks plain, plus (R-37 again) the correct diagnosis that the direction was missing. If you have no DESIGN.md, the agent labels its work "draft without direction" instead of passing it off as finished. That single labelling rule is, for many teams, worth the install on its own.
Repo Health: 4,200 Stars in Under Two Months
The repository was created on 7 August 2026 and passed 4,200 stars with 285 forks in under two months; the latest push landed two days before this page went up. The star history shows the usual early hockey stick, which is either a community forming or a trend passing, and two months is too early to say which. The more durable signals: releases are frequent (v3.2.20 is current, adding the opt-in mode preference above), the community files (CONTRIBUTING, CODE_OF_CONDUCT, issue and PR templates) are all in place, and the documentation is unusually willing to state what the tool will not do.
Good fit if you build pages or write copy with an AI agent and keep deleting the same tics by hand, or you have a backlog of AI-generated work that needs auditing against a written standard.
Poor fit if what you actually want is taste. Anti-Slop will hand you clean output and wait for you to say what good looks like. Pair it with a DESIGN.md and it earns its keep; treat it as a designer and it will disappoint you in exactly the way its own README predicts.
